Skip to main content

Tutorial Part 4 - Document Q&A

One of the most powerful capabilities of Large Language Models is the ability to answer questions about the contents of a set of documents. AI.JSX provides a powerful component called <DocsQA> that can answer a question about a corpus of documents using the LLM.

Here's an example of how this will look in your app:

<DocsQA question="When did the storm occur?" corpus={corpus} chunkLimit={5} chunkFormatter={GetChunk} />

How does Documents Q&A work?

There are many ways of using LLMs to perform some form of document question and answering. The particular approach implemented by DocsQA is described as follows.

First, we construct a corpus from a set of documents -- which might be plain text, HTML, Markdown, or some other format. We break the source documents into a series of chunks that are individually processed by the LLM. Each chunk is passed into the LLM, and the text embedding representing the chunk is computed. (For more details on text embeddings, check out the OpenAI text embeddings API docs.) The text embedding is a multidimensional vector, calculated by the LLM, representing the semantic content of the chunk. A corpus consists of a set of document chunks and their corresponding embedding vectors. For a more detailed explanation of how a corpus is prepared and used, see the guide DocsQA: Grounding Answers with a Source of Truth.

When a question is posed to the corpus, we first pass the question to the LLM and extract its own embedding vector. The chunks with the highest cosine similarity to the question are then gathered from the corpus, and the LLM is invoked with the contents of the most relevant document chunks and the original question.

The implementation of <DocsQA> in AI.JSX currently stores the chunks and embedding vectors in memory. One could also use a vector database to store the embedding vectors, and in the near future we'll be adding support for this to AI.JSX.

The <DocsQA> component takes in several props:

  • question - the query to submit to the LLM about this document corpus.
  • corpus - the corpus of documents to search for the answer.
  • chunkLimit - the maximum number of document chunks to consider in formulating the response.
  • chunkFormatter - a component that will format each document chunk as it is presented to the LLM.

Creating a corpus

The first step is to create a corpus that you want the LLM to answer questions about. In this example, we'll create a corpus that consists of a single Wikipedia article, one about Hurricane Katrina.

// Fetch the HTML for the Wikipedia article.
const URL = 'https://en.wikipedia.org/wiki/Hurricane_Katrina';
const html = await fetch(URL).then((response) => response.text());

// Convert the HTML to Markdown using the Turndown library.
const turndownService = new TurndownService();
const markdown = turndownService.turndown(html);

// Create a LocalCorpus from this content.
const docs = [
{
pageContent: [markdown],
name: 'Wikipedia Article about Hurricane Katrina',
},
];
const corpus = new LocalCorpus(staticLoader(docs), makeChunker(600, 100));
await corpus.load();

We first fetch the contents of the web page and convert it to Markdown using the Turndown library. We then create a LocalCorpus from this Markdown content.

Calling corpus.load() will cause the corpus to be loaded into memory. This is an async operation, since loading the documents and generating vectors can take a while, but for this demo we await the result.

Querying the corpus

Once we have a corpus, we use the <DocsQA> component to query it:

function GetChunk({ chunk }: { chunk: ScoredChunk }) {
return chunk.chunk.content;
}

function App() {
return (
<>
<DocsQA question="What was Hurricane Katrina?" corpus={corpus} chunkLimit={5} chunkFormatter={GetChunk} />
{'\n\n'}
<DocsQA question="Which dates did the storm occur?" corpus={corpus} chunkLimit={5} chunkFormatter={GetChunk} />
{'\n\n'}
<DocsQA
question="Where were the strongest winds reported?"
corpus={corpus}
chunkLimit={5}
chunkFormatter={GetChunk}
/>
</>
);
}

The <GetChunk> component is used to format each document chunk as it is presented to the LLM. In this case, we just return the raw text of the chunk, but if we wanted to include some additional metadata, or transformed the chunks in some way before processing them, we could have done so.

Using a Pinecone database

Until now, we have been using an in-memory Corpus, which is good for a demo, but in practice you might want to use a vector database like Pinecone or Chroma instead. A vector database allows you to scale to much larger datasets.

To do so, you can use LangchainCorpus to integrate with any VectorStore from LangChain.js:

const corpus = new LangchainCorpus(await getVectorStore());

Here is an example where we build a DocsQA component from an existing Pinecone index:

import { PineconeClient } from '@pinecone-database/pinecone';
import { PineconeStore } from 'langchain/vectorstores/pinecone';
import { OpenAIEmbeddings } from 'langchain/embeddings/openai';

const client = new PineconeClient();
await client.init({
apiKey: process.env.PINECONE_API_KEY,
environment: process.env.PINECONE_ENVIRONMENT,
});

const vectorStore = await PineconeStore.fromExistingIndex(new OpenAIEmbeddings(), {
pineconeIndex: client.Index(process.env.PINECONE_INDEX);,
namespace: process.env.PINECONE_NAMESPACE,
});

Once we have a VectorStore object, we wrap it with a LangchainCorpus:

const corpus = new LangchainCorpus(vectorStore);

The above assumes that the vector store is already populated. To load documents, you can use a LoadableLangchainCorpus:

const corpus = new LoadableLangchainCorpus(vectorStore, staticLoader(docs), makeChunker(600, 100));
// Loading docs into the Piencone database.
await corpus.load();

Once you have the corpus object, you can ask questions from it just as before:

function App() {
return (
<>
<DocsQA question="What was Hurricane Katrina?" corpus={corpus} chunkLimit={5} chunkFormatter={GetChunk} />
{'\n\n'}
<DocsQA question="Which dates did the storm occur?" corpus={corpus} chunkLimit={5} chunkFormatter={GetChunk} />
</>
);
}