perf: chat framwork

This commit is contained in:
archer
2023-05-03 15:28:25 +08:00
parent 91decc3683
commit 00a99261ae
23 changed files with 811 additions and 1011 deletions

View File

@@ -0,0 +1,104 @@
import { PgClient } from '@/service/pg';
import { ModelDataStatusEnum, ModelVectorSearchModeEnum, ChatModelMap } from '@/constants/model';
import { ModelSchema } from '@/types/mongoSchema';
import { openaiCreateEmbedding } from '../utils/chat/openai';
import { ChatRoleEnum } from '@/constants/chat';
import { sliceTextByToken } from '@/utils/chat';
/**
* use openai embedding search kb
*/
export const searchKb = async ({
userApiKey,
systemApiKey,
text,
similarity = 0.2,
model,
userId
}: {
userApiKey?: string;
systemApiKey: string;
text: string;
model: ModelSchema;
userId: string;
similarity?: number;
}): Promise<{
code: 200 | 201;
searchPrompt?: {
obj: `${ChatRoleEnum}`;
value: string;
};
}> => {
const modelConstantsData = ChatModelMap[model.chat.chatModel];
// 获取提示词的向量
const { vector: promptVector } = await openaiCreateEmbedding({
userApiKey,
systemApiKey,
userId,
text
});
const vectorSearch = await PgClient.select<{ q: string; a: string }>('modelData', {
fields: ['q', 'a'],
where: [
['status', ModelDataStatusEnum.ready],
'AND',
['model_id', model._id],
'AND',
`vector <=> '[${promptVector}]' < ${similarity}`
],
order: [{ field: 'vector', mode: `<=> '[${promptVector}]'` }],
limit: 20
});
const systemPrompts: string[] = vectorSearch.rows.map((item) => `${item.q}\n${item.a}`);
// filter system prompt
if (
systemPrompts.length === 0 &&
model.chat.searchMode === ModelVectorSearchModeEnum.hightSimilarity
) {
return {
code: 201,
searchPrompt: {
obj: ChatRoleEnum.AI,
value: '对不起,你的问题不在知识库中。'
}
};
}
/* 高相似度+无上下文,不添加额外知识,仅用系统提示词 */
if (systemPrompts.length === 0 && model.chat.searchMode === ModelVectorSearchModeEnum.noContext) {
return {
code: 200,
searchPrompt: model.chat.systemPrompt
? {
obj: ChatRoleEnum.System,
value: model.chat.systemPrompt
}
: undefined
};
}
// 有匹配情况下system 添加知识库内容。
// 系统提示词过滤,最多 65% tokens
const filterSystemPrompt = sliceTextByToken({
model: model.chat.chatModel,
text: systemPrompts.join('\n'),
length: Math.floor(modelConstantsData.contextMaxToken * 0.65)
});
return {
code: 200,
searchPrompt: {
obj: ChatRoleEnum.System,
value: `
${model.chat.systemPrompt}
${
model.chat.searchMode === ModelVectorSearchModeEnum.hightSimilarity ? `不回答知识库外的内容.` : ''
}
知识库内容为: ${filterSystemPrompt}'
`
}
};
};