feat(core): QAChain

packages/core/.gitignore

@@ -4,6 +4,9 @@ chain.d.ts
 chain/saveWebInfo.cjs
 chain/saveWebInfo.js
 chain/saveWebInfo.d.ts
+chain/qa.cjs
+chain/qa.js
+chain/qa.d.ts
 llm.cjs
 llm.js
 llm.d.ts

packages/core/package.json

@@ -29,6 +29,11 @@
       "require": "./chain/saveWebInfo.cjs",
       "import": "./chain/saveWebInfo.js"
     },
+    "./chain/qa": {
+      "types": "./chain/qa.d.ts",
+      "require": "./chain/qa.cjs",
+      "import": "./chain/qa.js"
+    },
     "./llm": {
       "types": "./llm.d.ts",
       "require": "./llm.cjs",
@@ -82,6 +87,9 @@
     "chain/saveWebInfo.cjs",
     "chain/saveWebInfo.js",
     "chain/saveWebInfo.d.ts",
+    "chain/qa.cjs",
+    "chain/qa.js",
+    "chain/qa.d.ts",
     "llm.cjs",
     "llm.js",
     "llm.d.ts",
@@ -123,9 +131,9 @@
     "test": "vitest"
   },
   "peerDependencies": {
-    "@supabase/supabase-js": "^2.22.0",
+    "@supabase/supabase-js": "^2.26.0",
     "langchain": "^0.0.96",
-    "ofetch": "^1.0.1"
+    "ofetch": "^1.1.1"
   },
   "peerDependenciesMeta": {
     "@supabase/supabase-js": {

packages/core/scripts/create-entrypoints.js

@@ -11,6 +11,7 @@ import * as path from 'node:path'
 export const entrypoints = {
   'chain': 'chain/index',
   'chain/saveWebInfo': 'chain/saveWebInfo',
+  'chain/qa': 'chain/qa',
   'llm': 'llm/index',
   'llm/openai': 'llm/openai/index',
   'storage': 'storage/index',

packages/core/src/chain/qa.ts

@@ -0,0 +1,49 @@
+import { RetrievalQAChain } from 'langchain/chains'
+import { PromptTemplate } from 'langchain/prompts'
+import type { IVectorStorage } from '../storage'
+import type { PromptsLanguage, QARes } from '../types'
+import { ANSWER_IN_LANGUAGE } from '../const'
+import { getPromptsByTemplate } from '../utils'
+
+const prompt_template = `Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.
+
+{context}
+
+Question: {question}
+Answer in {language}:`
+
+export class QAChain {
+  constructor(fields: {
+    vectorStorage: IVectorStorage
+  }) {
+    this.vectorStorage = fields.vectorStorage
+  }
+
+  private vectorStorage
+
+  async call(question: string, language: PromptsLanguage = 'zh-CN') {
+    const retriever = await this.vectorStorage.getRetriever()
+    const config = this.vectorStorage.getConfig()
+
+    const kv = {
+      language: ANSWER_IN_LANGUAGE[language],
+    }
+    const template = getPromptsByTemplate(prompt_template, kv)
+
+    const chain = RetrievalQAChain.fromLLM(
+      config.embeddingsInfo.llmModel,
+      retriever,
+      {
+        returnSourceDocuments: true,
+        prompt: new PromptTemplate({
+          template,
+          inputVariables: ['context', 'question'],
+        }),
+      },
+    )
+
+    const res = await chain.call({ query: question })
+
+    return res as QARes
+  }
+}

packages/core/src/llm/openai/index.ts

@@ -1,5 +1,6 @@
 import { $fetch } from 'ofetch'
 import { OpenAIEmbeddings } from 'langchain/embeddings/openai'
+import { ChatOpenAI } from 'langchain/chat_models/openai'
 import type { PromptsLanguage, SummarizeData, WebInfoData } from '../../types'
 import { countWord, getPromptsByTemplate, preprocessText } from '../../utils'
 import { ANSWER_IN_LANGUAGE, OPENAI_CHAT_API, SUMMARIZE_PROMPTS, USER_PROMPTS } from '../../const'
@@ -24,7 +25,11 @@ export class Openai extends CLLM<OpenaiConfig> {
     const embeddings = new OpenAIEmbeddings({ openAIApiKey: this.config.apiKey }, {
       basePath: `${this.config.apiHost}/v1`,
     })
+    const model = new ChatOpenAI({ openAIApiKey: this.config.apiKey, modelName: 'gpt-3.5-turbo', temperature: 0.2 }, {
+      basePath: `${this.config.apiHost}/v1`,
+    })
     return {
+      llmModel: model,
       embeddings,
       indexName: 'openai_documents',
       queryName: 'openai_match_documents',

packages/core/src/storage/supabase/index.ts

@@ -5,7 +5,7 @@ import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter'
 import { $fetch } from 'ofetch'
 import type { WebCardData, WebInfoData } from '../../types'
 import { ImageStorage, VectorStorage } from '../types'
-import type { SavedVector, StorageImage, StorageType, VectorConfig } from '../types'
+import type { StorageImage, StorageType, VectorConfig } from '../types'
 
 export interface SupabaseImageConfig {
   anonKey: string
@@ -16,7 +16,11 @@ export interface SupabaseImageConfig {
 export class SupabaseImageStorage extends ImageStorage<SupabaseImageConfig, WebCardData> {
   constructor(config: SupabaseImageConfig, data?: StorageImage) {
     super(config, data)
-    this.client = createClient(this.config.url, this.config.anonKey)
+    this.client = createClient(this.config.url, this.config.anonKey, {
+      auth: {
+        persistSession: false,
+      },
+    })
     this.storageUrl = `${this.config.url}/storage/v1/object/public/${this.config.bucket}`
   }
 
@@ -96,10 +100,14 @@ export interface SupabaseVectorConfig extends VectorConfig {
   anonKey: string
   url: string
 }
-export class SupabaseVectorStorage extends VectorStorage<SupabaseVectorConfig, Document<SavedVector>[]> {
+export class SupabaseVectorStorage extends VectorStorage<SupabaseVectorConfig> {
   constructor(config: SupabaseVectorConfig, data?: WebInfoData) {
     super(config, data)
-    this.client = createClient(this.config.url, this.config.anonKey)
+    this.client = createClient(this.config.url, this.config.anonKey, {
+      auth: {
+        persistSession: false,
+      },
+    })
   }
 
   private client
@@ -125,22 +133,15 @@ export class SupabaseVectorStorage extends VectorStorage<SupabaseVectorConfig, D
     await store.addDocuments(docs)
   }
 
-  async query(question: string) {
+  async getRetriever() {
     const vectorStore = await SupabaseVectorStore.fromExistingIndex(
       this.config.embeddingsInfo.embeddings,
       {
         client: this.client,
         tableName: this.config.embeddingsInfo.indexName,
         queryName: this.config.embeddingsInfo.queryName,
       })
-
-    const result = await vectorStore.similaritySearch(question, 2, {
-      appName: this.config.metaData.appName,
-      botId: this.config.metaData.botId,
-      userId: this.config.metaData.userId,
-    })
-
-    return result as Document<SavedVector>[]
+    return vectorStore.asRetriever()
   }
 
   getConfig(): SupabaseVectorConfig {

packages/core/src/storage/types.ts

@@ -1,5 +1,5 @@
-import type { Document } from 'langchain/document'
-import type { EmbeddingsInfo, SummarizeData, WebInfoData } from '../types'
+import type { BaseRetriever } from 'langchain/schema'
+import type { EmbeddingsInfo, SummarizeData, VectorMetaData, WebInfoData } from '../types'
 
 export type TStorage = 'DataStorage' | 'ImageStorage' | 'VectorStorage'
 export interface StorageType {
@@ -66,27 +66,17 @@ export abstract class ImageStorage<T, R extends SavedImage> implements IImageSto
   abstract getConfig(): T
 }
 
-export interface SavedVector {
-  metadata: VectorMetaData
-  pageContent: string
-}
-export interface VectorMetaData {
-  url: string
-  appName: string
-  botId: string
-  userId: string
-}
 export interface IVectorStorage {
   save(data?: WebInfoData): Promise<void>
-  query(question: string): Promise<Document<SavedVector>[]>
+  getRetriever(): Promise<BaseRetriever>
   getType(): StorageType
-  getConfig(): any
+  getConfig(): VectorConfig
 }
 export interface VectorConfig {
   embeddingsInfo: EmbeddingsInfo
   metaData: VectorMetaData
 }
-export abstract class VectorStorage<T extends VectorConfig, R extends Document<SavedVector>[]> implements IVectorStorage {
+export abstract class VectorStorage<T extends VectorConfig> implements IVectorStorage {
   constructor(config: T, data?: WebInfoData) {
     this.config = config
     this.data = data
@@ -96,7 +86,7 @@ export abstract class VectorStorage<T extends VectorConfig, R extends Document<S
   protected data?: WebInfoData
 
   abstract save(data?: WebInfoData): Promise<void>
-  abstract query(question: string): Promise<R>
+  abstract getRetriever(): Promise<BaseRetriever>
   abstract getType(): StorageType
   abstract getConfig(): T
 }

packages/core/src/types.ts

@@ -1,4 +1,6 @@
 import type { Embeddings } from 'langchain/embeddings/base'
+import type { BaseLanguageModel } from 'langchain/base_language'
+import type { Document } from 'langchain/document'
 
 export interface WebInfoData {
   title: string
@@ -105,7 +107,23 @@ export interface Routes {
 }
 
 export interface EmbeddingsInfo {
+  llmModel: BaseLanguageModel
   embeddings: Embeddings
   indexName: string
   queryName?: string
 }
+
+export interface QARes {
+  text: string
+  sourceDocuments?: Document<VectorMetaData>[]
+}
+export interface SavedVector {
+  metadata: VectorMetaData
+  pageContent: string
+}
+export interface VectorMetaData {
+  source?: string
+  appName: string
+  botId: string
+  userId: string
+}

packages/core/vite.config.ts

@@ -5,6 +5,7 @@ import vue from '@vitejs/plugin-vue'
 export const entrypoints = {
   'chain': 'chain/index',
   'chain/saveWebInfo': 'chain/saveWebInfo',
+  'chain/qa': 'chain/qa',
   'llm': 'llm/index',
   'llm/openai': 'llm/openai/index',
   'storage': 'storage/index',

server/nuxt3/package.json

@@ -22,6 +22,7 @@
   "dependencies": {
     "@stargram/core": "workspace:*",
     "@stargram/web-hub": "workspace:*",
+    "@supabase/supabase-js": "^2.26.0",
     "@vercel/kv": "^0.2.2",
     "@vue-flow/background": "^1.2.0",
     "@vue-flow/controls": "^1.1.0",