Merge pull request #434 from jacoblee93/jacob/update_versions

Update LangChain and Pinecone client, use expression language for chain

README.md

@@ -62,7 +62,7 @@ PINECONE_INDEX_NAME=
 
 1. Inside `docs` folder, add your pdf files or folders that contain pdf files.
 
-2. Run the script `npm run ingest` to 'ingest' and embed your docs. If you run into errors troubleshoot below.
+2. Run the script `yarn run ingest` to 'ingest' and embed your docs. If you run into errors troubleshoot below.
 
 3. Check Pinecone dashboard to verify your namespace and vectors have been added.
 

package.json

@@ -16,11 +16,11 @@
   },
   "dependencies": {
     "@microsoft/fetch-event-source": "^2.0.1",
-    "@pinecone-database/pinecone": "0.0.14",
+    "@pinecone-database/pinecone": "1.1.0",
     "@radix-ui/react-accordion": "^1.1.1",
     "clsx": "^1.2.1",
     "dotenv": "^16.0.3",
-    "langchain": "^0.0.125",
+    "langchain": "^0.0.186",
     "lucide-react": "^0.125.0",
     "next": "13.2.3",
     "pdf-parse": "1.1.1",

pages/_document.tsx

@@ -1,4 +1,4 @@
-import { Html, Head, Main, NextScript } from "next/document";
+import { Html, Head, Main, NextScript } from 'next/document';
 
 export default function Document() {
   return (

pages/api/chat.ts

@@ -1,7 +1,7 @@
 import type { NextApiRequest, NextApiResponse } from 'next';
+import type { Document } from 'langchain/document';
 import { OpenAIEmbeddings } from 'langchain/embeddings/openai';
 import { PineconeStore } from 'langchain/vectorstores/pinecone';
-import { AIMessage, HumanMessage } from 'langchain/schema';
 import { makeChain } from '@/utils/makechain';
 import { pinecone } from '@/utils/pinecone-client';
 import { PINECONE_INDEX_NAME, PINECONE_NAME_SPACE } from '@/config/pinecone';
@@ -40,25 +40,41 @@ export default async function handler(
       },
     );
 
+    // Use a callback to get intermediate sources from the middle of the chain
+    let resolveWithDocuments: (value: Document[]) => void;
+    const documentPromise = new Promise<Document[]>((resolve) => {
+      resolveWithDocuments = resolve;
+    });
+    const retriever = vectorStore.asRetriever({
+      callbacks: [
+        {
+          handleRetrieverEnd(documents) {
+            resolveWithDocuments(documents);
+          },
+        },
+      ],
+    });
+
     //create chain
-    const chain = makeChain(vectorStore);
+    const chain = makeChain(retriever);
 
-    const pastMessages = history.map((message: string, i: number) => {
-      if (i % 2 === 0) {
-        return new HumanMessage(message);
-      } else {
-        return new AIMessage(message);
-      }
-    });
+    const pastMessages = history
+      .map((message: [string, string]) => {
+        return [`Human: ${message[0]}`, `Assistant: ${message[1]}`].join('\n');
+      })
+      .join('\n');
+    console.log(pastMessages);
 
     //Ask a question using chat history
-    const response = await chain.call({
+    const response = await chain.invoke({
       question: sanitizedQuestion,
-      chat_history: pastMessages
+      chat_history: pastMessages,
     });
 
+    const sourceDocuments = await documentPromise;
+
     console.log('response', response);
-    res.status(200).json(response);
+    res.status(200).json({ text: response, sourceDocuments });
   } catch (error: any) {
     console.log('error', error);
     res.status(500).json({ error: error.message || 'Something went wrong' });

utils/customPDFLoader.ts

@@ -1,6 +1,6 @@
 import { Document } from 'langchain/document';
 import { readFile } from 'fs/promises';
-import { BaseDocumentLoader } from 'langchain/document_loaders';
+import { BaseDocumentLoader } from 'langchain/document_loaders/base';
 
 export abstract class BufferLoader extends BaseDocumentLoader {
   constructor(public filePathOrBlob: string | Blob) {

utils/makechain.ts

@@ -1,37 +1,85 @@
 import { ChatOpenAI } from 'langchain/chat_models/openai';
-import { PineconeStore } from 'langchain/vectorstores/pinecone';
-import { ConversationalRetrievalQAChain } from 'langchain/chains';
+import { ChatPromptTemplate } from 'langchain/prompts';
+import { RunnableSequence } from 'langchain/schema/runnable';
+import { StringOutputParser } from 'langchain/schema/output_parser';
+import type { Document } from 'langchain/document';
+import type { VectorStoreRetriever } from 'langchain/vectorstores/base';
 
 const CONDENSE_TEMPLATE = `Given the following conversation and a follow up question, rephrase the follow up question to be a standalone question.
 
-Chat History:
-{chat_history}
+<chat_history>
+  {chat_history}
+</chat_history>
+
 Follow Up Input: {question}
 Standalone question:`;
 
-const QA_TEMPLATE = `You are a helpful AI assistant. Use the following pieces of context to answer the question at the end.
+const QA_TEMPLATE = `You are an expert researcher. Use the following pieces of context to answer the question at the end.
 If you don't know the answer, just say you don't know. DO NOT try to make up an answer.
-If the question is not related to the context, politely respond that you are tuned to only answer questions that are related to the context.
+If the question is not related to the context or chat history, politely respond that you are tuned to only answer questions that are related to the context.
+
+<context>
+  {context}
+</context>
 
-{context}
+<chat_history>
+  {chat_history}
+</chat_history>
 
 Question: {question}
 Helpful answer in markdown:`;
 
-export const makeChain = (vectorstore: PineconeStore) => {
+const combineDocumentsFn = (docs: Document[], separator = '\n\n') => {
+  const serializedDocs = docs.map((doc) => doc.pageContent);
+  return serializedDocs.join(separator);
+};
+
+export const makeChain = (retriever: VectorStoreRetriever) => {
+  const condenseQuestionPrompt =
+    ChatPromptTemplate.fromTemplate(CONDENSE_TEMPLATE);
+  const answerPrompt = ChatPromptTemplate.fromTemplate(QA_TEMPLATE);
+
   const model = new ChatOpenAI({
-    temperature: 0, // increase temepreature to get more creative answers
+    temperature: 0, // increase temperature to get more creative answers
     modelName: 'gpt-3.5-turbo', //change this to gpt-4 if you have access
   });
 
-  const chain = ConversationalRetrievalQAChain.fromLLM(
+  // Rephrase the initial question into a dereferenced standalone question based on
+  // the chat history to allow effective vectorstore querying.
+  const standaloneQuestionChain = RunnableSequence.from([
+    condenseQuestionPrompt,
+    model,
+    new StringOutputParser(),
+  ]);
+
+  // Retrieve documents based on a query, then format them.
+  const retrievalChain = retriever.pipe(combineDocumentsFn);
+
+  // Generate an answer to the standalone question based on the chat history
+  // and retrieved documents. Additionally, we return the source documents directly.
+  const answerChain = RunnableSequence.from([
+    {
+      context: RunnableSequence.from([
+        (input) => input.question,
+        retrievalChain,
+      ]),
+      chat_history: (input) => input.chat_history,
+      question: (input) => input.question,
+    },
+    answerPrompt,
     model,
-    vectorstore.asRetriever(),
+    new StringOutputParser(),
+  ]);
+
+  // First generate a standalone question, then answer it based on
+  // chat history and retrieved context documents.
+  const conversationalRetrievalQAChain = RunnableSequence.from([
     {
-      qaTemplate: QA_TEMPLATE,
-      questionGeneratorTemplate: CONDENSE_TEMPLATE,
-      returnSourceDocuments: true, //The number of source documents returned is 4 by default
+      question: standaloneQuestionChain,
+      chat_history: (input) => input.chat_history,
     },
-  );
-  return chain;
+    answerChain,
+  ]);
+
+  return conversationalRetrievalQAChain;
 };

utils/pinecone-client.ts

@@ -1,14 +1,12 @@
-import { PineconeClient } from '@pinecone-database/pinecone';
+import { Pinecone } from '@pinecone-database/pinecone';
 
 if (!process.env.PINECONE_ENVIRONMENT || !process.env.PINECONE_API_KEY) {
   throw new Error('Pinecone environment or api key vars missing');
 }
 
 async function initPinecone() {
   try {
-    const pinecone = new PineconeClient();
-
-    await pinecone.init({
+    const pinecone = new Pinecone({
       environment: process.env.PINECONE_ENVIRONMENT ?? '', //this is in the dashboard
       apiKey: process.env.PINECONE_API_KEY ?? '',
     });