(Cont.) LangChain.js: Chatting with a PDF

Now, that we have done with the retriever module, the next steps are:

Step 4: Rephrase the Question

The question has to be reformed to be a standalone question, so that it can be understood independently without any prior knowledge. This ensures that the question is clear and can be answered accurately.

In the below code, we see a RunnableSequence, which takes in the prompt, passes it to the Google Model, and converts the output to a string.

import { RunnableWithMessageHistory } from "@langchain/core/runnables"; 
import { ChatPromptTemplate, MessagesPlaceholder } from "@langchain/core/prompts"; 

export function createRephraseQuestionChain() {
    const REPHRASE_QUESTION_SYSTEM_TEMPLATE = `
    Given the following conversation and a follow up question,
    rephrase the follow up question to be a standalone question.
    `;

    // Rephrase the question to be a standalone question
    // along with passing the chat history
    const rephraseQuestionChainPrompt = ChatPromptTemplate.fromMessages([
        ["system", REPHRASE_QUESTION_SYSTEM_TEMPLATE],
        new MessagesPlaceholder("history"),
        ["human", "Rephrase the following question as a standalone question:\n{question}"],
    ]);

    // Runnable to rephrase the question
    const rephraseQuestionChain = RunnableSequence.from([
        rephraseQuestionChainPrompt,
        new ChatGoogleGenerativeAI({
            modelName: "gemini-pro",
            maxOutputTokens: 2048,
            apiKey: process.env.google_api_key
        }),
        new StringOutputParser(),
    ]);
    return rephraseQuestionChain;
}

const rephraseQuestionChain = createRephraseQuestionChain();

Technical Terms:

• StringOutputParser: Converts the incoming input to a String.

Step 5: Create a prompt template for the response and initialize the chain to retrieve the response

We make a prompt template for the response, so that the LLM answers according to our use case. Using all our prompts we initialize a RunnableSequence to run all our prompts in sequence.

// make a prompt template for the response
const ANSWER_CHAIN_SYSTEM_TEMPLATE = `You are an experienced researcher,
expert at interpreting and answering questions based on provided sources.
Using the below provided context and chat history, 
answer the user's question to the best of your ability
using only the resources provided. Be verbose!

<context>
{context}
</context>`;

// Creat a prompt template from Messages
const answerGenerationChainPrompt = ChatPromptTemplate.fromMessages([
    ["system", ANSWER_CHAIN_SYSTEM_TEMPLATE],
    new MessagesPlaceholder("history"),
    [
        "human",
        `Now, answer this question using the previous context and chat history:

    {standalone_question}`
    ]
]);

// initialize the chain to retrieve the answer for the question
const conversationalRetrievalChain = RunnableSequence.from([
    RunnablePassthrough.assign({
        standalone_question: rephraseQuestionChain,
    }),
    RunnablePassthrough.assign({
        context: documentRetrievalChain,
    }),
    answerGenerationChainPrompt,
    new ChatGoogleGenerativeAI({
        modelName: "gemini-pro",
        maxOutputTokens: 2048,
        apiKey: process.env.google_api_key
    }),
]);

Technical Terms:

• RunnablePassthrough: It is a component or function in the Langchain framework that allows the passage of data or information without any modification or processing. It acts as a placeholder or a pass-through mechanism in a chain of operations or transformations.

Step 6: Initialize the chat history and run the final chain.

The final part is we should get the chat history and pass all the initialized runnables to the final chain.

import { ChatMessageHistory } from "langchain/stores/message/in_memory";

// we should create a new history object per session
const messageHistories = {};
const getMessageHistoryForSession = (sessionId) => {
    if (messageHistories[sessionId] !== undefined) {
        return messageHistories[sessionId];
    }
    const newChatSessionHistory = new ChatMessageHistory();
    messageHistories[sessionId] = newChatSessionHistory;
    return newChatSessionHistory;
};
const finalRetrievalChain = new RunnableWithMessageHistory({
    runnable: conversationalRetrievalChain,
    getMessageHistory: getMessageHistoryForSession,
    inputMessagesKey: "question",
    historyMessagesKey: "history",
}).pipe(new StringOutputParser())

const result = await finalRetrievalChain.invoke({
  question: "What are the pre-requisites of this course?",
}, {
  configurable: { sessionId: "test" }
});

console.log(result)

//Input: Our loaded PDF has details about a Machine Learning Course by Andrew NG
//Output:
/* 
To enroll in this class, you must meet the following prerequisites:

* Basic probability and statistics. A course like Statistics 116 would be more than enough.

* Basic linear algebra. Courses like Math 51, 103, 113, or CS205 at Stanford are recommended. Matrix and vector multiplication, matrix inverse, and eigenvectors are concepts you should be familiar with.

* Programming skills. You should be able to write a simple computer program.

* Understanding of big O notation and data structures such as queues, stacks, and binary trees.

If you have taken these courses or have equivalent knowledge, you should be prepared for this class. However, it is important to note that while programming will not be a major focus of the course, some programming in MATLAB or Octave will be required.
*/