diff --git a/.claude/design/core/ai/agentCall-declarative-tools.md b/.claude/design/core/ai/agentCall-declarative-tools.md new file mode 100644 index 0000000000..bc61756fb7 --- /dev/null +++ b/.claude/design/core/ai/agentCall-declarative-tools.md @@ -0,0 +1,331 @@ +# agentCall 声明式工具改造设计 + +## 1. 背景 + +当前 `packages/service/core/ai/llm/agentCall/index.ts` 的 `runAgentLoop` 通过四个分离的参数处理工具调度： + +- `body.tools`：喂给 LLM 的 schema 数组 +- `onToolCall`：LLM 识别到工具调用时的流式回调 +- `onToolParam`：工具参数流式增量回调 +- `onRunTool`：实际执行工具的总入口 + +调用方（`toolCall.ts` / `masterCall.ts`）在 `onRunTool` 内写了一长串 `if (toolId === X) else if (toolId === Y)` 分支，每个分支都要独立做 `parseJsonArgs + XxxSchema.safeParse + 错误处理`。新增工具必须改这个巨型函数，schema / 解析 / 执行三段逻辑被拆散在不同参数里。 + +## 2. 目标 + +1. **声明式**：一个工具自带 schema、参数解析、执行逻辑，三段聚合到一个对象里。 +2. **两阶段执行**：所有工具统一 `parseParams`（解析 + 校验）→ `execute`（执行）两个阶段，消除分支里重复的校验代码。 +3. **生命周期钩子**：流式事件（`onToolCall / onToolParam / onAfterToolCall`）保持全局，由 `runAgentLoop` 统一编排，工具定义不感知 UI 层。 +4. **`runAgentLoop` 自身不感知具体工具种类**：核心循环只负责调度，新增工具不需要改 `agentCall` 模块。 + +本文档只覆盖 **`agentCall` 模块自身** 的改造，应用层（`toolCall.ts` / `masterCall.ts`）如何迁移在后续文档单独讨论。 + +## 3. 目录结构与类型定义 + +声明式工具的**类型定义与执行服务**放在独立目录 `packages/service/core/ai/llm/toolCall/` 下管理，与 `agentCall/` 解耦（`agentCall` 负责多轮调度，`toolCall` 负责单次工具调用的解析与执行；后者是前者的依赖）： + +``` +packages/service/core/ai/llm/ +├── agentCall/ +│ └── index.ts # 多轮调度，import from ../toolCall +├── toolCall/ +│ ├── type.ts # ToolDefinition、ToolExecuteContext、ToolExecuteResult、ToolParseResult +│ └── index.ts # runTool 两阶段执行器 +├── request.ts +└── ... +``` + +因为类型不再属于 `agentCall` 私有命名空间，`AgentToolDefinition` 去掉 `Agent` 前缀，统一命名为 `ToolDefinition`（其他类型同理）。 + +新建 `packages/service/core/ai/llm/toolCall/type.ts`： + +```ts +import type { + ChatCompletionMessageParam, + ChatCompletionMessageToolCall, + ChatCompletionTool +} from '@fastgpt/global/core/ai/llm/type'; +import type { ChatNodeUsageType } from '@fastgpt/global/support/wallet/bill/type'; +import type { WorkflowInteractiveResponseType } from '@fastgpt/global/core/workflow/template/system/interactive/type'; + +// 参数解析结果：成功返回强类型 data，失败返回要回填给 LLM 的 errorMessage +export type ToolParseResult

= + | { success: true; data: P } + | { success: false; errorMessage: string }; + +// 执行上下文 +export type ToolExecuteContext

= { + call: ChatCompletionMessageToolCall; // 原始工具调用 + messages: ChatCompletionMessageParam[]; // 当前 requestMessages 快照 + params: P; // parseParams 输出 +}; + +// 执行结果（结构与现有 onRunTool 的返回值对齐） +export type ToolExecuteResult = { + response: string; + assistantMessages?: ChatCompletionMessageParam[]; + usages?: ChatNodeUsageType[]; + interactive?: WorkflowInteractiveResponseType; + stop?: boolean; +}; + +// 声明式工具定义 +export type ToolDefinition

= { + // 1. 喂给 LLM 的 schema（name/description/parameters） + schema: ChatCompletionTool; + + // 2. 参数解析阶段，可选；缺省走 parseJsonArgs，参数类型为 Record + parseParams?: (rawArgs: string) => ToolParseResult

; + + // 3. 执行阶段（必填） + execute: (ctx: ToolExecuteContext

) => Promise; +}; +``` + +设计要点： + +- `parseParams` 的返回类型强制调用方处理校验失败，失败文案会作为 `response` 回写给 LLM（保持当前代码行为，让模型能看到错误自行纠偏）。 +- `execute` 的 `params` 通过泛型 `P` 串联，从 `parseParams` 的 `data` 类型收窄而来；调用方编写 `execute` 时不再需要重复 `safeParse`。 +- 返回值沿用现有的 `response / assistantMessages / usages / interactive / stop` 字段，迁移时不需要对 `agentCall` 循环体里"如何消费这些字段"做任何改动。 + +## 4. `runAgentLoop` Props 变更 + +### 4.1 删除的 props + +```ts +body.tools: ChatCompletionTool[] +onToolCall: (e: { call }) => void +onToolParam: (e: { tool; params }) => void +onRunTool: (e: { call; messages }) => Promise<...> +``` + +### 4.2 新增 / 替换的 props + +```ts +type RunAgentCallProps = { + // ... 其他不变（maxRunAgentTimes、childrenInteractiveParams、handleInteractiveTool、 + // onAfterCompressContext、onToolCompress、usagePush、isAborted、userKey、onReasoning、onStreaming 等） + + body: CreateLLMResponseProps['body'] & { + // tools 字段被移除 + temperature?: number; + top_p?: number; + stream?: boolean; + }; + + // 声明式的工具集合（schema + 执行逻辑） + // 所有工具必须在调用 runAgentLoop 前完整枚举。LLM 能看到的工具集 ≡ 能执行的工具集。 + // 动态场景（用户 SubApp、capability 等）由调用方在构建 tools 数组时提前展开。 + tools: ToolDefinition[]; + + // 生命周期钩子（统一编排） + onToolCall?: (e: { call: ChatCompletionMessageToolCall }) => void; + onToolParam?: (e: { tool: ChatCompletionMessageToolCall; argsDelta: string }) => void; + onAfterToolCall?: (e: { + call: ChatCompletionMessageToolCall; + response: string; + }) => void; +}; +``` + +### 4.3 `onToolParam` 字段重命名 + +当前 `onToolParam` 的 `params` 字段传的是**本次增量** `arg`（参见 `request.ts:462`：`onToolParam?.({ tool: currentTool, params: arg })`），字段名容易误解为完整参数。本次一并重命名为 `argsDelta`： + +- `packages/service/core/ai/llm/request.ts:44`：类型定义 `params: string` → `argsDelta: string` +- `packages/service/core/ai/llm/request.ts:462`：调用处 `{ tool, params: arg }` → `{ tool, argsDelta: arg }` +- 所有调用方同步修改（调用方文档里列出） + +## 5. 内部实现 + +### 5.1 新建 `toolCall/index.ts` + +统一的两阶段执行器（对外暴露 `runTool` 作为 `toolCall` 服务的入口）： + +```ts +import { parseJsonArgs } from '../../utils'; +import { getErrText } from '@fastgpt/global/common/error/utils'; +import type { + ToolDefinition, + ToolExecuteResult, + ToolParseResult +} from './type'; +import type { + ChatCompletionMessageParam, + ChatCompletionMessageToolCall +} from '@fastgpt/global/core/ai/llm/type'; + +type RunToolArgs = { + call: ChatCompletionMessageToolCall; + messages: ChatCompletionMessageParam[]; + tools: ToolDefinition[]; +}; + +export const runTool = async ({ + call, + messages, + tools +}: RunToolArgs): Promise => { + const name = call.function.name; + const def = tools.find((t) => t.schema.function.name === name); + + // 1. 工具未找到（LLM hallucination 或 tools 配置漏项）：兜底 response，外层仍会触发 onAfterToolCall + if (!def) { + return { response: `Call tool not found: ${name}` }; + } + + // 2. 阶段一：解析 + const parseResult: ToolParseResult = def.parseParams + ? def.parseParams(call.function.arguments ?? '') + : { success: true, data: parseJsonArgs(call.function.arguments ?? '') }; + + if (!parseResult.success) { + return { response: parseResult.errorMessage }; + } + + // 3. 阶段二：执行（统一 try/catch） + try { + return await def.execute({ + call, + messages, + params: parseResult.data + }); + } catch (error) { + return { response: `Tool error: ${getErrText(error)}` }; + } +}; +``` + +要点： + +- `tools.find` 用 `schema.function.name` 查，未命中即兜底，不再提供动态解析通道。 +- 任何"失败"（未找到 / 解析失败 / 执行抛错）都归一成 `{ response: string }`，外层流程不区分。 +- `execute` 内部闭包捕获到的副作用（`childrenResponses.push` / `toolRunResponses.push` / `planResult = ...`）保持原样，`runner` 不感知。 + +### 5.2 改造 `agentCall/index.ts` + +从 `toolCall` 模块引入类型和 runner： + +```ts +import type { ToolDefinition } from '../toolCall/type'; +import { runTool } from '../toolCall'; +``` + +以下只列出"变化点"，其余不动： + +**1) LLM 请求部分的 `body.tools`** + +```ts +// 改造前 +tools, // 直接来自 props.body.tools + +// 改造后 +tools: tools.map((t) => t.schema), // 来自 props.tools，运行时 .map 提取 schema +``` + +**2) 循环体内部的工具调用** + +```ts +// 改造前（line 339-349） +for await (const tool of toolCalls) { + const { response, assistantMessages, usages, interactive, stop } = + await onRunTool({ + call: tool, + messages: cloneRequestMessages + }); + ... +} + +// 改造后 +for await (const toolCall of toolCalls) { + const result = await runTool({ + call: toolCall, + messages: cloneRequestMessages, + tools + }); + + onAfterToolCall?.({ call: toolCall, response: result.response }); + + const { + response, + assistantMessages: toolAssistantMessages = [], + usages: toolUsages = [], + interactive, + stop + } = result; + + // 以下压缩 / 消息追加 / interactive 处理逻辑完全不变 + ... +} +``` + +**3) `createLLMResponse` 的钩子透传** + +```ts +// 改造前 +onToolCall, +onToolParam + +// 改造后（字段名一致，内部定义改名后透传不变；外部 props 也保留 onToolCall/onToolParam 语义） +onToolCall, +onToolParam // 注意透传给 createLLMResponse 的结构里字段要同步改为 argsDelta +``` + +### 5.3 生命周期触发时机汇总 + +| 钩子 | 触发位置 | 参数 | +|---|---|---| +| `onToolCall` | `createLLMResponse` 解析出新 tool 时（`request.ts:452`）| `{ call }` | +| `onToolParam` | `createLLMResponse` 每次累积到 args 增量时（`request.ts:462`）| `{ tool, argsDelta }` | +| `onAfterToolCall` | `runTool` 返回后，压缩和消息追加之前 | `{ call, response }` | + +`onAfterToolCall` 在 notFound / parseParams 失败 / execute 抛错时一样会被触发——UI 层事件流不断档。 + +## 6. 文件清单 + +``` +新建目录： + packages/service/core/ai/llm/toolCall/ + ├── type.ts # ToolDefinition / ToolExecuteContext / ToolExecuteResult / ToolParseResult + └── index.ts # runTool 两阶段执行器（对外导出入口） + +改动： + packages/service/core/ai/llm/agentCall/index.ts + - 从 ../toolCall 引入 ToolDefinition 和 runTool + - props: 删 body.tools / onRunTool + - props: 加 tools / onAfterToolCall + - props: 保留 onToolCall / onToolParam 作为生命周期钩子（语义不变，字段名对齐 argsDelta） + - LLM body.tools 改为 props.tools.map(t => t.schema) + - 循环体 onRunTool → runTool + - onAfterToolCall 触发点 + + packages/service/core/ai/llm/request.ts + - onToolParam 类型：params: string → argsDelta: string（44 行） + - onToolParam 调用：params: arg → argsDelta: arg（462 行） +``` + +## 7. 与现有单测的关系 + +需要检查： + +- `test/cases/service/core/ai/llm/request.test.ts` 对 `onToolParam` 的断言是否使用 `params` 字段。 +- 改造完成后至少补一个 `packages/service/core/ai/llm/toolCall/` 的单测（建议测试文件放在 `test/cases/service/core/ai/llm/toolCall/` 下）覆盖：工具命中 / 未命中（LLM hallucination）/ parseParams 失败 / execute 抛错 四种路径。 + +## 8. 待确认问题 + +1. **`onAfterToolCall` 的触发粒度**：目前是 `runTool` 返回后触发一次，不含压缩后的 response。如果 UI 需要看到"压缩后的 tool response"，应该让 `onAfterToolCall` 接收压缩后的值 —— 但这会与现有 `onToolCompress`（已经单独推送压缩产物）重复。建议 `onAfterToolCall` 接收**原始 response**，与 `onToolCompress` 解耦。 + +2. **`body.tools` 去除后的类型收敛**：`CreateLLMResponseProps['body']` 这个类型本身可能没有 `tools` 字段，而是 agentCall 的扩展类型加进去的。需要确认并更新扩展类型定义。 + +## 9. 改造分步 TODO + +- [ ] 新建目录 `packages/service/core/ai/llm/toolCall/` +- [ ] 新建 `toolCall/type.ts` 定义 `ToolDefinition` / `ToolExecuteContext` / `ToolExecuteResult` / `ToolParseResult` +- [ ] 新建 `toolCall/index.ts` 实现并导出 `runTool` +- [ ] 改 `request.ts`：`onToolParam` 的 `params` → `argsDelta` +- [ ] 改 `agentCall/index.ts`：props 重构 + 从 `../toolCall` 引入 + LLM body.tools 提取 + 循环体接入 `runTool` + `onAfterToolCall` 触发 +- [ ] 为 `toolCall/` 补单测（四条路径：命中 / 未命中 / parseParams 失败 / execute 抛错） +- [ ] 跑一遍 `agentCall` 相关现有单测，确认类型编译通过 +- [ ] 调用方（`toolCall.ts`（workflow 层同名但不同路径的文件）/ `masterCall.ts` / 其他）的迁移放在**后续文档**里讨论，此步**先不动** + +> 命名冲突提示：`packages/service/core/workflow/dispatch/ai/tool/toolCall.ts` 是 workflow dispatch 层的文件名，与本次新建的 `packages/service/core/ai/llm/toolCall/` 目录同名但路径不同，不会产生 import 冲突。后续迁移时两者需要区分清楚。 diff --git a/.claude/design/core/ai/gradient-pricing-fix.md b/.claude/design/core/ai/gradient-pricing-fix.md index 17f51f9e7e..00296e73bd 100644 --- a/.claude/design/core/ai/gradient-pricing-fix.md +++ b/.claude/design/core/ai/gradient-pricing-fix.md @@ -65,12 +65,12 @@ const { totalPoints: modelTotalPoints } = formatModelChars2Points({ }); ``` -`runToolCall` 调用 `runAgentCall` 时**不传 `usagePush`**，所以单次计价全部丢失，只依赖这里的累加计算 → **实际计费错误**。 +`runToolCall` 调用 `runAgentLoop` 时**不传 `usagePush`**，所以单次计价全部丢失，只依赖这里的累加计算 → **实际计费错误**。 ### 3. `packages/service/core/workflow/dispatch/ai/agent/master/call.ts` (masterCall) — **展示 BUG** ```ts -// inputTokens = runAgentCall 返回的累加值 +// inputTokens = runAgentLoop 返回的累加值 const llmUsage = formatModelChars2Points({ inputTokens, // ❌ 累加值 outputTokens @@ -101,15 +101,15 @@ const { totalPoints } = formatModelChars2Points({ **不应用累加的 token 数计算价格，而应该每次 LLM 调用单独计价，再累加价格。** -### 方案：`runAgentCall` 返回预计算的 `llmTotalPoints` +### 方案：`runAgentLoop` 返回预计算的 `llmTotalPoints` -在 `runAgentCall` 的 while 循环中，每次 LLM 调用后立即计算该次的价格，并累加到 `llmTotalPoints`，最终将其作为返回值之一。调用方直接使用该预计算值，而不再重复调用 `formatModelChars2Points(累加 tokens)`。 +在 `runAgentLoop` 的 while 循环中，每次 LLM 调用后立即计算该次的价格，并累加到 `llmTotalPoints`，最终将其作为返回值之一。调用方直接使用该预计算值，而不再重复调用 `formatModelChars2Points(累加 tokens)`。 --- ## 具体修改 -### 修改 1：`runAgentCall` — 增加 `llmTotalPoints` 返回值 +### 修改 1：`runAgentLoop` — 增加 `llmTotalPoints` 返回值 **文件**：`packages/service/core/ai/llm/agentCall/index.ts` @@ -155,8 +155,8 @@ type ResponseType = { toolCallOutputTokens: number; // 保留展示用 }; -// runAgentCall 返回后 -const { inputTokens, outputTokens, llmTotalPoints, ... } = await runAgentCall(...); +// runAgentLoop 返回后 +const { inputTokens, outputTokens, llmTotalPoints, ... } = await runAgentLoop(...); return { ... @@ -189,8 +189,8 @@ const modelTotalPoints = toolCallTotalPoints; // 直接使用预计算值， **文件**：`packages/service/core/workflow/dispatch/ai/agent/master/call.ts` ```ts -// runAgentCall 返回 llmTotalPoints -const { inputTokens, outputTokens, llmTotalPoints, childrenUsages, ... } = await runAgentCall(...); +// runAgentLoop 返回 llmTotalPoints +const { inputTokens, outputTokens, llmTotalPoints, childrenUsages, ... } = await runAgentLoop(...); // 修改前（❌） const llmUsage = formatModelChars2Points({ model: agentModel, inputTokens, outputTokens }); @@ -265,7 +265,7 @@ usage.outputTokens += regenResult.usage.outputTokens; ## TODO -- [ ] 修改 `runAgentCall` 返回类型，新增 `llmTotalPoints` +- [ ] 修改 `runAgentLoop` 返回类型，新增 `llmTotalPoints` - [ ] 修改 `runToolCall` 返回类型，新增 `toolCallTotalPoints` - [ ] 修改 `dispatchRunTools` 使用预计算值 - [ ] 修改 `masterCall` 使用预计算值（修正展示） diff --git a/.claude/design/core/ai/sandbox/get-file-url.md b/.claude/design/core/ai/sandbox/get-file-url.md index 8d8f812b51..08133b96e9 100644 --- a/.claude/design/core/ai/sandbox/get-file-url.md +++ b/.claude/design/core/ai/sandbox/get-file-url.md @@ -200,7 +200,7 @@ export const dispatchSandboxGetFileUrl = async ({ #### 3.5.1 普通工作流：toolCall.ts -`handleToolResponse` 中合并 `SANDBOX_TOOL_NAME` 和 `SANDBOX_GET_FILE_URL_TOOL_NAME` 到同一拦截块： +`onRunTool` 中合并 `SANDBOX_TOOL_NAME` 和 `SANDBOX_GET_FILE_URL_TOOL_NAME` 到同一拦截块： ```typescript if ( diff --git a/document/content/docs/self-host/upgrading/4-15/4150.mdx b/document/content/docs/self-host/upgrading/4-15/4150.mdx new file mode 100644 index 0000000000..60d4ba4a60 --- /dev/null +++ b/document/content/docs/self-host/upgrading/4-15/4150.mdx @@ -0,0 +1,16 @@ +--- +title: 'V4.15.0(进行中)' +description: 'FastGPT V4.15.0 更新说明' +--- + +## 🚀 新增内容 + + +## ⚙️ 优化 + + +## 🐛 修复 + +## 代码优化 + +1. 优化 Agent tool 声明和运行，统一所有 tool 的声明和运行方式。 \ No newline at end of file diff --git a/document/content/docs/self-host/upgrading/4-15/meta.en.json b/document/content/docs/self-host/upgrading/4-15/meta.en.json new file mode 100644 index 0000000000..1fc1786916 --- /dev/null +++ b/document/content/docs/self-host/upgrading/4-15/meta.en.json @@ -0,0 +1,5 @@ +{ + "title": "4.15.x", + "description": "", + "pages": ["4150"] +} diff --git a/document/content/docs/self-host/upgrading/4-15/meta.json b/document/content/docs/self-host/upgrading/4-15/meta.json new file mode 100644 index 0000000000..1fc1786916 --- /dev/null +++ b/document/content/docs/self-host/upgrading/4-15/meta.json @@ -0,0 +1,5 @@ +{ + "title": "4.15.x", + "description": "", + "pages": ["4150"] +} diff --git a/document/content/docs/self-host/upgrading/meta.en.json b/document/content/docs/self-host/upgrading/meta.en.json index 64eb6a2b89..d26749060f 100644 --- a/document/content/docs/self-host/upgrading/meta.en.json +++ b/document/content/docs/self-host/upgrading/meta.en.json @@ -1,5 +1,5 @@ { "title": "Version History", "description": "FastGPT version history", - "pages": ["4-14", "4-13", "4-12", "outdated"] + "pages": ["4-15", "4-14", "4-13", "4-12", "outdated"] } diff --git a/document/content/docs/self-host/upgrading/meta.json b/document/content/docs/self-host/upgrading/meta.json index e73955a160..73238bb3ed 100644 --- a/document/content/docs/self-host/upgrading/meta.json +++ b/document/content/docs/self-host/upgrading/meta.json @@ -1,5 +1,5 @@ { "title": "版本列表", "description": "FastGPT 版本列表", - "pages": ["4-14", "4-13", "4-12", "outdated"] + "pages": ["4-15", "4-14", "4-13", "4-12", "outdated"] } diff --git a/document/content/docs/toc.mdx b/document/content/docs/toc.mdx index 2faadf602b..8f07c74636 100644 --- a/document/content/docs/toc.mdx +++ b/document/content/docs/toc.mdx @@ -127,6 +127,7 @@ description: FastGPT 文档目录 - [/docs/self-host/upgrading/4-14/4148](/docs/self-host/upgrading/4-14/4148) - [/docs/self-host/upgrading/4-14/41481](/docs/self-host/upgrading/4-14/41481) - [/docs/self-host/upgrading/4-14/4149](/docs/self-host/upgrading/4-14/4149) +- [/docs/self-host/upgrading/4-15/4150](/docs/self-host/upgrading/4-15/4150) - [/docs/self-host/upgrading/outdated/40](/docs/self-host/upgrading/outdated/40) - [/docs/self-host/upgrading/outdated/41](/docs/self-host/upgrading/outdated/41) - [/docs/self-host/upgrading/outdated/4100](/docs/self-host/upgrading/outdated/4100) diff --git a/document/data/doc-last-modified.json b/document/data/doc-last-modified.json index ee02468543..d042276cf6 100644 --- a/document/data/doc-last-modified.json +++ b/document/data/doc-last-modified.json @@ -224,8 +224,12 @@ "document/content/docs/self-host/upgrading/4-14/4141.mdx": "2026-03-03T17:39:47+08:00", "document/content/docs/self-host/upgrading/4-14/41410.en.mdx": "2026-03-31T23:15:29+08:00", "document/content/docs/self-host/upgrading/4-14/41410.mdx": "2026-04-18T20:47:39+08:00", + "document/content/docs/self-host/upgrading/4-14/41411.en.mdx": "2026-04-21T23:04:26+08:00", "document/content/docs/self-host/upgrading/4-14/41411.mdx": "2026-04-20T20:18:35+08:00", - "document/content/docs/self-host/upgrading/4-14/41412.mdx": "2026-04-20T20:18:35+08:00", + "document/content/docs/self-host/upgrading/4-14/41412.en.mdx": "2026-04-21T23:04:26+08:00", + "document/content/docs/self-host/upgrading/4-14/41412.mdx": "2026-04-21T23:04:26+08:00", + "document/content/docs/self-host/upgrading/4-14/41413.en.mdx": "2026-04-21T23:04:26+08:00", + "document/content/docs/self-host/upgrading/4-14/41413.mdx": "2026-04-21T23:04:26+08:00", "document/content/docs/self-host/upgrading/4-14/4142.en.mdx": "2026-03-03T17:39:47+08:00", "document/content/docs/self-host/upgrading/4-14/4142.mdx": "2026-03-03T17:39:47+08:00", "document/content/docs/self-host/upgrading/4-14/4143.en.mdx": "2026-03-03T17:39:47+08:00", @@ -386,8 +390,8 @@ "document/content/docs/self-host/upgrading/outdated/499.mdx": "2026-03-03T17:39:47+08:00", "document/content/docs/self-host/upgrading/upgrade-intruction.en.mdx": "2026-03-03T17:39:47+08:00", "document/content/docs/self-host/upgrading/upgrade-intruction.mdx": "2026-04-20T13:51:34+08:00", - "document/content/docs/toc.en.mdx": "2026-04-17T23:28:43+08:00", - "document/content/docs/toc.mdx": "2026-04-20T17:45:22+08:00", + "document/content/docs/toc.en.mdx": "2026-04-21T23:04:26+08:00", + "document/content/docs/toc.mdx": "2026-04-21T23:04:26+08:00", "document/content/docs/use-cases/app-cases/dalle3.en.mdx": "2026-02-26T22:14:30+08:00", "document/content/docs/use-cases/app-cases/dalle3.mdx": "2025-07-23T21:35:03+08:00", "document/content/docs/use-cases/app-cases/english_essay_correction_bot.en.mdx": "2026-02-26T22:14:30+08:00", diff --git a/packages/global/core/ai/sandbox/constants.ts b/packages/global/core/ai/sandbox/constants.ts index a2ffbd5f5f..f5d32f7ce4 100644 --- a/packages/global/core/ai/sandbox/constants.ts +++ b/packages/global/core/ai/sandbox/constants.ts @@ -15,10 +15,10 @@ export const SANDBOX_SUSPEND_MINUTES = 5; // ---- sandboxId 生成 ---- export const generateSandboxId = (appId: string, userId: string, chatId: string): string => { - return hashStr(`${appId}-${userId}-${chatId}`).slice(0, 16); + return hashStr(`${String(appId)}-${String(userId)}-${String(chatId)}`).slice(0, 16); }; -// Tool +// Shell Tool export const SANDBOX_NAME: I18nStringType = { 'zh-CN': '虚拟机', 'zh-Hant': '虛擬機', @@ -26,10 +26,6 @@ export const SANDBOX_NAME: I18nStringType = { }; export const SANDBOX_ICON = 'core/app/sandbox/sandbox' as const; export const SANDBOX_TOOL_NAME = 'sandbox_shell'; -export const SandboxShellToolSchema = z.object({ - command: z.string(), - timeout: z.number().optional() -}); export const SANDBOX_SHELL_TOOL: ChatCompletionTool = { type: 'function', function: { @@ -51,15 +47,13 @@ export const SANDBOX_SHELL_TOOL: ChatCompletionTool = { } }; +// Get File URL Tool export const SANDBOX_READ_FILE_TOOL_NAME: I18nStringType = { 'zh-CN': '虚拟机/获取文件链接', 'zh-Hant': '虛擬機/獲取文件鏈接', en: 'Sandbox/Get File URL' }; export const SANDBOX_GET_FILE_URL_TOOL_NAME = 'sandbox_get_file_url'; -export const SandboxGetFileUrlToolSchema = z.object({ - paths: z.array(z.string()) -}); export const SANDBOX_GET_FILE_URL_TOOL: ChatCompletionTool = { type: 'function', function: { @@ -82,10 +76,30 @@ export const SANDBOX_GET_FILE_URL_TOOL: ChatCompletionTool = { } }; -export const SANDBOX_TOOLS: ChatCompletionTool[] = [SANDBOX_SHELL_TOOL, SANDBOX_GET_FILE_URL_TOOL]; - +// Prompt export const SANDBOX_SYSTEM_PROMPT = `你拥有一个独立的 Linux 沙盒环境（Ubuntu 22.04），可通过 ${SANDBOX_TOOL_NAME} 工具执行命令： - 预装：bash / python3 / node / bun / git / curl - 可自行安装软件包（apt / pip / npm） - 生成的文件内容都保存在当前目录下即可 - 若需要将生成的文件分享给用户，可使用 ${SANDBOX_GET_FILE_URL_TOOL_NAME} 工具获取文件的临时访问链接`; + +// 聚合 +export const sandboxToolMap: Record< + string, + { schema: ChatCompletionTool; name: I18nStringType; avatar: string; toolDescription: string } +> = { + [SANDBOX_TOOL_NAME]: { + schema: SANDBOX_SHELL_TOOL, + name: SANDBOX_NAME, + avatar: SANDBOX_ICON, + toolDescription: SANDBOX_SHELL_TOOL.function.description! + }, + [SANDBOX_GET_FILE_URL_TOOL_NAME]: { + schema: SANDBOX_GET_FILE_URL_TOOL, + name: SANDBOX_READ_FILE_TOOL_NAME, + avatar: SANDBOX_ICON, + toolDescription: SANDBOX_GET_FILE_URL_TOOL.function.description! + } +}; + +export const SANDBOX_TOOLS = Object.values(sandboxToolMap).map((item) => item.schema); diff --git a/packages/global/core/chat/adapt.ts b/packages/global/core/chat/adapt.ts index 53df999de7..62ed7e6ae6 100644 --- a/packages/global/core/chat/adapt.ts +++ b/packages/global/core/chat/adapt.ts @@ -217,7 +217,7 @@ export const GPTMessages2Chats = ({ messages: ChatCompletionMessageParam[]; reserveTool?: boolean; reserveReason?: boolean; - getToolInfo?: (name: string) => { name: string; avatar: string }; + getToolInfo?: (name: string) => { name: string; avatar?: string } | undefined; }): ChatItemMiniType[] => { const chatMessages = messages .map((item) => { diff --git a/packages/global/core/workflow/node/agent/constants.ts b/packages/global/core/workflow/node/agent/constants.ts index 0e00a4a09e..ac52639bbc 100644 --- a/packages/global/core/workflow/node/agent/constants.ts +++ b/packages/global/core/workflow/node/agent/constants.ts @@ -1,37 +1,20 @@ -import { - SANDBOX_GET_FILE_URL_TOOL, - SANDBOX_ICON, - SANDBOX_NAME, - SANDBOX_READ_FILE_TOOL_NAME, - SANDBOX_SHELL_TOOL -} from '../../../ai/sandbox/constants'; -import type { I18nStringType } from '../../../../common/i18n/type'; +import type { I18nStringType, localeType } from '../../../../common/i18n/type'; +import { sandboxToolMap } from '../../../ai/sandbox/constants'; import { skillToolsMap } from './skillTools'; +import { parseI18nString } from '../../../../common/i18n/utils'; export enum SubAppIds { plan = 'plan_agent', ask = 'ask_agent', model = 'model_agent', fileRead = 'file_read', - datasetSearch = 'dataset_search', - sandboxTool = 'sandbox_shell', - sandboxGetFileUrl = 'sandbox_get_file_url' + datasetSearch = 'dataset_search' } export const systemSubInfo: Record< string, { name: I18nStringType; avatar: string; toolDescription: string } > = { - [SubAppIds.sandboxTool]: { - name: SANDBOX_NAME, - avatar: SANDBOX_ICON, - toolDescription: SANDBOX_SHELL_TOOL.function.description! - }, - [SubAppIds.sandboxGetFileUrl]: { - name: SANDBOX_READ_FILE_TOOL_NAME, - avatar: SANDBOX_ICON, - toolDescription: SANDBOX_GET_FILE_URL_TOOL.function.description! - }, [SubAppIds.plan]: { name: { 'zh-CN': '规划Agent', @@ -78,5 +61,16 @@ export const systemSubInfo: Record< avatar: 'core/workflow/template/agent', toolDescription: '调用 LLM 模型完成一些通用任务。' }, + ...sandboxToolMap, ...skillToolsMap }; +export const getSystemToolInfo = (id: string, lang: localeType = 'en') => { + if (id in systemSubInfo) { + const info = systemSubInfo[id]; + return { + name: parseI18nString(info.name, lang), + avatar: info.avatar, + toolDescription: info.toolDescription + }; + } +}; diff --git a/packages/global/core/workflow/node/agent/skillTools.ts b/packages/global/core/workflow/node/agent/skillTools.ts index 96077748bd..e78b9c6f31 100644 --- a/packages/global/core/workflow/node/agent/skillTools.ts +++ b/packages/global/core/workflow/node/agent/skillTools.ts @@ -1,5 +1,7 @@ import z from 'zod'; import type { ChatCompletionTool } from '../../../ai/llm/type'; +import type { I18nStringType, localeType } from '../../../../common/i18n/type'; +import { parseI18nString } from '../../../../common/i18n/utils'; export enum SandboxToolIds { readFile = 'sandbox_read_file', @@ -10,7 +12,10 @@ export enum SandboxToolIds { fetchUserFile = 'sandbox_fetch_user_file' } -export const skillToolsMap = { +export const skillToolsMap: Record< + string, + { name: I18nStringType; avatar: string; toolDescription: string } +> = { // Sandbox tools [SandboxToolIds.readFile]: { name: { @@ -73,6 +78,19 @@ export const skillToolsMap = { 'Download a user-uploaded file (document or image) from the conversation and write it as a binary file into the sandbox filesystem. Use this when a skill script needs to process a raw file. Workflow: call this tool first to place the file at target_path (relative to workspace), then run skill scripts that read from that path.' } }; +export const getSkillToolInfo = ( + id: string, + lang: localeType = 'en' +): { name: string; avatar: string; toolDescription: string } | undefined => { + const toolInfo = skillToolsMap[id]; + if (toolInfo) { + return { + name: parseI18nString(toolInfo.name, lang), + avatar: toolInfo.avatar, + toolDescription: toolInfo.toolDescription + }; + } +}; // Zod parameter schemas (runtime validation) export const SandboxReadFileSchema = z.object({ diff --git a/packages/service/common/logger/categories.ts b/packages/service/common/logger/categories.ts index 3ba708245b..d0b82913e0 100644 --- a/packages/service/common/logger/categories.ts +++ b/packages/service/common/logger/categories.ts @@ -76,6 +76,7 @@ export const LogCategories = { }), AI: Object.assign(['ai'], { AGENT: ['ai', 'agent'], + TOOL_CALL: ['ai', 'tool-call'], HELPERBOT: ['ai', 'helperbot'], CONFIG: ['ai', 'config'], EMBEDDING: ['ai', 'embedding'], diff --git a/packages/service/core/ai/llm/agentCall/index.ts b/packages/service/core/ai/llm/agentLoop/index.ts similarity index 58% rename from packages/service/core/ai/llm/agentCall/index.ts rename to packages/service/core/ai/llm/agentLoop/index.ts index e0c3de5a4a..f7d3533dc4 100644 --- a/packages/service/core/ai/llm/agentCall/index.ts +++ b/packages/service/core/ai/llm/agentLoop/index.ts @@ -1,7 +1,7 @@ import type { ChatCompletionMessageParam, - ChatCompletionTool, ChatCompletionMessageToolCall, + ChatCompletionTool, CompletionFinishReason } from '@fastgpt/global/core/ai/llm/type'; import { ChatCompletionRequestMessageRoleEnum } from '@fastgpt/global/core/ai/constants'; @@ -19,14 +19,12 @@ import { filterEmptyAssistantMessages } from './utils'; import { countGptMessagesTokens } from '../../../../common/string/tiktoken/index'; import { formatModelChars2Points } from '../../../../support/wallet/usage/utils'; import { i18nT } from '../../../../../web/i18n/utils'; +import type { LLMModelItemType } from '@fastgpt/global/core/ai/model.schema'; type RunAgentCallProps = { maxRunAgentTimes: number; - compressTaskDescription?: string; - body: CreateLLMResponseProps['body'] & { tools: ChatCompletionTool[]; - temperature?: number; top_p?: number; stream?: boolean; @@ -38,25 +36,15 @@ type RunAgentCallProps = { childrenInteractiveParams?: ToolCallChildrenInteractive['params']; // LLM 压缩后回调 - onCompressContext?: (usage: { + onAfterCompressContext?: (usage: { modelName: string; inputTokens?: number; outputTokens?: number; totalPoints: number; seconds: number; }) => void; - // 工具压缩后回调 - onToolCompress?: (e: { - call: ChatCompletionMessageToolCall; - response: string; - usage: { - inputTokens: number; - outputTokens: number; - totalPoints: number; - }; - }) => void; // 处理交互工具 - handleInteractiveTool: (e: ToolCallChildrenInteractive['params']) => Promise<{ + onRunInteractiveTool: (e: ToolCallChildrenInteractive['params']) => Promise<{ response: string; assistantMessages: ChatCompletionMessageParam[]; usages: ChatNodeUsageType[]; @@ -64,7 +52,7 @@ type RunAgentCallProps = { stop?: boolean; }>; // 处理工具响应 - handleToolResponse: (e: { + onRunTool: (e: { call: ChatCompletionMessageToolCall; messages: ChatCompletionMessageParam[]; }) => Promise<{ @@ -95,39 +83,64 @@ type RunAgentResponse = { finish_reason: CompletionFinishReason | undefined; }; -/* - 一个循环进行工具调用的 LLM 请求封装。 +/** + * 上下文压缩，内部会判断是否需要压缩 + */ +export const onCompressContext = async ({ + isAborted, + requestMessages, + modelData, + userKey +}: { + isAborted: RunAgentCallProps['isAborted']; + requestMessages: ChatCompletionMessageParam[]; + modelData: LLMModelItemType; + userKey: RunAgentCallProps['userKey']; +}) => { + const compressStartTime = Date.now(); + const result = await compressRequestMessages({ + checkIsStopping: isAborted, + messages: requestMessages, + model: modelData, + userKey + }); + if (result.usage) { + return { + messages: result.messages, + usage: result.usage, + seconds: +((Date.now() - compressStartTime) / 1000).toFixed(2) + }; + } +}; - AssistantMessages 组成： - 1. 调用 AI 时生成的 messages - 2. tool 内部调用产生的 messages - 3. tool 响应的值，role=tool，content=tool response - - RequestMessages 为模型请求的消息，组成: - 1. 历史对话记录 - 2. 调用 AI 时生成的 messages - 3. tool 响应的值，role=tool，content=tool response - - memoryRequestMessages 为上一轮中断时，requestMessages 的内容 -*/ -export const runAgentCall = async ({ +/** + * 一个循环调用工具的 LLM 请求封装。 + * 每次循环会进行以下操作： + * 1. 压缩请求消息: 如果满足条件则压缩请求消息 + * 2. 请求 LLM + * 3. 调用工具（如有）： Call、Compress response + * 4. 检查是否循环结束 + */ +export const runAgentLoop = async ({ maxRunAgentTimes, - body: { model, messages, max_tokens, tools, ...body }, + body: { model, messages, max_tokens, ...body }, userKey, usagePush, isAborted, - onCompressContext, + onAfterCompressContext, childrenInteractiveParams, - handleInteractiveTool, - handleToolResponse, - onToolCompress, + onRunInteractiveTool, + + onToolCall, + onToolParam, + onAfterToolResponseCompress, + onAfterToolCall, + onRunTool, onReasoning, - onStreaming, - onToolCall, - onToolParam + onStreaming }: RunAgentCallProps): Promise => { const modelData = getLLMModel(model); @@ -174,7 +187,7 @@ export const runAgentCall = async ({ usages, interactive, stop - } = await handleInteractiveTool(childrenInteractiveParams); + } = await onRunInteractiveTool(childrenInteractiveParams); // 将 requestMessages 复原成上一轮中断时的内容，并附上 tool response requestMessages = childrenInteractiveParams.toolParams.memoryRequestMessages.map((item) => @@ -226,37 +239,42 @@ export const runAgentCall = async ({ // 正常完成该工具的响应，继续进行工具调用 } - // 自循环运行 + // Agent loop const requestIds: string[] = []; let consecutiveRequestToolTimes = 0; // 连续多次工具调用后会强制回答，避免模型自身死循环。 while (runTimes < maxRunAgentTimes) { - // TODO: 费用检测 + let stopAgentLoop = false; + // TODO: 费用检测 runTimes++; // 1. Compress request messages - const compressStartTime = Date.now(); - const result = await compressRequestMessages({ - checkIsStopping: isAborted, - messages: requestMessages, - model: modelData, - userKey - }); - requestMessages = result.messages; - if (result.usage) { - compressInputTokens += result.usage.inputTokens || 0; - compressOutputTokens += result.usage.outputTokens || 0; - childrenUsages.push(result.usage); - usagePush?.([result.usage]); - onCompressContext?.({ - modelName: modelData.name, - inputTokens: result.usage.inputTokens, - outputTokens: result.usage.outputTokens, - totalPoints: result.usage.totalPoints, - seconds: +((Date.now() - compressStartTime) / 1000).toFixed(2) + { + const compressResult = await onCompressContext({ + isAborted, + requestMessages, + modelData, + userKey }); + if (compressResult) { + requestMessages = compressResult.messages; + compressInputTokens += compressResult.usage.inputTokens || 0; + compressOutputTokens += compressResult.usage.outputTokens || 0; + childrenUsages.push(compressResult.usage); + usagePush?.([compressResult.usage]); + onAfterCompressContext?.({ + modelName: modelData.name, + inputTokens: compressResult.usage.inputTokens, + outputTokens: compressResult.usage.outputTokens, + totalPoints: compressResult.usage.totalPoints, + seconds: compressResult.seconds + }); + } } + // 拷贝一份 requestMessages 用于后续操作 + const cloneRequestMessages = requestMessages.slice(); + // 2. Request LLM let { requestId, @@ -276,7 +294,6 @@ export const runAgentCall = async ({ messages: requestMessages, tool_choice: consecutiveRequestToolTimes > 5 ? 'none' : 'auto', toolCallMode: modelData.toolChoice ? 'toolChoice' : 'prompt', - tools, parallel_tool_calls: true }, userKey, @@ -286,105 +303,65 @@ export const runAgentCall = async ({ onToolCall, onToolParam }); + // 请求后赋值操作 + { + finish_reason = finishReason; + requestError = error; + requestIds.push(requestId); - finish_reason = finishReason; - requestError = error; - requestIds.push(requestId); + if (requestError) { + break; + } + if (responseEmptyTip) { + return Promise.reject(responseEmptyTip); + } + if (toolCalls.length) { + consecutiveRequestToolTimes++; + } + if (answer) { + consecutiveRequestToolTimes = 0; + } - if (requestError) { - break; - } - if (responseEmptyTip) { - return Promise.reject(responseEmptyTip); - } - if (toolCalls.length) { - consecutiveRequestToolTimes++; - } - if (answer) { - consecutiveRequestToolTimes = 0; - } - - // Record usage - inputTokens += usage.inputTokens; - outputTokens += usage.outputTokens; - const totalPoints = userKey - ? 0 - : formatModelChars2Points({ - model: modelData, + // Record usage + inputTokens += usage.inputTokens; + outputTokens += usage.outputTokens; + const totalPoints = userKey + ? 0 + : formatModelChars2Points({ + model: modelData, + inputTokens: usage.inputTokens, + outputTokens: usage.outputTokens + }).totalPoints; + llmTotalPoints += totalPoints; // 每次调用单独计价后累加，保证梯度计费正确 + usagePush?.([ + { + moduleName: i18nT('account_usage:agent_call'), + model: modelData.name, + totalPoints, inputTokens: usage.inputTokens, outputTokens: usage.outputTokens - }).totalPoints; - llmTotalPoints += totalPoints; // 每次调用单独计价后累加，保证梯度计费正确 + } + ]); - usagePush?.([ - { - moduleName: i18nT('account_usage:agent_call'), - model: modelData.name, - totalPoints, - inputTokens: usage.inputTokens, - outputTokens: usage.outputTokens + // 推送 AI 生成后的 assistantMessages + if (llmAssistantMessage) { + assistantMessages.push(llmAssistantMessage); + requestMessages.push(llmAssistantMessage); } - ]); - - // 3. 更新 messages - const cloneRequestMessages = requestMessages.slice(); - // 推送 AI 生成后的 assistantMessages - if (llmAssistantMessage) { - assistantMessages.push(llmAssistantMessage); - requestMessages.push(llmAssistantMessage); } - // 4. Call tools - let toolCallStep = false; + // 3. Call tools for await (const tool of toolCalls) { const { response, assistantMessages: toolAssistantMessages, usages: toolUsages, interactive, - stop - } = await handleToolResponse({ + stop: stopLoop + } = await onRunTool({ call: tool, messages: cloneRequestMessages }); - childrenUsages.push(...toolUsages); - usagePush(toolUsages); - - // 5. Add tool response to messages - // 获取当前 messages 的 token 数，用于动态调整 tool response 的压缩阈值（防止下一个工具直接打爆上下文） - const currentMessagesTokens = await countGptMessagesTokens(requestMessages); - - const { compressed: compressed_context, usage: compressionUsage } = - await compressToolResponse({ - response, - model: modelData, - currentMessagesTokens, - toolLength: toolCalls.length, - reservedTokens: 8000, // 预留 8k tokens 给输出 - userKey - }); - if (compressionUsage) { - childrenUsages.push(compressionUsage); - usagePush?.([compressionUsage]); - onToolCompress?.({ - call: tool, - response: compressed_context, - usage: { - inputTokens: compressionUsage.inputTokens!, - outputTokens: compressionUsage.outputTokens!, - totalPoints: compressionUsage.totalPoints! - } - }); - } - - const toolMessage: ChatCompletionMessageParam = { - tool_call_id: tool.id, - role: ChatCompletionRequestMessageRoleEnum.Tool, - content: compressed_context - }; - assistantMessages.push(toolMessage); - requestMessages.push(toolMessage); - assistantMessages.push(...filterEmptyAssistantMessages(toolAssistantMessages)); // 因为 toolAssistantMessages 也需要记录成 AI 响应，所以这里需要推送。 if (interactive) { interactiveResponse = { @@ -398,11 +375,68 @@ export const runAgentCall = async ({ } }; } - if (stop) { - toolCallStep = true; + if (stopLoop) { + stopAgentLoop = true; + } + + // Push usages + { + childrenUsages.push(...toolUsages); + usagePush(toolUsages); + } + + // Compress tool response + const toolFinalResponse = await (async () => { + const currentMessagesTokens = await countGptMessagesTokens(requestMessages); + const { compressed: compressed_context, usage: compressionUsage } = + await compressToolResponse({ + response, + model: modelData, + currentMessagesTokens, + toolLength: toolCalls.length, + reservedTokens: 8000, // 预留 8k tokens 给输出 + userKey + }); + if (compressionUsage) { + childrenUsages.push(compressionUsage); + usagePush([compressionUsage]); + onAfterToolResponseCompress?.({ + call: tool, + response: compressed_context, + usage: { + inputTokens: compressionUsage.inputTokens!, + outputTokens: compressionUsage.outputTokens!, + totalPoints: compressionUsage.totalPoints! + } + }); + } + + return compressed_context; + })(); + + onAfterToolCall?.({ success: true, call: tool, response: toolFinalResponse }); + + // Push messages + { + const toolMessage: ChatCompletionMessageParam = { + tool_call_id: tool.id, + role: ChatCompletionRequestMessageRoleEnum.Tool, + content: toolFinalResponse + }; + assistantMessages.push(toolMessage); + requestMessages.push(toolMessage); + assistantMessages.push(...filterEmptyAssistantMessages(toolAssistantMessages)); // 因为 toolAssistantMessages 也需要记录成 AI 响应，所以这里需要推送。 } } - if (toolCalls.length === 0 || !!interactiveResponse || toolCallStep || isAborted?.()) { + + /** + * 检查是否 loop 结束 + * 1. 没有工具调用 + * 2. 有交互工具 + * 3. 特殊的工具，要求结束当前 loop + * 4. 用户主动暂停 + */ + if (toolCalls.length === 0 || !!interactiveResponse || stopAgentLoop || isAborted?.()) { break; } } diff --git a/packages/service/core/ai/llm/agentCall/prompt.ts b/packages/service/core/ai/llm/agentLoop/prompt.ts similarity index 100% rename from packages/service/core/ai/llm/agentCall/prompt.ts rename to packages/service/core/ai/llm/agentLoop/prompt.ts diff --git a/packages/service/core/ai/llm/agentCall/utils.ts b/packages/service/core/ai/llm/agentLoop/utils.ts similarity index 100% rename from packages/service/core/ai/llm/agentCall/utils.ts rename to packages/service/core/ai/llm/agentLoop/utils.ts diff --git a/packages/service/core/ai/llm/request.ts b/packages/service/core/ai/llm/request.ts index 5a9a4dae6d..82594ef1af 100644 --- a/packages/service/core/ai/llm/request.ts +++ b/packages/service/core/ai/llm/request.ts @@ -31,6 +31,7 @@ import { getErrText } from '@fastgpt/global/common/error/utils'; import json5 from 'json5'; import { getLogger, LogCategories } from '../../../common/logger'; import { saveLLMRequestRecord } from '../record/controller'; +import type { ToolCallEventType } from './toolCall/type'; const getRequestId = () => { return customNanoid('abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ1234567890_-', 16); @@ -38,11 +39,9 @@ const getRequestId = () => { const logger = getLogger(LogCategories.MODULE.AI.LLM); -export type ResponseEvents = { +export type ResponseEvents = ToolCallEventType & { onStreaming?: (e: { text: string }) => void; onReasoning?: (e: { text: string }) => void; - onToolCall?: (e: { call: ChatCompletionMessageToolCall }) => void; - onToolParam?: (e: { tool: ChatCompletionMessageToolCall; params: string }) => void; }; export type CreateLLMResponseProps< @@ -459,7 +458,7 @@ export const createStreamResponse = async ({ if (currentTool && arg) { currentTool.function.arguments += arg; - onToolParam?.({ tool: currentTool, params: arg }); + onToolParam?.({ call: currentTool, argsDelta: arg }); } } }); diff --git a/packages/service/core/ai/llm/toolCall/type.ts b/packages/service/core/ai/llm/toolCall/type.ts new file mode 100644 index 0000000000..e20e8e2a8a --- /dev/null +++ b/packages/service/core/ai/llm/toolCall/type.ts @@ -0,0 +1,24 @@ +import type { ChatCompletionMessageToolCall } from '@fastgpt/global/core/ai/llm/type'; + +export type ToolCallEventType = { + onToolCall?: (e: { call: ChatCompletionMessageToolCall }) => void; + onToolParam?: (e: { call: ChatCompletionMessageToolCall; argsDelta: string }) => void; + // 工具执行完成后的生命周期钩子（含未找到 / parseParams 失败 / execute 抛错的兜底） + onAfterToolCall?: (e: { + success: boolean; + call: ChatCompletionMessageToolCall; + response?: string; + errorMessage?: string; + }) => void; + + // 工具压缩后回调 + onAfterToolResponseCompress?: (e: { + call: ChatCompletionMessageToolCall; + response: string; + usage: { + inputTokens: number; + outputTokens: number; + totalPoints: number; + }; + }) => void; +}; diff --git a/packages/service/core/ai/sandbox/toolCall.ts b/packages/service/core/ai/sandbox/toolCall.ts deleted file mode 100644 index ae3b65d6c1..0000000000 --- a/packages/service/core/ai/sandbox/toolCall.ts +++ /dev/null @@ -1,133 +0,0 @@ -import { - SANDBOX_TOOL_NAME, - SANDBOX_GET_FILE_URL_TOOL_NAME, - SandboxShellToolSchema, - SandboxGetFileUrlToolSchema -} from '@fastgpt/global/core/ai/sandbox/constants'; -import { getErrText } from '@fastgpt/global/common/error/utils'; -import { parseJsonArgs } from '../utils'; -import { getSandboxClient } from './controller'; -import { getS3ChatSource } from '../../../common/s3/sources/chat'; -import path from 'path'; -import { jwtSignS3ObjectKey } from '../../../common/s3/utils'; -import { addHours } from 'date-fns'; -import { Readable } from 'stream'; -import { getLogger } from '@fastgpt-sdk/otel'; -import { LogCategories } from '../../../common/logger'; - -type SandboxToolCallParams = { - toolName: string; - rawArgs: string; - appId: string; - userId: string; - chatId: string; -}; - -export type SandboxToolCallResult = { - input: Record; - response: string; - durationSeconds: number; -}; - -/** - * 纯沙盒工具执行层。 - * 只负责调用沙盒、上传 S3 等底层操作，返回统一的执行结果，不绑定任何业务响应格式。 - */ -export const callSandboxTool = async ({ - toolName, - rawArgs, - appId, - userId, - chatId -}: SandboxToolCallParams): Promise => { - const startTime = Date.now(); - const getDuration = () => +((Date.now() - startTime) / 1000).toFixed(2); - - if (toolName === SANDBOX_TOOL_NAME) { - const parsed = SandboxShellToolSchema.safeParse(parseJsonArgs(rawArgs)); - if (!parsed.success) { - return { input: {}, response: parsed.error.message, durationSeconds: getDuration() }; - } - const { command, timeout } = parsed.data; - - try { - const instance = await getSandboxClient({ appId, userId, chatId }); - const result = await instance.exec(command, timeout); - - return { - input: { command, timeout }, - response: JSON.stringify({ - stdout: result.stdout, - stderr: result.stderr, - exitCode: result.exitCode - }), - durationSeconds: getDuration() - }; - } catch (error: any) { - getLogger(LogCategories.MODULE.AI.AGENT).error('[Sandbox Shell] Execution failed', { error }); - return { - input: { command, timeout }, - response: getErrText(error), - durationSeconds: getDuration() - }; - } - } - - if (toolName === SANDBOX_GET_FILE_URL_TOOL_NAME) { - const parsed = SandboxGetFileUrlToolSchema.safeParse(parseJsonArgs(rawArgs)); - if (!parsed.success) { - return { input: {}, response: parsed.error.message, durationSeconds: getDuration() }; - } - - const { paths } = parsed.data; - - try { - const instance = await getSandboxClient({ appId, userId, chatId }); - - const result = await Promise.all( - paths.map(async (url) => { - const filename = path.basename(url); - const stream = instance.provider.readFileStream(url); - const readable = Readable.from(stream); // AsyncIterable → Readable - - const chatBucket = getS3ChatSource(); - const expiredTime = addHours(new Date(), 2); - const { key } = await chatBucket.uploadChatFile({ - appId, - chatId, - uId: userId, - filename, - body: readable, - expiredTime: expiredTime - }); - const fileUrl = jwtSignS3ObjectKey(key, expiredTime); - - return { - fileUrl, - filename - }; - }) - ); - - return { - input: { paths }, - response: JSON.stringify(result), - durationSeconds: getDuration() - }; - } catch (error) { - getLogger(LogCategories.MODULE.AI.AGENT).error('[Sandbox Get File URL] failed', { error }); - - return { - input: { paths }, - response: `Get file URL error: ${getErrText(error)}`, - durationSeconds: getDuration() - }; - } - } - - return { - input: {}, - response: `Unknown sandbox tool: ${toolName}`, - durationSeconds: getDuration() - }; -}; diff --git a/packages/service/core/ai/sandbox/toolCall/getFileUrl.tool.ts b/packages/service/core/ai/sandbox/toolCall/getFileUrl.tool.ts new file mode 100644 index 0000000000..dd9ebd537c --- /dev/null +++ b/packages/service/core/ai/sandbox/toolCall/getFileUrl.tool.ts @@ -0,0 +1,45 @@ +import z from 'zod'; +import path from 'path'; +import { Readable } from 'stream'; +import { addHours } from 'date-fns'; +import { defineTool } from './type'; +import { getS3ChatSource } from '../../../../common/s3/sources/chat'; +import { jwtSignS3ObjectKey } from '../../../../common/s3/utils'; +import { SANDBOX_GET_FILE_URL_TOOL_NAME } from '@fastgpt/global/core/ai/sandbox/constants'; + +const SandboxGetFileUrlToolSchema = z.object({ + paths: z.array(z.string()) +}); + +export const sandboxGetFileUrlTool = defineTool({ + zodSchema: SandboxGetFileUrlToolSchema, + execute: async ({ appId, userId, chatId, sandboxInstance, params }) => { + const result = await Promise.all( + params.paths.map(async (filePath) => { + const filename = path.basename(filePath); + const stream = sandboxInstance.provider.readFileStream(filePath); + const readable = Readable.from(stream); + + const chatBucket = getS3ChatSource(); + const expiredTime = addHours(new Date(), 2); + const { key } = await chatBucket.uploadChatFile({ + appId, + chatId, + uId: userId, + filename, + body: readable, + expiredTime + }); + const fileUrl = jwtSignS3ObjectKey(key, expiredTime); + + return { fileUrl, filename }; + }) + ); + + return { response: JSON.stringify(result) }; + } +}); + +export const toolMap = { + [SANDBOX_GET_FILE_URL_TOOL_NAME]: sandboxGetFileUrlTool +}; diff --git a/packages/service/core/ai/sandbox/toolCall/index.ts b/packages/service/core/ai/sandbox/toolCall/index.ts new file mode 100644 index 0000000000..0580adcb0d --- /dev/null +++ b/packages/service/core/ai/sandbox/toolCall/index.ts @@ -0,0 +1,86 @@ +import { sandboxToolMap } from '@fastgpt/global/core/ai/sandbox/constants'; +import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; +import type { localeType } from '@fastgpt/global/common/i18n/type'; +import { LangEnum } from '@fastgpt/global/common/i18n/type'; +import { toolMap as getFileUrlToolMap } from './getFileUrl.tool'; +import { toolMap as shellToolMap } from './shell.tool'; +import { getSandboxClient } from '../controller'; +import { parseJsonArgs } from '../../utils'; + +const ToolMap = { + ...getFileUrlToolMap, + ...shellToolMap +}; + +export type SandboxToolCallResult = { + success: boolean; + input: Record; + response: string; + durationSeconds: number; +}; + +export const runSandboxTools = async ({ + appId, + userId, + chatId, + toolName, + args +}: { + appId: string; + userId: string; + chatId: string; + toolName: string; + args: string; +}): Promise => { + const startTime = Date.now(); + const getDuration = () => +((Date.now() - startTime) / 1000).toFixed(2); + + const tool = ToolMap[toolName as keyof typeof ToolMap]; + + if (!tool) { + return { + success: false, + input: {}, + response: `Unknown sandbox tool: ${toolName}`, + durationSeconds: getDuration() + }; + } + + // Parse args + const parsedArgs = tool.zodSchema.safeParse(parseJsonArgs(args)); + if (!parsedArgs.success) { + return { + success: false, + input: {}, + response: parsedArgs.error.message, + durationSeconds: getDuration() + }; + } + + const instance = await getSandboxClient({ appId, userId, chatId }); + const result = await tool.execute({ + appId, + userId, + chatId, + sandboxInstance: instance, + params: parsedArgs.data as any + }); + + return { + success: true, + input: parsedArgs.data, + response: result.response, + durationSeconds: getDuration() + }; +}; + +export const getSandboxToolInfo = (name: string, lang: localeType = LangEnum.en) => { + if (name in sandboxToolMap) { + const info = sandboxToolMap[name]; + return { + name: parseI18nString(info.name, lang), + avatar: info.avatar, + toolDescription: info.toolDescription + }; + } +}; diff --git a/packages/service/core/ai/sandbox/toolCall/shell.tool.ts b/packages/service/core/ai/sandbox/toolCall/shell.tool.ts new file mode 100644 index 0000000000..e9ceb9b3bb --- /dev/null +++ b/packages/service/core/ai/sandbox/toolCall/shell.tool.ts @@ -0,0 +1,26 @@ +import z from 'zod'; +import { defineTool } from './type'; +import { SANDBOX_TOOL_NAME } from '@fastgpt/global/core/ai/sandbox/constants'; + +const SandboxShellToolSchema = z.object({ + command: z.string(), + timeout: z.number().optional() +}); + +export const sandboxShellTool = defineTool({ + zodSchema: SandboxShellToolSchema, + execute: async ({ sandboxInstance, params }) => { + const result = await sandboxInstance.exec(params.command, params.timeout); + return { + response: JSON.stringify({ + stdout: result.stdout, + stderr: result.stderr, + exitCode: result.exitCode + }) + }; + } +}); + +export const toolMap = { + [SANDBOX_TOOL_NAME]: sandboxShellTool +}; diff --git a/packages/service/core/ai/sandbox/toolCall/type.ts b/packages/service/core/ai/sandbox/toolCall/type.ts new file mode 100644 index 0000000000..751ef4f035 --- /dev/null +++ b/packages/service/core/ai/sandbox/toolCall/type.ts @@ -0,0 +1,18 @@ +import type { z } from 'zod'; +import type { SandboxClient } from '../controller'; + +type ToolExecuteContext

= { + appId: string; + userId: string; + chatId: string; + sandboxInstance: SandboxClient; + params: P; +}; +// 声明式工具定义 +export type ToolDefinition = { + zodSchema: S; + execute: (ctx: ToolExecuteContext>) => Promise<{ response: string }>; +}; + +export const defineTool = (def: ToolDefinition): ToolDefinition => + def; diff --git a/packages/service/core/chat/HelperBot/dispatch/topAgent/utils.ts b/packages/service/core/chat/HelperBot/dispatch/topAgent/utils.ts index 0c00efda8c..aaac31655f 100644 --- a/packages/service/core/chat/HelperBot/dispatch/topAgent/utils.ts +++ b/packages/service/core/chat/HelperBot/dispatch/topAgent/utils.ts @@ -8,6 +8,7 @@ import { MongoResourcePermission } from '../../../../../support/permission/schem import { PerResourceTypeEnum } from '@fastgpt/global/support/permission/constant'; import { getGroupsByTmbId } from '../../../../../support/permission/memberGroup/controllers'; import { getOrgIdSetWithParentByTmbId } from '../../../../../support/permission/org/controllers'; +import { SANDBOX_TOOL_NAME } from '@fastgpt/global/core/ai/sandbox/constants'; const getAccessibleDatasets = async ({ teamId, tmbId }: { teamId: string; tmbId: string }) => { const [roleList, myGroupMap, myOrgSet] = await Promise.all([ @@ -110,7 +111,7 @@ ${dataset} }) ]); - const builtinTools = [SubAppIds.fileRead, SubAppIds.sandboxTool].map((id) => { + const builtinTools = [SubAppIds.fileRead, SANDBOX_TOOL_NAME].map((id) => { const info = systemSubInfo[id]; return `- **${id}** [工具]: ${parseI18nString(info.name, lang)} - ${info.toolDescription}`; }); diff --git a/packages/service/core/dataset/search/controller.ts b/packages/service/core/dataset/search/controller.ts index 45a5eb8170..b0cea1788f 100644 --- a/packages/service/core/dataset/search/controller.ts +++ b/packages/service/core/dataset/search/controller.ts @@ -52,7 +52,7 @@ export type SearchDatasetDataProps = { [NodeInputKeyEnum.datasetSimilarity]?: number; // min distance [NodeInputKeyEnum.datasetMaxTokens]: number; // max Token limit - [NodeInputKeyEnum.datasetSearchMode]?: `${DatasetSearchModeEnum}`; + [NodeInputKeyEnum.datasetSearchMode]?: DatasetSearchModeEnum; [NodeInputKeyEnum.datasetSearchEmbeddingWeight]?: number; [NodeInputKeyEnum.datasetSearchUsingReRank]?: boolean; diff --git a/packages/service/core/workflow/dispatch/ai/agent/index.ts b/packages/service/core/workflow/dispatch/ai/agent/index.ts index 0f54bab15a..6f17ab9435 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/index.ts @@ -22,8 +22,7 @@ import { } from '@fastgpt/global/core/chat/adapt'; import { getPlanCallResponseText } from '@fastgpt/global/core/chat/utils'; import { filterMemoryMessages } from '../utils'; -import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; -import { systemSubInfo } from '@fastgpt/global/core/workflow/node/agent/constants'; +import { getSystemToolInfo } from '@fastgpt/global/core/workflow/node/agent/constants'; import type { DispatchPlanAgentResponse } from './sub/plan'; import { dispatchPlanAgent } from './sub/plan'; @@ -272,13 +271,12 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise }; } - const systemToolNode = systemSubInfo[id] || systemSubInfo[formatId]; - const systemDisplayName = parseI18nString(systemToolNode?.name, lang); + const systemToolNode = getSystemToolInfo(id, lang) || getSystemToolInfo(formatId, lang); return { - name: systemDisplayName || '', + name: systemToolNode?.name || '', avatar: systemToolNode?.avatar || '', - toolDescription: systemToolNode?.toolDescription || systemDisplayName || '' + toolDescription: systemToolNode?.toolDescription || systemToolNode?.name || '' }; }; const getSubApp = (id: string) => { diff --git a/packages/service/core/workflow/dispatch/ai/agent/master/call.ts b/packages/service/core/workflow/dispatch/ai/agent/master/call.ts index 18a5f06a22..1255f9f05f 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/master/call.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/master/call.ts @@ -2,27 +2,21 @@ import type { ChatCompletionMessageParam, ChatCompletionTool } from '@fastgpt/global/core/ai/llm/type'; -import { runAgentCall } from '../../../../../ai/llm/agentCall'; +import { runAgentLoop } from '../../../../../ai/llm/agentLoop'; import { chats2GPTMessages, runtimePrompt2ChatsValue } from '@fastgpt/global/core/chat/adapt'; import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants'; -import { addFilePrompt2Input, ReadFileToolSchema } from '../sub/file/utils'; +import { addFilePrompt2Input } from '../sub/file/utils'; import { type AgentStepItemType } from '@fastgpt/global/core/ai/agent/type'; import type { GetSubAppInfoFnType, SubAppRuntimeType } from '../type'; import { SseResponseEventEnum } from '@fastgpt/global/core/workflow/runtime/constants'; import { textAdaptGptResponse } from '@fastgpt/global/core/workflow/runtime/utils'; import { SubAppIds } from '@fastgpt/global/core/workflow/node/agent/constants'; -import { parseJsonArgs } from '../../../../../ai/utils'; -import { dispatchFileRead } from '../sub/file'; -import { dispatchTool } from '../sub/tool'; import { getErrText } from '@fastgpt/global/common/error/utils'; -import { DatasetSearchToolSchema } from '../sub/dataset/utils'; -import { dispatchAgentDatasetSearch } from '../sub/dataset'; import type { DispatchAgentModuleProps } from '..'; import { getLLMModel } from '../../../../../ai/model'; import { getStepCallQuery, getStepDependon } from './dependon'; import { getOneStepResponseSummary } from './responseSummary'; import type { DispatchPlanAgentResponse } from '../sub/plan'; -import { dispatchPlanAgent } from '../sub/plan'; import type { WorkflowResponseItemType } from '../../../type'; import type { AIChatItemValueItemType, @@ -32,18 +26,9 @@ import { getNanoid } from '@fastgpt/global/common/string/tools'; import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; import { i18nT } from '../../../../../../../web/i18n/utils'; import { getMasterSystemPrompt } from './prompt'; -import { PlanAgentParamsSchema } from '../sub/plan/constants'; import { filterMemoryMessages } from '../../utils'; -import { dispatchApp, dispatchPlugin } from '../sub/app'; -import { getLogger, LogCategories } from '../../../../../../common/logger'; -import { - SandboxShellToolSchema, - SANDBOX_TOOL_NAME, - SANDBOX_GET_FILE_URL_TOOL_NAME, - SandboxGetFileUrlToolSchema -} from '@fastgpt/global/core/ai/sandbox/constants'; -import { dispatchSandboxShell, dispatchSandboxGetFileUrl } from '../sub/sandbox'; import type { CapabilityToolCallHandlerType } from '../capability/type'; +import { getExecuteTool } from '../utils'; type Response = { stepResponse?: { @@ -215,6 +200,16 @@ export const masterCall = async ({ let planResult: DispatchPlanAgentResponse | undefined; + const executeTool = getExecuteTool({ + ...props, + streamResponseFn: stepStreamResponse, + getSubAppInfo, + getSubApp, + completionTools, + filesMap, + capabilityToolCallHandler + }); + const { model: agentModel, assistantMessages, @@ -226,7 +221,7 @@ export const masterCall = async ({ finish_reason, requestIds, error: agentError - } = await runAgentCall({ + } = await runAgentLoop({ maxRunAgentTimes: 100, body: { messages: requestMessages, @@ -280,18 +275,18 @@ export const masterCall = async ({ } }); }, - onToolParam({ tool, params }) { + onToolParam({ call, argsDelta }) { stepStreamResponse?.({ - id: tool.id, + id: call.id, event: SseResponseEventEnum.toolParams, data: { tool: { - params + params: argsDelta } } }); }, - onCompressContext: ({ modelName, inputTokens, outputTokens, totalPoints, seconds }) => { + onAfterCompressContext: ({ modelName, inputTokens, outputTokens, totalPoints, seconds }) => { childrenResponses.push({ nodeId: getNanoid(6), id: getNanoid(6), @@ -305,369 +300,41 @@ export const masterCall = async ({ runningTime: seconds }); }, - handleToolResponse: async ({ call, messages }) => { + onRunTool: async ({ call }) => { const toolId = call.function.name; const callId = call.id; const { response, usages = [], - stop = false - } = await (async () => { - try { - if (toolId === SubAppIds.fileRead) { - const toolParams = ReadFileToolSchema.safeParse(parseJsonArgs(call.function.arguments)); - if (!toolParams.success) { - return { - response: toolParams.error.message, - usages: [] - }; - } - const params = toolParams.data; + stop = false, + nodeResponse, + planResult: execPlanResult, + capabilityAssistantResponses: execCapabilityAssistantResponses + } = await executeTool({ callId, toolId, args: call.function.arguments }); - const files = params.file_indexes.map((index) => ({ - index, - url: filesMap[index] - })); - const result = await dispatchFileRead({ - files, - teamId: runningUserInfo.teamId, - tmbId: runningUserInfo.tmbId, - customPdfParse: chatConfig?.fileSelectConfig?.customPdfParse, - model, - userKey: externalProvider.openaiAccount - }); - - if (result.nodeResponse) { - childrenResponses.push(result.nodeResponse); - } - return { - response: result.response, - usages: result.usages - }; - } - if (toolId === SubAppIds.datasetSearch) { - const toolParams = DatasetSearchToolSchema.safeParse( - parseJsonArgs(call.function.arguments) - ); - if (!toolParams.success) { - return { - response: toolParams.error.message, - usages: [] - }; - } - - if (!datasetParams || datasetParams.datasets.length === 0) { - return { - response: 'No dataset selected', - usages: [] - }; - } - - const params = toolParams.data; - - const result = await dispatchAgentDatasetSearch({ - query: params.query, - config: { - datasets: datasetParams.datasets, - similarity: datasetParams.similarity || 0.4, - maxTokens: datasetParams.limit || 5000, - searchMode: datasetParams.searchMode, - embeddingWeight: datasetParams.embeddingWeight, - usingReRank: datasetParams.usingReRank ?? false, - rerankModel: datasetParams.rerankModel, - rerankWeight: datasetParams.rerankWeight || 0.5, - usingExtensionQuery: datasetParams.datasetSearchUsingExtensionQuery ?? false, - extensionModel: datasetParams.datasetSearchExtensionModel, - extensionBg: datasetParams.datasetSearchExtensionBg - }, - teamId: runningUserInfo.teamId, - tmbId: runningUserInfo.tmbId, - llmModel: model - }); - - if (result.nodeResponse) { - childrenResponses.push(result.nodeResponse); - } - - return { - response: result.response, - usages: result.usages - }; - } - if (toolId === SANDBOX_TOOL_NAME) { - const toolParams = SandboxShellToolSchema.safeParse( - parseJsonArgs(call.function.arguments) - ); - if (!toolParams.success) { - return { - response: toolParams.error.message, - usages: [] - }; - } - - const result = await dispatchSandboxShell({ - command: toolParams.data.command, - timeout: toolParams.data.timeout, - appId: runningAppInfo.id, - userId: props.uid, - chatId, - lang: props.lang - }); - - childrenResponses.push(result.nodeResponse); - - return { - response: result.response, - usages: result.usages - }; - } - if (toolId === SANDBOX_GET_FILE_URL_TOOL_NAME) { - const toolParams = SandboxGetFileUrlToolSchema.safeParse( - parseJsonArgs(call.function.arguments) - ); - if (!toolParams.success) { - return { - response: toolParams.error.message, - usages: [] - }; - } - - const result = await dispatchSandboxGetFileUrl({ - paths: toolParams.data.paths, - appId: runningAppInfo.id, - userId: props.uid, - chatId, - lang: props.lang - }); - - childrenResponses.push(result.nodeResponse); - - return { - response: result.response, - usages: result.usages - }; - } - if (toolId === SubAppIds.plan) { - try { - const toolArgs = await PlanAgentParamsSchema.safeParseAsync( - parseJsonArgs(call.function.arguments) - ); - - if (!toolArgs.success) { - return { - response: 'Tool arguments is not valid', - usages: [] - }; - } - - // plan: 1,3 场景 - planResult = await dispatchPlanAgent({ - checkIsStopping, - completionTools, - getSubAppInfo, - systemPrompt, - model, - stream, - mode: 'initial', - ...toolArgs.data, - planId: call.id - }); - - return { - response: '', - stop: true, - usages: [] // 外部会单独对 plan 计费 - }; - } catch (error) { - getLogger(LogCategories.MODULE.AI.AGENT).error('dispatchPlanAgent error', { error }); - return { - response: `Plan error: ${getErrText(error)}`, - stop: false - }; - } - } - - // TODO: 所有内置工具，合并成一个 function - // Capability tools (e.g. sandbox skills) - const capResult = await capabilityToolCallHandler?.( - toolId, - call.function.arguments ?? '', - callId - ); - if (capResult != null) { - if (capResult.assistantResponses?.length) { - capabilityAssistantResponses.push(...capResult.assistantResponses); - } - const subInfo = getSubAppInfo(toolId); - childrenResponses.push({ - nodeId: callId, - id: callId, - moduleType: FlowNodeTypeEnum.tool, - moduleName: subInfo.name, - moduleLogo: subInfo.avatar, - toolInput: parseJsonArgs(call.function.arguments), - toolRes: capResult.response - }); - return { - response: capResult.response, - usages: capResult.usages || [] - }; - } - - // User Sub App - const tool = getSubApp(toolId); - if (!tool) { - return { - response: `Can't find the tool ${toolId}`, - usages: [] - }; - } - const toolCallParams = parseJsonArgs(call.function.arguments); - - if (call.function.arguments && !toolCallParams) { - return { - response: 'Params is not object', - usages: [] - }; - } - - // Get params - const requestParams = { - ...tool.params, - ...toolCallParams - }; - // Remove sensitive data - - if (tool.type === 'tool') { - const { response, usages, runningTime, toolParams, result } = await dispatchTool({ - tool: { - name: tool.name, - version: tool.version, - toolConfig: tool.toolConfig - }, - params: requestParams, - runningUserInfo, - runningAppInfo, - chatId, - uid, - variables, - workflowStreamResponse: stepStreamResponse - }); - - childrenResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.tool, - moduleName: tool.name, - moduleLogo: tool.avatar, - toolInput: toolParams, - toolRes: result || response, - totalPoints: usages?.reduce((sum, item) => sum + item.totalPoints, 0) - }); - return { - response, - usages - }; - } else if (tool.type === 'workflow') { - const { userChatInput, ...params } = requestParams; - - const { response, runningTime, usages } = await dispatchApp({ - appId: tool.id, - userChatInput: userChatInput, - customAppVariables: params, - checkIsStopping, - lang: props.lang, - requestOrigin: props.requestOrigin, - mode: props.mode, - timezone: props.timezone, - externalProvider: props.externalProvider, - runningAppInfo: props.runningAppInfo, - runningUserInfo: props.runningUserInfo, - retainDatasetCite: props.retainDatasetCite, - maxRunTimes: props.maxRunTimes, - workflowDispatchDeep: props.workflowDispatchDeep, - variables: props.variables - }); - - childrenResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.appModule, - moduleName: tool.name, - moduleLogo: tool.avatar, - toolInput: requestParams, - toolRes: response, - totalPoints: usages?.reduce((sum, item) => sum + item.totalPoints, 0) - }); - - return { - response, - usages, - runningTime - }; - } else if (tool.type === 'toolWorkflow') { - const { response, result, runningTime, usages } = await dispatchPlugin({ - appId: tool.id, - userChatInput: '', - customAppVariables: requestParams, - checkIsStopping, - lang: props.lang, - requestOrigin: props.requestOrigin, - mode: props.mode, - timezone: props.timezone, - externalProvider: props.externalProvider, - runningAppInfo: props.runningAppInfo, - runningUserInfo: props.runningUserInfo, - retainDatasetCite: props.retainDatasetCite, - maxRunTimes: props.maxRunTimes, - workflowDispatchDeep: props.workflowDispatchDeep, - variables: props.variables - }); - - childrenResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.pluginModule, - moduleName: tool.name, - moduleLogo: tool.avatar, - toolInput: requestParams, - toolRes: result, - totalPoints: usages?.reduce((sum, item) => sum + item.totalPoints, 0) - }); - - return { - response, - usages, - runningTime - }; - } else { - return { - response: 'Invalid tool type', - usages: [] - }; - } - } catch (error) { - return { - response: `Tool error: ${getErrText(error)}`, - usages: [] - }; + // 赋值操作 + { + if (execPlanResult) { + planResult = execPlanResult; } - })(); - - // Push stream response - stepStreamResponse?.({ - id: call.id, - event: SseResponseEventEnum.toolResponse, - data: { - tool: { - response - } + if (execCapabilityAssistantResponses) { + capabilityAssistantResponses.push(...execCapabilityAssistantResponses); } - }); + if (nodeResponse) { + childrenResponses.push(nodeResponse); + } + // Push stream response + stepStreamResponse?.({ + id: call.id, + event: SseResponseEventEnum.toolResponse, + data: { + tool: { + response + } + } + }); + } return { response, @@ -676,7 +343,7 @@ export const masterCall = async ({ stop }; }, - onToolCompress: ({ call, response, usage }) => { + onAfterToolResponseCompress: ({ call, response, usage }) => { const callId = call.id; const nodeResponse = childrenResponses.findLast((item) => item.id === callId); if (nodeResponse) { @@ -687,7 +354,7 @@ export const masterCall = async ({ nodeResponse.toolRes = response; } }, - handleInteractiveTool: async ({ toolParams }) => { + onRunInteractiveTool: async ({}) => { return { response: 'Interactive tool not supported', assistantMessages: [], // TODO @@ -696,7 +363,7 @@ export const masterCall = async ({ } }); - // llmTotalPoints 是 runAgentCall 内每次 LLM 调用单独计价后的累计值，保证梯度计费正确 + // llmTotalPoints 是 runAgentLoop 内每次 LLM 调用单独计价后的累计值，保证梯度计费正确 const llmUsage = { modelName: getLLMModel(agentModel).name, totalPoints: llmTotalPoints diff --git a/packages/service/core/workflow/dispatch/ai/agent/piAgent/index.ts b/packages/service/core/workflow/dispatch/ai/agent/piAgent/index.ts index 72cfaf6392..5c789b503e 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/piAgent/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/piAgent/index.ts @@ -15,12 +15,13 @@ import { formatFileInput } from '../sub/file/utils'; import { normalizeSkillIds } from '@fastgpt/global/core/app/formEdit/type'; import { systemSubInfo } from '@fastgpt/global/core/workflow/node/agent/constants'; import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; +import type { ToolDispatchContext } from '../utils'; import { getSubapps } from '../utils'; import { createCapabilityToolCallHandler, type AgentCapability } from '../capability/type'; import { createSandboxSkillsCapability } from '../capability/sandboxSkills'; import { textAdaptGptResponse } from '@fastgpt/global/core/workflow/runtime/utils'; import { buildPiModel, getModelApiKey } from './modelBridge'; -import { buildAgentTools, type ToolDispatchContext } from './toolAdapter'; +import { buildAgentTools } from './toolAdapter'; import { getLogger, LogCategories } from '../../../../../../common/logger'; import { env } from '../../../../../../env'; import type { DispatchAgentModuleProps } from '..'; @@ -166,35 +167,19 @@ export const dispatchPiAgent = async (props: DispatchAgentModuleProps): Promise< const apiKey = getModelApiKey(model); const toolCtx: ToolDispatchContext = { - checkIsStopping, - chatConfig, - runningUserInfo: props.runningUserInfo, - runningAppInfo, - chatId, - uid: props.uid, - variables: props.variables, - externalProvider: props.externalProvider, - workflowStreamResponse, - lang, - requestOrigin, - mode, - timezone: props.timezone, - retainDatasetCite: props.retainDatasetCite, - maxRunTimes: props.maxRunTimes, - workflowDispatchDeep: props.workflowDispatchDeep, - usagePush, - model, - datasetParams + ...props, + streamResponseFn: workflowStreamResponse, + getSubAppInfo, + getSubApp, + completionTools: agentCompletionTools, + filesMap, + capabilityToolCallHandler }; const piTools = await buildAgentTools({ - completionTools: agentCompletionTools, ctx: toolCtx, - filesMap, - getSubApp, - getSubAppInfo, - capabilityToolCallHandler, - nodeResponses + nodeResponses, + usagePush }); /* ===== Restore session messages from last AI history ===== */ diff --git a/packages/service/core/workflow/dispatch/ai/agent/piAgent/toolAdapter.ts b/packages/service/core/workflow/dispatch/ai/agent/piAgent/toolAdapter.ts index 0905405ca3..7594aba996 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/piAgent/toolAdapter.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/piAgent/toolAdapter.ts @@ -1,328 +1,56 @@ -import type { ChatCompletionTool } from '@fastgpt/global/core/ai/llm/type'; import type { ChatHistoryItemResType } from '@fastgpt/global/core/chat/type'; import { SubAppIds } from '@fastgpt/global/core/workflow/node/agent/constants'; -import { - SANDBOX_TOOL_NAME, - SANDBOX_GET_FILE_URL_TOOL_NAME, - SandboxShellToolSchema, - SandboxGetFileUrlToolSchema -} from '@fastgpt/global/core/ai/sandbox/constants'; -import { ReadFileToolSchema } from '../sub/file/utils'; -import { DatasetSearchToolSchema } from '../sub/dataset/utils'; -import { dispatchFileRead } from '../sub/file'; -import { dispatchAgentDatasetSearch } from '../sub/dataset'; -import { dispatchSandboxShell, dispatchSandboxGetFileUrl } from '../sub/sandbox'; -import { dispatchTool } from '../sub/tool'; -import { dispatchApp, dispatchPlugin } from '../sub/app'; -import { parseJsonArgs } from '../../../../../ai/utils'; -import { getErrText } from '@fastgpt/global/common/error/utils'; -import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; import { SseResponseEventEnum } from '@fastgpt/global/core/workflow/runtime/constants'; -import type { GetSubAppInfoFnType, SubAppRuntimeType } from '../type'; -import type { CapabilityToolCallHandlerType } from '../capability/type'; import type { DispatchAgentModuleProps } from '..'; -import type { AppFormEditFormType } from '@fastgpt/global/core/app/formEdit/type'; -import type { OpenaiAccountType } from '@fastgpt/global/support/user/team/type'; +import { getExecuteTool, type ToolDispatchContext } from '../utils'; type AgentTool = import('@mariozechner/pi-agent-core').AgentTool; -// Flatten context for tool dispatch (avoids NodeInputKeyEnum computed-key Pick issues) -export type ToolDispatchContext = Pick< - DispatchAgentModuleProps, - | 'checkIsStopping' - | 'chatConfig' - | 'runningUserInfo' - | 'runningAppInfo' - | 'chatId' - | 'uid' - | 'variables' - | 'externalProvider' - | 'workflowStreamResponse' - | 'lang' - | 'requestOrigin' - | 'mode' - | 'timezone' - | 'retainDatasetCite' - | 'maxRunTimes' - | 'workflowDispatchDeep' - | 'usagePush' -> & { - model: string; - datasetParams?: AppFormEditFormType['dataset']; -}; - export async function buildAgentTools({ - completionTools, ctx, - filesMap, - getSubApp, - getSubAppInfo, - capabilityToolCallHandler, - nodeResponses + nodeResponses, + usagePush }: { - completionTools: ChatCompletionTool[]; ctx: ToolDispatchContext; - filesMap: Record; - getSubApp: (id: string) => SubAppRuntimeType | undefined; - getSubAppInfo: GetSubAppInfoFnType; - capabilityToolCallHandler?: CapabilityToolCallHandlerType; nodeResponses: ChatHistoryItemResType[]; + usagePush: DispatchAgentModuleProps['usagePush']; }): Promise { const { Type } = await import('@mariozechner/pi-ai'); - const { - checkIsStopping, - chatConfig, - runningUserInfo, - runningAppInfo, - chatId, - uid, - variables, - externalProvider, - workflowStreamResponse, - lang, - requestOrigin, - mode, - timezone, - retainDatasetCite, - maxRunTimes, - workflowDispatchDeep, - usagePush, - model, - datasetParams - } = ctx; - + const executeTool = getExecuteTool(ctx); const tools: AgentTool[] = []; - for (const tool of completionTools) { + for (const tool of ctx.completionTools) { const toolId = tool.function.name; // pi-agent-core manages multi-turn reasoning; skip the plan tool if (toolId === SubAppIds.plan) continue; - const execute = async ( - callId: string, - args: Record, - _signal?: AbortSignal - ): Promise<{ content: { type: 'text'; text: string }[]; details: Record }> => { + const execute = async (callId: string, args: Record, _signal?: AbortSignal) => { const argStr = JSON.stringify(args); - try { - const { response, usages = [] } = await (async (): Promise<{ - response: string; - usages?: any[]; - }> => { - if (toolId === SubAppIds.fileRead) { - const toolParams = ReadFileToolSchema.safeParse(args); - if (!toolParams.success) return { response: toolParams.error.message }; - const files = toolParams.data.file_indexes.map((index) => ({ - index, - url: filesMap[index] - })); - const result = await dispatchFileRead({ - files, - teamId: runningUserInfo.teamId, - tmbId: runningUserInfo.tmbId, - customPdfParse: chatConfig?.fileSelectConfig?.customPdfParse, - model, - userKey: externalProvider.openaiAccount as OpenaiAccountType | undefined - }); - if (result.nodeResponse) nodeResponses.push(result.nodeResponse); - return { response: result.response, usages: result.usages }; - } + const { + response, + usages = [], + nodeResponse + } = await executeTool({ + callId, + toolId, + args: argStr + }); - if (toolId === SubAppIds.datasetSearch) { - const toolParams = DatasetSearchToolSchema.safeParse(args); - if (!toolParams.success) return { response: toolParams.error.message }; - if (!datasetParams || datasetParams.datasets.length === 0) { - return { response: 'No dataset selected' }; - } - const result = await dispatchAgentDatasetSearch({ - query: toolParams.data.query, - config: { - datasets: datasetParams.datasets, - similarity: datasetParams.similarity || 0.4, - maxTokens: datasetParams.limit || 5000, - searchMode: datasetParams.searchMode, - embeddingWeight: datasetParams.embeddingWeight, - usingReRank: datasetParams.usingReRank ?? false, - rerankModel: datasetParams.rerankModel, - rerankWeight: datasetParams.rerankWeight || 0.5, - usingExtensionQuery: datasetParams.datasetSearchUsingExtensionQuery ?? false, - extensionModel: datasetParams.datasetSearchExtensionModel, - extensionBg: datasetParams.datasetSearchExtensionBg - }, - teamId: runningUserInfo.teamId, - tmbId: runningUserInfo.tmbId, - llmModel: model - }); - if (result.nodeResponse) nodeResponses.push(result.nodeResponse); - return { response: result.response, usages: result.usages }; - } + { + if (nodeResponse) nodeResponses.push(nodeResponse); + if (usages.length > 0) usagePush(usages); - if (toolId === SANDBOX_TOOL_NAME) { - const toolParams = SandboxShellToolSchema.safeParse(args); - if (!toolParams.success) return { response: toolParams.error.message }; - const result = await dispatchSandboxShell({ - command: toolParams.data.command, - timeout: toolParams.data.timeout, - appId: runningAppInfo.id, - userId: uid, - chatId, - lang - }); - nodeResponses.push(result.nodeResponse); - return { response: result.response, usages: result.usages }; - } - - if (toolId === SANDBOX_GET_FILE_URL_TOOL_NAME) { - const toolParams = SandboxGetFileUrlToolSchema.safeParse(args); - if (!toolParams.success) return { response: toolParams.error.message }; - const result = await dispatchSandboxGetFileUrl({ - paths: toolParams.data.paths, - appId: runningAppInfo.id, - userId: uid, - chatId, - lang - }); - nodeResponses.push(result.nodeResponse); - return { response: result.response, usages: result.usages }; - } - - // Capability tools (e.g. sandbox skills) - const capResult = await capabilityToolCallHandler?.(toolId, argStr, callId); - if (capResult != null) { - const subInfo = getSubAppInfo(toolId); - nodeResponses.push({ - nodeId: callId, - id: callId, - moduleType: FlowNodeTypeEnum.tool, - moduleName: subInfo.name, - moduleLogo: subInfo.avatar, - toolInput: parseJsonArgs(argStr), - toolRes: capResult.response - }); - if (capResult.usages?.length) usagePush(capResult.usages); - return { response: capResult.response, usages: capResult.usages }; - } - - // User sub-apps - const subApp = getSubApp(toolId); - if (!subApp) return { response: `Can't find the tool ${toolId}` }; - - const requestParams = { ...subApp.params, ...args }; - - if (subApp.type === 'tool') { - const { response, usages, runningTime, toolParams, result } = await dispatchTool({ - tool: { - name: subApp.name, - version: subApp.version, - toolConfig: subApp.toolConfig - }, - params: requestParams, - runningUserInfo, - runningAppInfo, - chatId, - uid, - variables, - workflowStreamResponse - }); - nodeResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.tool, - moduleName: subApp.name, - moduleLogo: subApp.avatar, - toolInput: toolParams, - toolRes: result || response, - totalPoints: usages?.reduce((sum: number, item: any) => sum + item.totalPoints, 0) - }); - return { response, usages }; - } - - if (subApp.type === 'workflow') { - const { userChatInput, ...params } = requestParams; - const { response, runningTime, usages } = await dispatchApp({ - appId: subApp.id, - userChatInput: userChatInput ?? '', - customAppVariables: params, - checkIsStopping, - lang, - requestOrigin, - mode, - timezone, - externalProvider, - runningAppInfo, - runningUserInfo, - retainDatasetCite, - maxRunTimes, - workflowDispatchDeep, - variables - }); - nodeResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.appModule, - moduleName: subApp.name, - moduleLogo: subApp.avatar, - toolInput: requestParams, - toolRes: response, - totalPoints: usages?.reduce((sum: number, item: any) => sum + item.totalPoints, 0) - }); - return { response, usages }; - } - - if (subApp.type === 'toolWorkflow') { - const { response, result, runningTime, usages } = await dispatchPlugin({ - appId: subApp.id, - userChatInput: '', - customAppVariables: requestParams, - checkIsStopping, - lang, - requestOrigin, - mode, - timezone, - externalProvider, - runningAppInfo, - runningUserInfo, - retainDatasetCite, - maxRunTimes, - workflowDispatchDeep, - variables - }); - nodeResponses.push({ - nodeId: callId, - id: callId, - runningTime, - moduleType: FlowNodeTypeEnum.pluginModule, - moduleName: subApp.name, - moduleLogo: subApp.avatar, - toolInput: requestParams, - toolRes: result, - totalPoints: usages?.reduce((sum: number, item: any) => sum + item.totalPoints, 0) - }); - return { response, usages }; - } - - return { response: 'Invalid tool type' }; - })(); - - if (usages && usages.length > 0) usagePush(usages); - - // SSE tool response - workflowStreamResponse?.({ + ctx.streamResponseFn?.({ id: callId, event: SseResponseEventEnum.toolResponse, data: { tool: { response } } }); - - return { content: [{ type: 'text' as const, text: response }], details: {} }; - } catch (error) { - const errText = `Tool error: ${getErrText(error)}`; - return { content: [{ type: 'text' as const, text: errText }], details: {} }; } + + return { content: [{ type: 'text' as const, text: response }], details: {} }; }; // Wrap execute to also emit SSE toolCall event before execution @@ -331,8 +59,8 @@ export async function buildAgentTools({ args: Record, signal?: AbortSignal ) => { - const subAppInfo = getSubAppInfo(toolId); - workflowStreamResponse?.({ + const subAppInfo = ctx.getSubAppInfo(toolId); + ctx.streamResponseFn?.({ id: callId, event: SseResponseEventEnum.toolCall, data: { diff --git a/packages/service/core/workflow/dispatch/ai/agent/sub/app/index.ts b/packages/service/core/workflow/dispatch/ai/agent/sub/app/index.ts index 9d33726ebb..68f729b9c5 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/sub/app/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/sub/app/index.ts @@ -37,7 +37,11 @@ type Props = Pick< | 'responseDetail' | 'variables' > & { - appId: string; + app: { + name: string; + avatar?: string; + id: string; + }; userChatInput: string; customAppVariables: Record; }; @@ -46,25 +50,21 @@ export const dispatchApp = async (props: Props): Promise const { runningAppInfo, runningUserInfo, - appId, + app, variables, customAppVariables, userChatInput, ...data } = props; - if (!appId) { - return Promise.reject(new Error('AppId is empty')); - } - // Auth the app by tmbId(Not the user, but the workflow user) const { app: appData } = await authAppByTmbId({ - appId, + appId: app.id, tmbId: runningAppInfo.tmbId, per: ReadPermissionVal }); const { nodes, edges, chatConfig } = await getAppVersionById({ - appId, + appId: app.id, app: appData }); @@ -86,7 +86,7 @@ export const dispatchApp = async (props: Props): Promise ); const runtimeEdges = storeEdges2RuntimeEdges(edges); - const { assistantResponses, flowUsages, runTimes } = await runWorkflow({ + const { assistantResponses, flowUsages } = await runWorkflow({ ...data, uid: variables.userId, chatId: variables.chatId, @@ -119,9 +119,17 @@ export const dispatchApp = async (props: Props): Promise return { response: text, - result: {}, - runningTime: runTimes || 0, - usages: flowUsages + usages: flowUsages, + nodeResponse: { + moduleType: FlowNodeTypeEnum.appModule, + moduleName: app.name, + moduleLogo: app.avatar, + toolInput: { + userChatInput, + ...customAppVariables + }, + toolRes: text + } }; }; @@ -129,25 +137,21 @@ export const dispatchPlugin = async (props: Props): Promise => { - const startTime = Date.now(); - getLogger(LogCategories.MODULE.AI.AGENT).debug('[Agent Dataset Search] Starting', { +}: DatasetSearchParams): Promise => { + if (!datasetParams || datasetParams.datasets.length === 0) { + return { + response: 'No dataset selected' + }; + } + + const toolParams = DatasetSearchToolSchema.safeParse(parseJsonArgs(args)); + if (!toolParams.success) { + return { + response: toolParams.error.message + }; + } + + const query = toolParams.data.query; + + logger.debug('[Agent Dataset Search] Starting', { query, - config + datasetParams }); try { - const datasetIds = await Promise.resolve(config.datasets.map((item) => item.datasetId)); - - if (datasetIds.length === 0) { - return { - response: 'No dataset selected', - usages: [] - }; - } + const datasetIds = await Promise.resolve(datasetParams.datasets.map((item) => item.datasetId)); // Get vector model const vectorModel = getEmbeddingModel( (await MongoDataset.findById(datasetIds[0], 'vectorModel').lean())?.vectorModel ); // Get Rerank Model - const rerankModelData = getRerankModel(config.rerankModel); + const rerankModelData = getRerankModel(datasetParams.rerankModel); const searchData: DefaultSearchDatasetDataProps = { histories: [], @@ -196,17 +188,17 @@ export const dispatchAgentDatasetSearch = async ({ reRankQuery: query, queries: [query], model: vectorModel.model, - similarity: config.similarity, - limit: config.maxTokens, + similarity: datasetParams.similarity ?? 0.4, + limit: datasetParams.limit || 5000, datasetIds, - searchMode: config.searchMode, - embeddingWeight: config.embeddingWeight, - usingReRank: config.usingReRank, + searchMode: datasetParams.searchMode, + embeddingWeight: datasetParams.embeddingWeight, + usingReRank: datasetParams.usingReRank, rerankModel: rerankModelData, - rerankWeight: config.rerankWeight, - datasetSearchUsingExtensionQuery: config.usingExtensionQuery, - datasetSearchExtensionModel: config.extensionModel, - datasetSearchExtensionBg: config.extensionBg + rerankWeight: datasetParams.rerankWeight ?? 0.5, + datasetSearchUsingExtensionQuery: datasetParams.datasetSearchUsingExtensionQuery ?? false, + datasetSearchExtensionModel: datasetParams.datasetSearchExtensionModel, + datasetSearchExtensionBg: datasetParams.datasetSearchExtensionBg }; const { searchRes, @@ -295,29 +287,24 @@ export const dispatchAgentDatasetSearch = async ({ }); } } - const totalPoints = usages.reduce((acc, item) => acc + item.totalPoints, 0); - const id = getNanoid(6); - const nodeResponse: ChatHistoryItemResType = { - nodeId: id, - id: id, + const nodeResponse: DispatchSubAppResponse['nodeResponse'] = { moduleType: FlowNodeTypeEnum.datasetSearchNode, moduleName: i18nT('chat:dataset_search'), - totalPoints, query, embeddingModel: vectorModel.name, embeddingTokens, - similarity: usingSimilarityFilter ? config.similarity : undefined, - limit: config.maxTokens, - searchMode: config.searchMode, + similarity: usingSimilarityFilter ? searchData.similarity : undefined, + limit: searchData.limit, + searchMode: searchData.searchMode, embeddingWeight: - config.searchMode === DatasetSearchModeEnum.mixedRecall - ? config.embeddingWeight + searchData.searchMode === DatasetSearchModeEnum.mixedRecall + ? searchData.embeddingWeight : undefined, // Rerank ...(searchUsingReRank && { rerankModel: rerankModelData?.name, - rerankWeight: config.rerankWeight, + rerankWeight: searchData.rerankWeight, reRankInputTokens }), searchUsingReRank, @@ -330,8 +317,7 @@ export const dispatchAgentDatasetSearch = async ({ } : undefined, // Results - quoteList: searchResults, - runningTime: +((Date.now() - startTime) / 1000).toFixed(2) + quoteList: searchResults }; return { @@ -340,10 +326,9 @@ export const dispatchAgentDatasetSearch = async ({ nodeResponse }; } catch (error) { - getLogger(LogCategories.MODULE.AI.AGENT).error('[Agent Dataset Search] Failed', { error }); + logger.error('[Agent Dataset Search] Failed', { error }); return { - response: `Failed to search dataset: ${getErrText(error)}`, - usages: [] + response: `Failed to search dataset: ${getErrText(error)}` }; } }; diff --git a/packages/service/core/workflow/dispatch/ai/agent/sub/file/index.ts b/packages/service/core/workflow/dispatch/ai/agent/sub/file/index.ts index 722f3e433c..a4d3cf3528 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/sub/file/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/sub/file/index.ts @@ -13,12 +13,11 @@ import { getLLMModel } from '../../../../../../ai/model'; import { compressLargeContent } from '../../../../../../ai/llm/compress'; import { calculateCompressionThresholds } from '../../../../../../ai/llm/compress/constants'; import type { ChatNodeUsageType } from '@fastgpt/global/support/wallet/bill/type'; -import { getNanoid } from '@fastgpt/global/common/string/tools'; import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; import { i18nT } from '../../../../../../../../web/i18n/utils'; -import type { ChatHistoryItemResType } from '@fastgpt/global/core/chat/type'; import { getLogger, LogCategories } from '../../../../../../../common/logger'; import type { OpenaiAccountType } from '@fastgpt/global/support/user/team/type'; +import type { DispatchSubAppResponse } from '../../type'; type FileReadParams = { files: { index: string; url: string }[]; @@ -37,12 +36,7 @@ export const dispatchFileRead = async ({ customPdfParse, model, userKey -}: FileReadParams): Promise<{ - response: string; - usages: ChatNodeUsageType[]; - nodeResponse?: ChatHistoryItemResType; -}> => { - const startTime = Date.now(); +}: FileReadParams): Promise => { try { const usages: ChatNodeUsageType[] = []; const readFilesResult = await Promise.all( @@ -162,12 +156,8 @@ export const dispatchFileRead = async ({ response: responseText, usages, nodeResponse: { - nodeId: getNanoid(6), - id: getNanoid(6), moduleType: FlowNodeTypeEnum.readFiles, moduleName: i18nT('chat:read_file'), - totalPoints: usages.reduce((acc, item) => acc + item.totalPoints, 0), - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), compressTextAgent: result.usage ? { inputTokens: result.usage.inputTokens || 0, diff --git a/packages/service/core/workflow/dispatch/ai/agent/sub/sandbox/index.ts b/packages/service/core/workflow/dispatch/ai/agent/sub/sandbox/index.ts index 61385fede6..f9f12bfa4a 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/sub/sandbox/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/sub/sandbox/index.ts @@ -1,72 +1,28 @@ -import type { ChatNodeUsageType } from '@fastgpt/global/support/wallet/bill/type'; -import { getNanoid } from '@fastgpt/global/common/string/tools'; import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; -import type { ChatHistoryItemResType } from '@fastgpt/global/core/chat/type'; -import { - SANDBOX_ICON, - SANDBOX_NAME, - SANDBOX_TOOL_NAME, - SANDBOX_GET_FILE_URL_TOOL_NAME -} from '@fastgpt/global/core/ai/sandbox/constants'; +import { SANDBOX_ICON, SANDBOX_NAME } from '@fastgpt/global/core/ai/sandbox/constants'; import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; import type { localeType } from '@fastgpt/global/common/i18n/type'; -import { callSandboxTool } from '../../../../../../ai/sandbox/toolCall'; +import { runSandboxTools } from '../../../../../../ai/sandbox/toolCall'; +import type { DispatchSubAppResponse } from '../../type'; -type SandboxDispatchParams = { +export const dispatchSandboxTool = async ({ + toolName, + rawArgs, + appId, + userId, + chatId, + lang +}: { + toolName: string; + rawArgs: string; appId: string; userId: string; chatId: string; lang?: localeType; -}; - -type SandboxDispatchResult = { - response: string; - usages: ChatNodeUsageType[]; - nodeResponse: ChatHistoryItemResType; -}; - -const buildNodeResponse = ({ - toolId, - input, - response, - durationSeconds, - lang -}: { - toolId: string; - input: Record; - response: string; - durationSeconds: number; - lang?: localeType; -}): ChatHistoryItemResType => { - const nodeId = getNanoid(6); - return { - nodeId, - id: nodeId, - moduleType: FlowNodeTypeEnum.tool, - moduleName: parseI18nString(SANDBOX_NAME, lang), - moduleLogo: SANDBOX_ICON, - toolId, - toolInput: input, - toolRes: response, - totalPoints: 0, - runningTime: durationSeconds - }; -}; - -export const dispatchSandboxShell = async ({ - command, - timeout, - appId, - userId, - chatId, - lang -}: SandboxDispatchParams & { - command: string; - timeout?: number; -}): Promise => { - const { input, response, durationSeconds } = await callSandboxTool({ - toolName: SANDBOX_TOOL_NAME, - rawArgs: JSON.stringify({ command, timeout }), +}): Promise => { + const { input, response } = await runSandboxTools({ + toolName, + args: rawArgs, appId, userId, chatId @@ -74,44 +30,14 @@ export const dispatchSandboxShell = async ({ return { response, - usages: [], - nodeResponse: buildNodeResponse({ - toolId: SANDBOX_TOOL_NAME, - input, - response, - durationSeconds, - lang - }) - }; -}; - -export const dispatchSandboxGetFileUrl = async ({ - paths, - appId, - userId, - chatId, - lang -}: SandboxDispatchParams & { - paths: string[]; -}): Promise => { - const { input, response, durationSeconds } = await callSandboxTool({ - toolName: SANDBOX_GET_FILE_URL_TOOL_NAME, - rawArgs: JSON.stringify({ paths }), - appId, - userId, - chatId - }); - - return { - response, - usages: [], - nodeResponse: buildNodeResponse({ - toolId: SANDBOX_GET_FILE_URL_TOOL_NAME, - input, - response, - durationSeconds, - lang - }) + nodeResponse: { + moduleType: FlowNodeTypeEnum.tool, + moduleName: parseI18nString(SANDBOX_NAME, lang), + moduleLogo: SANDBOX_ICON, + toolId: toolName, + toolInput: input, + toolRes: response + } }; }; diff --git a/packages/service/core/workflow/dispatch/ai/agent/sub/tool/index.ts b/packages/service/core/workflow/dispatch/ai/agent/sub/tool/index.ts index eadc8528f8..b82e8ab55e 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/sub/tool/index.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/sub/tool/index.ts @@ -1,7 +1,6 @@ import type { StoreSecretValueType } from '@fastgpt/global/common/secret/type'; import { SystemToolSecretInputTypeEnum } from '@fastgpt/global/core/app/tool/systemTool/constants'; import type { DispatchSubAppResponse } from '../../type'; -import { splitCombineToolId } from '@fastgpt/global/core/app/tool/utils'; import { getSystemToolById } from '../../../../../../app/tool/controller'; import { getSecretValue } from '../../../../../../../common/secret/utils'; import { MongoSystemTool } from '../../../../../../plugin/tool/systemToolSchema'; @@ -21,6 +20,9 @@ import { MCPClient } from '../../../../../../app/mcp'; import { runHTTPTool } from '../../../../../../app/http'; import { getS3ChatSource } from '../../../../../../../common/s3/sources/chat'; import { parseToolId } from '../../../../child/runTool'; +import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; +import { getNanoid } from '@fastgpt/global/common/string/tools'; +import type { RequireOnlyOne } from '@fastgpt/global/common/type/utils'; type SystemInputConfigType = { type: SystemToolSecretInputTypeEnum; @@ -29,6 +31,7 @@ type SystemInputConfigType = { export type Props = { tool: { name: string; + avatar?: string; version?: string; toolConfig: RuntimeNodeItemType['toolConfig']; }; @@ -45,7 +48,7 @@ export type Props = { }; export const dispatchTool = async ({ - tool: { name, version, toolConfig }, + tool: { name, avatar, version, toolConfig }, params: { system_input_config, ...params }, runningUserInfo, runningAppInfo, @@ -53,19 +56,29 @@ export const dispatchTool = async ({ uid, variables, workflowStreamResponse -}: Props): Promise< - DispatchSubAppResponse & { - toolParams: Record; - } -> => { - const startTime = Date.now(); - - const getErrResponse = (error: any) => { +}: Props): Promise => { + const getNodeResponse = ({ + result, + response + }: RequireOnlyOne<{ + result?: any; + response?: string; + }>): DispatchSubAppResponse['nodeResponse'] => { return { - toolParams: params, - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), - response: getErrText(error, 'Call tool error'), - usages: [] + moduleType: FlowNodeTypeEnum.tool, + moduleName: name, + moduleLogo: avatar, + toolInput: params, + toolRes: result || response + }; + }; + const getErrResponse = (error: any): DispatchSubAppResponse => { + const response = getErrText(error, 'Call tool error'); + return { + response, + nodeResponse: getNodeResponse({ + response + }) }; }; @@ -165,9 +178,9 @@ export const dispatchTool = async ({ return { response: JSON.stringify(result), - toolParams: params, - result, - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), + nodeResponse: getNodeResponse({ + result + }), usages: [ { moduleName: name, @@ -196,11 +209,10 @@ export const dispatchTool = async ({ params }); return { - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), response: JSON.stringify(result), - toolParams: params, - result, - usages: [] + nodeResponse: getNodeResponse({ + result: result + }) }; } else if (toolConfig?.httpTool?.toolId) { const { parentId, toolName } = parseToolId(toolConfig.httpTool.toolId); @@ -242,19 +254,18 @@ export const dispatchTool = async ({ if (errorMsg) { return { - toolParams: params, - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), - response: errorMsg, - usages: [] + nodeResponse: getNodeResponse({ + response: errorMsg + }), + response: errorMsg }; } return { - toolParams: params, - runningTime: +((Date.now() - startTime) / 1000).toFixed(2), - response: typeof data === 'object' ? JSON.stringify(data) : data, - result: data, - usages: [] + nodeResponse: getNodeResponse({ + result: data + }), + response: typeof data === 'object' ? JSON.stringify(data) : data }; } else { return getErrResponse("Can't find the tool"); diff --git a/packages/service/core/workflow/dispatch/ai/agent/type.ts b/packages/service/core/workflow/dispatch/ai/agent/type.ts index 491f1966d2..30813b6542 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/type.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/type.ts @@ -3,6 +3,7 @@ import type { JSONSchemaInputType } from '@fastgpt/global/core/app/jsonschema'; import type { ChatNodeUsageType } from '@fastgpt/global/support/wallet/bill/type'; import z from 'zod'; import { NodeToolConfigTypeSchema } from '@fastgpt/global/core/workflow/type/node'; +import type { ChatHistoryItemResType } from '@fastgpt/global/core/chat/type'; export type ToolNodeItemType = RuntimeNodeItemType & { toolParams: RuntimeNodeItemType['inputs']; @@ -10,10 +11,9 @@ export type ToolNodeItemType = RuntimeNodeItemType & { }; export type DispatchSubAppResponse = { - response: string; - result?: any; - runningTime: number; + response: string; // 返回给 LLM 的响应 usages?: ChatNodeUsageType[]; + nodeResponse?: Omit; // 部分字段外层会自动根据 usages 计算。 }; export const SubAppRuntimeSchema = z.object({ diff --git a/packages/service/core/workflow/dispatch/ai/agent/utils.ts b/packages/service/core/workflow/dispatch/ai/agent/utils.ts index 856bd7fcc3..55b7927513 100644 --- a/packages/service/core/workflow/dispatch/ai/agent/utils.ts +++ b/packages/service/core/workflow/dispatch/ai/agent/utils.ts @@ -1,12 +1,30 @@ import type { localeType } from '@fastgpt/global/common/i18n/type'; import type { SkillToolType } from '@fastgpt/global/core/ai/skill/type'; -import type { SubAppRuntimeType } from './type'; +import type { DispatchSubAppResponse, GetSubAppInfoFnType, SubAppRuntimeType } from './type'; import { getAgentRuntimeTools } from './sub/tool/utils'; import type { ChatCompletionTool } from '@fastgpt/global/core/ai/llm/type'; -import { readFileTool } from './sub/file/utils'; -import { PlanAgentTool } from './sub/plan/constants'; +import { readFileTool, ReadFileToolSchema } from './sub/file/utils'; +import { PlanAgentParamsSchema, PlanAgentTool } from './sub/plan/constants'; import { datasetSearchTool } from './sub/dataset/utils'; -import { SANDBOX_TOOLS } from '@fastgpt/global/core/ai/sandbox/constants'; +import { SANDBOX_TOOLS, sandboxToolMap } from '@fastgpt/global/core/ai/sandbox/constants'; +import type { ChatNodeUsageType } from '@fastgpt/global/support/wallet/bill/type'; +import { SubAppIds } from '@fastgpt/global/core/workflow/node/agent/constants'; +import { dispatchFileRead } from './sub/file'; +import type { DispatchAgentModuleProps } from '.'; +import { dispatchAgentDatasetSearch } from './sub/dataset'; +import { dispatchSandboxTool } from './sub/sandbox'; +import type { CapabilityToolCallHandlerType } from './capability/type'; +import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant'; +import { parseJsonArgs } from '../../../../ai/utils'; +import type { DispatchPlanAgentResponse } from './sub/plan'; +import { dispatchPlanAgent } from './sub/plan'; +import { getLogger, LogCategories } from '../../../../../common/logger'; +import { getErrText } from '@fastgpt/global/common/error/utils'; +import { dispatchTool } from './sub/tool'; +import type { WorkflowResponseItemType } from '../../type'; +import { dispatchApp, dispatchPlugin } from './sub/app'; +import { SseResponseEventEnum } from '@fastgpt/global/core/workflow/runtime/constants'; +import type { AIChatItemValueItemType } from '@fastgpt/global/core/chat/type'; export const getSubapps = async ({ tmbId, @@ -30,34 +48,37 @@ export const getSubapps = async ({ completionTools: ChatCompletionTool[]; subAppsMap: Map; }> => { - const subAppsMap = new Map(); const completionTools: ChatCompletionTool[] = []; - /* Plan */ - if (getPlanTool) { - completionTools.push(PlanAgentTool); - } - /* File */ - if (hasFiles) { - completionTools.push(readFileTool); + // system tools + { + /* Plan */ + if (getPlanTool) { + completionTools.push(PlanAgentTool); + } + /* File */ + if (hasFiles) { + completionTools.push(readFileTool); + } + + /* Dataset Search */ + if (hasDataset) { + completionTools.push(datasetSearchTool); + } + + /* Sandbox Shell */ + if (useAgentSandbox && global.feConfigs?.show_agent_sandbox) { + completionTools.push(...SANDBOX_TOOLS); + } + + /* Capability extra tools (e.g. sandbox skills) */ + if (extraTools && extraTools.length > 0) { + completionTools.push(...extraTools); + } } - /* Dataset Search */ - if (hasDataset) { - completionTools.push(datasetSearchTool); - } - - /* Sandbox Shell */ - if (useAgentSandbox && global.feConfigs?.show_agent_sandbox) { - completionTools.push(...SANDBOX_TOOLS); - } - - /* Capability extra tools (e.g. sandbox skills) */ - if (extraTools && extraTools.length > 0) { - completionTools.push(...extraTools); - } - - /* System tool */ + /* User tools */ + const subAppsMap = new Map(); const formatTools = await getAgentRuntimeTools({ tools, tmbId, @@ -81,3 +102,330 @@ export const getSubapps = async ({ subAppsMap }; }; + +export type ToolDispatchContext = Pick< + DispatchAgentModuleProps, + | 'checkIsStopping' + | 'chatConfig' + | 'runningUserInfo' + | 'runningAppInfo' + | 'chatId' + | 'uid' + | 'variables' + | 'externalProvider' + | 'lang' + | 'requestOrigin' + | 'mode' + | 'timezone' + | 'retainDatasetCite' + | 'maxRunTimes' + | 'workflowDispatchDeep' + | 'params' + | 'stream' +> & { + systemPrompt?: string; + getSubAppInfo: GetSubAppInfoFnType; + getSubApp: (id: string) => SubAppRuntimeType | undefined; + completionTools: ChatCompletionTool[]; + filesMap: Record; + capabilityToolCallHandler?: CapabilityToolCallHandlerType; + streamResponseFn?: (args: WorkflowResponseItemType) => void | undefined; +}; +export const getExecuteTool = ({ + systemPrompt, + getSubAppInfo, + getSubApp, + completionTools, + filesMap, + capabilityToolCallHandler, + checkIsStopping, + chatConfig, + runningUserInfo, + runningAppInfo, + chatId, + uid, + variables, + externalProvider, + stream, + streamResponseFn, + params: { + model, + // Dataset search configuration + agent_datasetParams: datasetParams + }, + lang, + requestOrigin, + mode, + timezone, + retainDatasetCite, + maxRunTimes, + workflowDispatchDeep +}: ToolDispatchContext) => { + return async ({ callId, toolId, args }: { callId: string; toolId: string; args: string }) => { + let planResult: DispatchPlanAgentResponse | undefined; + const capabilityAssistantResponses: AIChatItemValueItemType[] = []; + const startTime = Date.now(); + + const { + response, + usages = [], + stop = false, + nodeResponse + } = await (async (): Promise<{ + response: string; + usages?: ChatNodeUsageType[]; + stop?: boolean; + nodeResponse?: DispatchSubAppResponse['nodeResponse']; + }> => { + try { + if (toolId in sandboxToolMap) { + const result = await dispatchSandboxTool({ + toolName: toolId, + rawArgs: args, + appId: runningAppInfo.id, + userId: uid, + chatId, + lang + }); + + return { + response: result.response, + usages: result.usages, + nodeResponse: result.nodeResponse + }; + } + + if (toolId === SubAppIds.fileRead) { + const toolParams = ReadFileToolSchema.safeParse(parseJsonArgs(args)); + if (!toolParams.success) { + return { + response: toolParams.error.message, + usages: [] + }; + } + const params = toolParams.data; + + const files = params.file_indexes.map((index) => ({ + index, + url: filesMap[index] + })); + const result = await dispatchFileRead({ + files, + teamId: runningUserInfo.teamId, + tmbId: runningUserInfo.tmbId, + customPdfParse: chatConfig?.fileSelectConfig?.customPdfParse, + model, + userKey: externalProvider.openaiAccount + }); + + return { + response: result.response, + usages: result.usages, + nodeResponse: result.nodeResponse + }; + } + if (toolId === SubAppIds.datasetSearch) { + const result = await dispatchAgentDatasetSearch({ + args: args, + datasetParams, + teamId: runningUserInfo.teamId, + tmbId: runningUserInfo.tmbId, + llmModel: model + }); + + return { + response: result.response, + usages: result.usages, + nodeResponse: result.nodeResponse + }; + } + if (toolId === SubAppIds.plan) { + try { + const toolArgs = await PlanAgentParamsSchema.safeParseAsync(parseJsonArgs(args)); + + if (!toolArgs.success) { + return { + response: 'Tool arguments is not valid' + }; + } + + // plan: 1,3 场景 + planResult = await dispatchPlanAgent({ + checkIsStopping, + completionTools, + getSubAppInfo, + systemPrompt, + model, + stream, + mode: 'initial', + ...toolArgs.data, + planId: callId + }); + + return { + response: '', + stop: true + }; + } catch (error) { + getLogger(LogCategories.MODULE.AI.AGENT).error('dispatchPlanAgent error', { error }); + return { + response: `Plan error: ${getErrText(error)}`, + stop: false + }; + } + } + + // TODO: 所有skill工具，合并成一个 function，不要依赖 capabilityToolCallHandler + // Capability tools (e.g. sandbox skills) + const capResult = await capabilityToolCallHandler?.(toolId, args ?? '', callId); + if (capResult != null) { + if (capResult.assistantResponses?.length) { + capabilityAssistantResponses.push(...capResult.assistantResponses); + } + const subInfo = getSubAppInfo(toolId); + return { + response: capResult.response, + usages: capResult.usages, + nodeResponse: { + moduleType: FlowNodeTypeEnum.tool, + moduleName: subInfo.name, + moduleLogo: subInfo.avatar, + toolInput: parseJsonArgs(args), + toolRes: capResult.response + } + }; + } + + // User Sub App + const tool = getSubApp(toolId); + if (!tool) { + return { + response: `Can't find the tool ${toolId}`, + usages: [] + }; + } + + // Get params + const toolCallParams = parseJsonArgs(args); + if (args && !toolCallParams) { + return { + response: 'Params is not object' + }; + } + const requestParams = { + ...tool.params, + ...toolCallParams + }; + + if (tool.type === 'tool') { + const { response, usages, nodeResponse } = await dispatchTool({ + tool: { + name: tool.name, + avatar: tool.avatar, + version: tool.version, + toolConfig: tool.toolConfig + }, + params: requestParams, + runningUserInfo, + runningAppInfo, + chatId, + uid, + variables, + workflowStreamResponse: streamResponseFn + }); + + return { + response, + usages, + nodeResponse + }; + } else if (tool.type === 'workflow') { + const { userChatInput, ...params } = requestParams; + + const { response, usages, nodeResponse } = await dispatchApp({ + app: { + name: tool.name, + avatar: tool.avatar, + id: tool.id + }, + userChatInput: userChatInput, + customAppVariables: params, + checkIsStopping, + lang, + requestOrigin, + mode, + timezone, + externalProvider, + runningAppInfo, + runningUserInfo, + retainDatasetCite, + maxRunTimes, + workflowDispatchDeep, + variables + }); + + return { + response, + usages, + nodeResponse + }; + } else if (tool.type === 'toolWorkflow') { + const { response, usages, nodeResponse } = await dispatchPlugin({ + app: { + name: tool.name, + avatar: tool.avatar, + id: tool.id + }, + userChatInput: '', + customAppVariables: requestParams, + checkIsStopping, + lang, + requestOrigin, + mode, + timezone, + externalProvider, + runningAppInfo, + runningUserInfo, + retainDatasetCite, + maxRunTimes, + workflowDispatchDeep, + variables + }); + + return { + response, + usages, + nodeResponse + }; + } else { + return { + response: 'Invalid tool type' + }; + } + } catch (error) { + return { + response: `Tool error: ${getErrText(error)}` + }; + } + })(); + + const formatNodeResponse = nodeResponse + ? { + ...nodeResponse, + nodeId: callId, + id: callId, + runningTime: +((Date.now() - startTime) / 1000).toFixed(2), + totalPoints: usages?.reduce((sum, item) => sum + item.totalPoints, 0) + } + : undefined; + + return { + response, + usages, + stop, + nodeResponse: formatNodeResponse, + planResult, + capabilityAssistantResponses + }; + }; +}; diff --git a/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts b/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts index 8c568e9499..a176c05eb2 100644 --- a/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts +++ b/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts @@ -14,20 +14,12 @@ import { parseJsonArgs } from '../../../../ai/utils'; import { sliceStrStartEnd } from '@fastgpt/global/common/string/tools'; import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants'; import { toolValueTypeList, valueTypeJsonSchemaMap } from '@fastgpt/global/core/workflow/constants'; -import { runAgentCall } from '../../../../ai/llm/agentCall'; +import { runAgentLoop } from '../../../../ai/llm/agentLoop'; import type { ToolCallChildrenInteractive } from '@fastgpt/global/core/workflow/template/system/interactive/type'; import type { JsonSchemaPropertiesItemType } from '@fastgpt/global/core/app/jsonschema'; -import { - SANDBOX_SYSTEM_PROMPT, - SANDBOX_ICON, - SANDBOX_TOOL_NAME, - SANDBOX_GET_FILE_URL_TOOL_NAME, - SANDBOX_TOOLS -} from '@fastgpt/global/core/ai/sandbox/constants'; +import { SANDBOX_SYSTEM_PROMPT, SANDBOX_TOOLS } from '@fastgpt/global/core/ai/sandbox/constants'; import { getSandboxToolWorkflowResponse } from './constants'; -import { callSandboxTool } from '../../../../ai/sandbox/toolCall'; -import { systemSubInfo } from '@fastgpt/global/core/workflow/node/agent/constants'; -import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; +import { getSandboxToolInfo, runSandboxTools } from '../../../../ai/sandbox/toolCall'; type ResponseType = { requestIds: string[]; @@ -75,6 +67,11 @@ export const runToolCall = async (props: DispatchToolModuleProps): Promise(); const tools: ChatCompletionTool[] = toolNodes.map((item) => { @@ -117,8 +114,7 @@ export const runToolCall = async (props: DispatchToolModuleProps): Promise { - const systemTool = systemSubInfo[name]; - if (systemTool) { + const sandboxToolInfo = getSandboxToolInfo(name, workflowProps.lang); + if (sandboxToolInfo) { return { - name: parseI18nString(systemTool.name, workflowProps.lang), - avatar: systemTool.avatar + type: 'sandbox' as const, + name: sandboxToolInfo.name, + avatar: sandboxToolInfo.avatar }; } const toolNode = toolNodesMap.get(name); - return { - name: toolNode?.name || '', - avatar: toolNode?.avatar || '', - rawData: toolNode - }; + if (toolNode) { + return { + type: 'user' as const, + name: toolNode.name, + avatar: toolNode.avatar, + rawData: toolNode + }; + } }; - // 工具响应原始值 - const toolRunResponses: ChildResponseItemType[] = []; - const { inputTokens, outputTokens, @@ -164,7 +161,7 @@ export const runToolCall = async (props: DispatchToolModuleProps): Promise { - const tool = getToolInfo(call.function?.name); + onRunTool: async ({ call }) => { + const toolInfo = getToolInfo(call.function?.name); + if (!toolInfo) { + return { + response: 'Call tool not found', + assistantMessages: [], + usages: [], + interactive: undefined, + stop: false + }; + } const { response, @@ -251,21 +257,18 @@ export const runToolCall = async (props: DispatchToolModuleProps): Promise { // 拦截 sandbox 工具调用 - if ( - call.function?.name === SANDBOX_TOOL_NAME || - call.function?.name === SANDBOX_GET_FILE_URL_TOOL_NAME - ) { - const { input, response, durationSeconds } = await callSandboxTool({ + if (toolInfo.type === 'sandbox') { + const { input, response, durationSeconds } = await runSandboxTools({ toolName: call.function.name, - rawArgs: call.function.arguments ?? '', - appId: String(workflowProps.runningAppInfo.id), - userId: String(workflowProps.uid), + args: call.function.arguments ?? '', + appId: workflowProps.runningAppInfo.id, + userId: workflowProps.uid, chatId: workflowProps.chatId }); const flowResponse = getSandboxToolWorkflowResponse({ - name: tool.name, - logo: SANDBOX_ICON, + name: toolInfo.name, + logo: toolInfo.avatar, toolId: call.function.name, input, response, @@ -274,13 +277,7 @@ export const runToolCall = async (props: DispatchToolModuleProps): Promise { + onRunInteractiveTool: async ({ childrenResponse, toolParams }) => { initToolNodes(runtimeNodes, childrenResponse.entryNodeIds); initToolCallEdges(runtimeEdges, childrenResponse.entryNodeIds); diff --git a/projects/app/package.json b/projects/app/package.json index f451cb285b..8b333e7a01 100644 --- a/projects/app/package.json +++ b/projects/app/package.json @@ -1,6 +1,6 @@ { "name": "app", - "version": "4.14.11", + "version": "4.14.13", "private": false, "scripts": { "dev": "NODE_OPTIONS='--max-old-space-size=8192' npm run build:workers && next dev", diff --git a/projects/app/src/pageComponents/app/detail/Edit/ChatAgent/hooks/useSkillManager.tsx b/projects/app/src/pageComponents/app/detail/Edit/ChatAgent/hooks/useSkillManager.tsx index f1244e9380..5719a770b6 100644 --- a/projects/app/src/pageComponents/app/detail/Edit/ChatAgent/hooks/useSkillManager.tsx +++ b/projects/app/src/pageComponents/app/detail/Edit/ChatAgent/hooks/useSkillManager.tsx @@ -31,6 +31,7 @@ import { import { useLatest } from 'ahooks'; import { SubAppIds, systemSubInfo } from '@fastgpt/global/core/workflow/node/agent/constants'; import { parseI18nString } from '@fastgpt/global/common/i18n/utils'; +import { SANDBOX_TOOL_NAME } from '@fastgpt/global/core/ai/sandbox/constants'; const ConfigToolModal = dynamic(() => import('../../component/ConfigToolModal')); @@ -111,10 +112,10 @@ export const useSkillManager = ({ }); } - const sandboxToolInfo = systemSubInfo[SubAppIds.sandboxTool]; + const sandboxToolInfo = systemSubInfo[SANDBOX_TOOL_NAME]; if (sandboxToolInfo) { apiTools.unshift({ - id: SubAppIds.sandboxTool, + id: SANDBOX_TOOL_NAME, label: parseI18nString(sandboxToolInfo.name, i18n.language), icon: sandboxToolInfo.avatar, description: sandboxToolInfo.toolDescription, @@ -338,11 +339,11 @@ export const useSkillManager = ({ } // Merge sandbox tool - const sandboxToolInfo = systemSubInfo[SubAppIds.sandboxTool]; + const sandboxToolInfo = systemSubInfo[SANDBOX_TOOL_NAME]; if (sandboxToolInfo) { tools.push({ - id: SubAppIds.sandboxTool, - pluginId: SubAppIds.sandboxTool, + id: SANDBOX_TOOL_NAME, + pluginId: SANDBOX_TOOL_NAME, name: parseI18nString(sandboxToolInfo.name, i18n.language), avatar: sandboxToolInfo.avatar, intro: sandboxToolInfo.toolDescription, diff --git a/test/cases/service/core/ai/llm/request.test.ts b/test/cases/service/core/ai/llm/request.test.ts index 5fc51bca92..74515d4996 100644 --- a/test/cases/service/core/ai/llm/request.test.ts +++ b/test/cases/service/core/ai/llm/request.test.ts @@ -747,8 +747,8 @@ describe('createLLMResponse', () => { onToolCall: ({ call }) => { toolCallResults.push(call); }, - onToolParam: ({ params }) => { - toolParamResults.push(params); + onToolParam: ({ argsDelta }) => { + toolParamResults.push(argsDelta); } });