mirror of
https://github.com/labring/FastGPT.git
synced 2025-07-23 13:03:50 +00:00

* Training status (#4424) * dataset data training state (#4311) * dataset data training state * fix * fix ts * fix * fix api format * fix * fix * perf: count training * format * fix: dataset training state (#4417) * fix * add test * fix * fix * fix test * fix test * perf: training count * count * loading status --------- Co-authored-by: heheer <heheer@sealos.io> * doc * website sync feature (#4429) * perf: introduce BullMQ for website sync (#4403) * perf: introduce BullMQ for website sync * feat: new redis module * fix: remove graceful shutdown * perf: improve UI in dataset detail - Updated the "change" icon SVG file. - Modified i18n strings. - Added new i18n string "immediate_sync". - Improved UI in dataset detail page, including button icons and background colors. * refactor: Add chunkSettings to DatasetSchema * perf: website sync ux * env template * fix: clean up website dataset when updating chunk settings (#4420) * perf: check setting updated * perf: worker currency * feat: init script for website sync refactor (#4425) * website feature doc --------- Co-authored-by: a.e. <49438478+I-Info@users.noreply.github.com> * pro migration (#4388) (#4433) * pro migration * reuse customPdfParseType Co-authored-by: gggaaallleee <91131304+gggaaallleee@users.noreply.github.com> * perf: remove loading ui * feat: config chat file expired time * Redis cache (#4436) * perf: add Redis cache for vector counting (#4432) * feat: cache * perf: get cache key --------- Co-authored-by: a.e. <49438478+I-Info@users.noreply.github.com> * perf: mobile voice input (#4437) * update:Mobile voice interaction (#4362) * Add files via upload * Add files via upload * Update ollama.md * Update ollama.md * Add files via upload * Update useSpeech.ts * Update ChatInput.tsx * Update useSpeech.ts * Update ChatInput.tsx * Update useSpeech.ts * Update constants.ts * Add files via upload * Update ChatInput.tsx * Update useSpeech.ts * Update useSpeech.ts * Update useSpeech.ts * Update ChatInput.tsx * Add files via upload * Update common.json * Update VoiceInput.tsx * Update ChatInput.tsx * Update VoiceInput.tsx * Update useSpeech.ts * Update useSpeech.ts * Update common.json * Update common.json * Update common.json * Update VoiceInput.tsx * Update VoiceInput.tsx * Update ChatInput.tsx * Update VoiceInput.tsx * Update ChatInput.tsx * Update VoiceInput.tsx * Update ChatInput.tsx * Update useSpeech.ts * Update common.json * Update chat.json * Update common.json * Update chat.json * Update common.json * Update chat.json * Update VoiceInput.tsx * Update ChatInput.tsx * Update useSpeech.ts * Update VoiceInput.tsx * speech ui * 优化语音输入组件,调整输入框显示逻辑,修复语音输入遮罩层样式,更新画布背景透明度,增强用户交互体验。 (#4435) * perf: mobil voice input --------- Co-authored-by: dreamer6680 <1468683855@qq.com> * Test completion v2 (#4438) * add v2 completions (#4364) * add v2 completions * completion config * config version * fix * frontend * doc * fix * fix: completions v2 api --------- Co-authored-by: heheer <heheer@sealos.io> * package * Test mongo log (#4443) * feat: mongodb-log (#4426) * perf: mongo log * feat: completions stop reasoner * mongo db log --------- Co-authored-by: Finley Ge <32237950+FinleyGe@users.noreply.github.com> * update doc * Update doc * fix external var ui (#4444) * action * fix: ts (#4458) * preview doc action add docs preview permission update preview action udpate action * update doc (#4460) * update preview action * update doc * remove * update * schema * update mq export;perf: redis cache (#4465) * perf: redis cache * update mq export * perf: website sync error tip * add error worker * website sync ui (#4466) * Updated the dynamic display of the voice input pop-up (#4469) * Update VoiceInput.tsx * Update VoiceInput.tsx * Update VoiceInput.tsx * fix: voice input --------- Co-authored-by: heheer <heheer@sealos.io> Co-authored-by: a.e. <49438478+I-Info@users.noreply.github.com> Co-authored-by: gggaaallleee <91131304+gggaaallleee@users.noreply.github.com> Co-authored-by: dreamer6680 <1468683855@qq.com> Co-authored-by: Finley Ge <32237950+FinleyGe@users.noreply.github.com>
292 lines
8.0 KiB
TypeScript
292 lines
8.0 KiB
TypeScript
import { LLMModelItemType } from '@fastgpt/global/core/ai/model.d';
|
||
import {
|
||
ChatCompletionCreateParamsNonStreaming,
|
||
ChatCompletionCreateParamsStreaming,
|
||
CompletionFinishReason,
|
||
StreamChatType
|
||
} from '@fastgpt/global/core/ai/type';
|
||
import { getLLMModel } from './model';
|
||
|
||
/*
|
||
Count response max token
|
||
*/
|
||
export const computedMaxToken = ({
|
||
maxToken,
|
||
model
|
||
}: {
|
||
maxToken?: number;
|
||
model: LLMModelItemType;
|
||
}) => {
|
||
if (maxToken === undefined) return;
|
||
|
||
maxToken = Math.min(maxToken, model.maxResponse);
|
||
return maxToken;
|
||
};
|
||
|
||
// FastGPT temperature range: [0,10], ai temperature:[0,2],{0,1]……
|
||
export const computedTemperature = ({
|
||
model,
|
||
temperature
|
||
}: {
|
||
model: LLMModelItemType;
|
||
temperature: number;
|
||
}) => {
|
||
if (typeof model.maxTemperature !== 'number') return undefined;
|
||
temperature = +(model.maxTemperature * (temperature / 10)).toFixed(2);
|
||
temperature = Math.max(temperature, 0.01);
|
||
|
||
return temperature;
|
||
};
|
||
|
||
type CompletionsBodyType =
|
||
| ChatCompletionCreateParamsNonStreaming
|
||
| ChatCompletionCreateParamsStreaming;
|
||
type InferCompletionsBody<T> = T extends { stream: true }
|
||
? ChatCompletionCreateParamsStreaming
|
||
: T extends { stream: false }
|
||
? ChatCompletionCreateParamsNonStreaming
|
||
: ChatCompletionCreateParamsNonStreaming | ChatCompletionCreateParamsStreaming;
|
||
|
||
export const llmCompletionsBodyFormat = <T extends CompletionsBodyType>(
|
||
body: T & {
|
||
response_format?: any;
|
||
json_schema?: string;
|
||
stop?: string;
|
||
},
|
||
model: string | LLMModelItemType
|
||
): InferCompletionsBody<T> => {
|
||
const modelData = typeof model === 'string' ? getLLMModel(model) : model;
|
||
if (!modelData) {
|
||
return body as unknown as InferCompletionsBody<T>;
|
||
}
|
||
|
||
const response_format = body.response_format;
|
||
const json_schema = body.json_schema ?? undefined;
|
||
const stop = body.stop ?? undefined;
|
||
|
||
const requestBody: T = {
|
||
...body,
|
||
model: modelData.model,
|
||
temperature:
|
||
typeof body.temperature === 'number'
|
||
? computedTemperature({
|
||
model: modelData,
|
||
temperature: body.temperature
|
||
})
|
||
: undefined,
|
||
...modelData?.defaultConfig,
|
||
response_format: response_format
|
||
? {
|
||
type: response_format,
|
||
json_schema
|
||
}
|
||
: undefined,
|
||
stop: stop?.split('|')
|
||
};
|
||
|
||
// field map
|
||
if (modelData.fieldMap) {
|
||
Object.entries(modelData.fieldMap).forEach(([sourceKey, targetKey]) => {
|
||
// @ts-ignore
|
||
requestBody[targetKey] = body[sourceKey];
|
||
// @ts-ignore
|
||
delete requestBody[sourceKey];
|
||
});
|
||
}
|
||
|
||
return requestBody as unknown as InferCompletionsBody<T>;
|
||
};
|
||
|
||
export const llmStreamResponseToAnswerText = async (response: StreamChatType) => {
|
||
let answer = '';
|
||
for await (const part of response) {
|
||
const content = part.choices?.[0]?.delta?.content || '';
|
||
answer += content;
|
||
}
|
||
return parseReasoningContent(answer)[1];
|
||
};
|
||
|
||
// Parse <think></think> tags to think and answer - unstream response
|
||
export const parseReasoningContent = (text: string): [string, string] => {
|
||
const regex = /<think>([\s\S]*?)<\/think>/;
|
||
const match = text.match(regex);
|
||
|
||
if (!match) {
|
||
return ['', text];
|
||
}
|
||
|
||
const thinkContent = match[1].trim();
|
||
|
||
// Add answer (remaining text after think tag)
|
||
const answerContent = text.slice(match.index! + match[0].length);
|
||
|
||
return [thinkContent, answerContent];
|
||
};
|
||
|
||
// Parse <think></think> tags to think and answer - stream response
|
||
export const parseReasoningStreamContent = () => {
|
||
let isInThinkTag: boolean | undefined;
|
||
|
||
const startTag = '<think>';
|
||
let startTagBuffer = '';
|
||
|
||
const endTag = '</think>';
|
||
let endTagBuffer = '';
|
||
|
||
/*
|
||
parseThinkTag - 只控制是否主动解析 <think></think>,如果接口已经解析了,则不再解析。
|
||
*/
|
||
const parsePart = (
|
||
part: {
|
||
choices: {
|
||
delta: {
|
||
content?: string;
|
||
reasoning_content?: string;
|
||
};
|
||
finish_reason?: CompletionFinishReason;
|
||
}[];
|
||
},
|
||
parseThinkTag = false
|
||
): {
|
||
reasoningContent: string;
|
||
content: string;
|
||
finishReason: CompletionFinishReason;
|
||
} => {
|
||
const content = part.choices?.[0]?.delta?.content || '';
|
||
const finishReason = part.choices?.[0]?.finish_reason || null;
|
||
|
||
// @ts-ignore
|
||
const reasoningContent = part.choices?.[0]?.delta?.reasoning_content || '';
|
||
if (reasoningContent || !parseThinkTag) {
|
||
isInThinkTag = false;
|
||
return { reasoningContent, content, finishReason };
|
||
}
|
||
|
||
if (!content) {
|
||
return {
|
||
reasoningContent: '',
|
||
content: '',
|
||
finishReason
|
||
};
|
||
}
|
||
|
||
// 如果不在 think 标签中,或者有 reasoningContent(接口已解析),则返回 reasoningContent 和 content
|
||
if (isInThinkTag === false) {
|
||
return {
|
||
reasoningContent: '',
|
||
content,
|
||
finishReason
|
||
};
|
||
}
|
||
|
||
// 检测是否为 think 标签开头的数据
|
||
if (isInThinkTag === undefined) {
|
||
// Parse content think and answer
|
||
startTagBuffer += content;
|
||
// 太少内容时候,暂时不解析
|
||
if (startTagBuffer.length < startTag.length) {
|
||
return {
|
||
reasoningContent: '',
|
||
content: '',
|
||
finishReason
|
||
};
|
||
}
|
||
|
||
if (startTagBuffer.startsWith(startTag)) {
|
||
isInThinkTag = true;
|
||
return {
|
||
reasoningContent: startTagBuffer.slice(startTag.length),
|
||
content: '',
|
||
finishReason
|
||
};
|
||
}
|
||
|
||
// 如果未命中 think 标签,则认为不在 think 标签中,返回 buffer 内容作为 content
|
||
isInThinkTag = false;
|
||
return {
|
||
reasoningContent: '',
|
||
content: startTagBuffer,
|
||
finishReason
|
||
};
|
||
}
|
||
|
||
// 确认是 think 标签内容,开始返回 think 内容,并实时检测 </think>
|
||
/*
|
||
检测 </think> 方案。
|
||
存储所有疑似 </think> 的内容,直到检测到完整的 </think> 标签或超出 </think> 长度。
|
||
content 返回值包含以下几种情况:
|
||
abc - 完全未命中尾标签
|
||
abc<th - 命中一部分尾标签
|
||
abc</think> - 完全命中尾标签
|
||
abc</think>abc - 完全命中尾标签
|
||
</think>abc - 完全命中尾标签
|
||
k>abc - 命中一部分尾标签
|
||
*/
|
||
// endTagBuffer 专门用来记录疑似尾标签的内容
|
||
if (endTagBuffer) {
|
||
endTagBuffer += content;
|
||
if (endTagBuffer.includes(endTag)) {
|
||
isInThinkTag = false;
|
||
const answer = endTagBuffer.slice(endTag.length);
|
||
return {
|
||
reasoningContent: '',
|
||
content: answer,
|
||
finishReason
|
||
};
|
||
} else if (endTagBuffer.length >= endTag.length) {
|
||
// 缓存内容超出尾标签长度,且仍未命中 </think>,则认为本次猜测 </think> 失败,仍处于 think 阶段。
|
||
const tmp = endTagBuffer;
|
||
endTagBuffer = '';
|
||
return {
|
||
reasoningContent: tmp,
|
||
content: '',
|
||
finishReason
|
||
};
|
||
}
|
||
return {
|
||
reasoningContent: '',
|
||
content: '',
|
||
finishReason
|
||
};
|
||
} else if (content.includes(endTag)) {
|
||
// 返回内容,完整命中</think>,直接结束
|
||
isInThinkTag = false;
|
||
const [think, answer] = content.split(endTag);
|
||
return {
|
||
reasoningContent: think,
|
||
content: answer,
|
||
finishReason
|
||
};
|
||
} else {
|
||
// 无 buffer,且未命中 </think>,开始疑似 </think> 检测。
|
||
for (let i = 1; i < endTag.length; i++) {
|
||
const partialEndTag = endTag.slice(0, i);
|
||
// 命中一部分尾标签
|
||
if (content.endsWith(partialEndTag)) {
|
||
const think = content.slice(0, -partialEndTag.length);
|
||
endTagBuffer += partialEndTag;
|
||
return {
|
||
reasoningContent: think,
|
||
content: '',
|
||
finishReason
|
||
};
|
||
}
|
||
}
|
||
}
|
||
|
||
// 完全未命中尾标签,还是 think 阶段。
|
||
return {
|
||
reasoningContent: content,
|
||
content: '',
|
||
finishReason
|
||
};
|
||
};
|
||
|
||
const getStartTagBuffer = () => startTagBuffer;
|
||
|
||
return {
|
||
parsePart,
|
||
getStartTagBuffer
|
||
};
|
||
};
|