Build a Query Analysis System

Prerequisites

This guide assumes familiarity with the following concepts:

• Document loaders
• Chat models
• Embeddings
• Vector stores
• Retrieval

This page will show how to use query analysis in a basic end-to-end example. This will cover creating a simple search engine, showing a failure mode that occurs when passing a raw user question to that search, and then an example of how query analysis can help address that issue. There are MANY different query analysis techniques and this end-to-end example will not show all of them.

For the purpose of this example, we will do retrieval over the LangChain YouTube videos.

Setup

Install dependencies

tip

See this section for general instructions on installing integration packages.

npm

npm i langchain @langchain/community @langchain/openai youtubei.js chromadb youtube-transcript

yarn

yarn add langchain @langchain/community @langchain/openai youtubei.js chromadb youtube-transcript

pnpm

pnpm add langchain @langchain/community @langchain/openai youtubei.js chromadb youtube-transcript

Set environment variables

We’ll use OpenAI in this example:

OPENAI_API_KEY=your-api-key

# Optional, use LangSmith for best-in-class observability
LANGSMITH_API_KEY=your-api-key
LANGCHAIN_TRACING_V2=true

Load documents

We can use the YouTubeLoader to load transcripts of a few LangChain videos:

import { DocumentInterface } from "@langchain/core/documents";
import { YoutubeLoader } from "@langchain/community/document_loaders/web/youtube";
import { getYear } from "date-fns";

const urls = [
  "https://www.youtube.com/watch?v=HAn9vnJy6S4",
  "https://www.youtube.com/watch?v=dA1cHGACXCo",
  "https://www.youtube.com/watch?v=ZcEMLz27sL4",
  "https://www.youtube.com/watch?v=hvAPnpSfSGo",
  "https://www.youtube.com/watch?v=EhlPDL4QrWY",
  "https://www.youtube.com/watch?v=mmBo8nlu2j0",
  "https://www.youtube.com/watch?v=rQdibOsL1ps",
  "https://www.youtube.com/watch?v=28lC4fqukoc",
  "https://www.youtube.com/watch?v=es-9MgxB-uc",
  "https://www.youtube.com/watch?v=wLRHwKuKvOE",
  "https://www.youtube.com/watch?v=ObIltMaRJvY",
  "https://www.youtube.com/watch?v=DjuXACWYkkU",
  "https://www.youtube.com/watch?v=o7C9ld6Ln-M",
];

let docs: Array<DocumentInterface> = [];

for await (const url of urls) {
  const doc = await YoutubeLoader.createFromUrl(url, {
    language: "en",
    addVideoInfo: true,
  }).load();
  docs = docs.concat(doc);
}

console.log(docs.length);
/*
13
 */

// Add some additional metadata: what year the video was published
// The JS API does not provide publish date, so we can use a
// hardcoded array with the dates instead.
const dates = [
  new Date("Jan 31, 2024"),
  new Date("Jan 26, 2024"),
  new Date("Jan 24, 2024"),
  new Date("Jan 23, 2024"),
  new Date("Jan 16, 2024"),
  new Date("Jan 5, 2024"),
  new Date("Jan 2, 2024"),
  new Date("Dec 20, 2023"),
  new Date("Dec 19, 2023"),
  new Date("Nov 27, 2023"),
  new Date("Nov 22, 2023"),
  new Date("Nov 16, 2023"),
  new Date("Nov 2, 2023"),
];
docs.forEach((doc, idx) => {
  // eslint-disable-next-line no-param-reassign
  doc.metadata.publish_year = getYear(dates[idx]);
  // eslint-disable-next-line no-param-reassign
  doc.metadata.publish_date = dates[idx];
});

// Here are the titles of the videos we've loaded:
console.log(docs.map((doc) => doc.metadata.title));
/*
[
  'OpenGPTs',
  'Building a web RAG chatbot: using LangChain, Exa (prev. Metaphor), LangSmith, and Hosted Langserve',
  'Streaming Events: Introducing a new `stream_events` method',
  'LangGraph: Multi-Agent Workflows',
  'Build and Deploy a RAG app with Pinecone Serverless',
  'Auto-Prompt Builder (with Hosted LangServe)',
  'Build a Full Stack RAG App With TypeScript',
  'Getting Started with Multi-Modal LLMs',
  'SQL Research Assistant',
  'Skeleton-of-Thought: Building a New Template from Scratch',
  'Benchmarking RAG over LangChain Docs',
  'Building a Research Assistant from Scratch',
  'LangServe and LangChain Templates Webinar'
] 
 */

API Reference:

• DocumentInterface from @langchain/core/documents
• YoutubeLoader from @langchain/community/document_loaders/web/youtube

Here’s the metadata associated with each video.

We can see that each document also has a title, view count, publication date, and length:

import { getDocs } from "./docs.js";

const docs = await getDocs();

console.log(docs[0].metadata);

/**
{
  source: 'HAn9vnJy6S4',
  description: 'OpenGPTs is an open-source platform aimed at recreating an experience like the GPT Store - but with any model, any tools, and that you can self-host.\n' +
    '\n' +
    'This video covers both how to use it as well as how to build it.\n' +
    '\n' +
    'GitHub: https://github.com/langchain-ai/opengpts',
  title: 'OpenGPTs',
  view_count: 7262,
  author: 'LangChain'
}
 */

// And here's a sample from a document's contents:

console.log(docs[0].pageContent.slice(0, 500));

/*
hello today I want to talk about open gpts open gpts is a project that we built here at linkchain uh that replicates the GPT store in a few ways so it creates uh end user-facing friendly interface to create different Bots and these Bots can have access to different tools and they can uh be given files to retrieve things over and basically it's a way to create a variety of bots and expose the configuration of these Bots to end users it's all open source um it can be used with open AI it can be us
 */

API Reference:

Indexing documents

Whenever we perform retrieval we need to create an index of documents that we can query. We’ll use a vector store to index our documents, and we’ll chunk them first to make our retrievals more concise and precise:

import { RecursiveCharacterTextSplitter } from "@langchain/textsplitters";
import { OpenAIEmbeddings } from "@langchain/openai";
import { Chroma } from "@langchain/community/vectorstores/chroma";
import { getDocs } from "./docs.js";

const docs = await getDocs();
const textSplitter = new RecursiveCharacterTextSplitter({ chunkSize: 2000 });
const chunkedDocs = await textSplitter.splitDocuments(docs);
const embeddings = new OpenAIEmbeddings({
  model: "text-embedding-3-small",
});
const vectorStore = await Chroma.fromDocuments(chunkedDocs, embeddings, {
  collectionName: "yt-videos",
});

API Reference:

• RecursiveCharacterTextSplitter from @langchain/textsplitters
• OpenAIEmbeddings from @langchain/openai
• Chroma from @langchain/community/vectorstores/chroma

Then later, you can retrieve the index without having to re-query and embed:

import "chromadb";
import { OpenAIEmbeddings } from "@langchain/openai";
import { Chroma } from "@langchain/community/vectorstores/chroma";

const embeddings = new OpenAIEmbeddings({
  model: "text-embedding-3-small",
});
const vectorStore = await Chroma.fromExistingCollection(embeddings, {
  collectionName: "yt-videos",
});

[Module: null prototype] {
  AdminClient: [class AdminClient],
  ChromaClient: [class ChromaClient],
  CloudClient: [class CloudClient extends ChromaClient],
  CohereEmbeddingFunction: [class CohereEmbeddingFunction],
  Collection: [class Collection],
  DefaultEmbeddingFunction: [class _DefaultEmbeddingFunction],
  GoogleGenerativeAiEmbeddingFunction: [class _GoogleGenerativeAiEmbeddingFunction],
  HuggingFaceEmbeddingServerFunction: [class HuggingFaceEmbeddingServerFunction],
  IncludeEnum: {
    Documents: "documents",
    Embeddings: "embeddings",
    Metadatas: "metadatas",
    Distances: "distances"
  },
  JinaEmbeddingFunction: [class JinaEmbeddingFunction],
  OpenAIEmbeddingFunction: [class _OpenAIEmbeddingFunction],
  TransformersEmbeddingFunction: [class _TransformersEmbeddingFunction]
}

Retrieval without query analysis

We can perform similarity search on a user question directly to find chunks relevant to the question:

const searchResults = await vectorStore.similaritySearch(
  "how do I build a RAG agent"
);
console.log(searchResults[0].metadata.title);
console.log(searchResults[0].pageContent.slice(0, 500));

OpenGPTs
hardcoded that it will always do a retrieval step here the assistant decides whether to do a retrieval step or not sometimes this is good sometimes this is bad sometimes it you don't need to do a retrieval step when I said hi it didn't need to call it tool um but other times you know the the llm might mess up and not realize that it needs to do a retrieval step and so the rag bot will always do a retrieval step so it's more focused there because this is also a simpler architecture so it's always

This works pretty okay! Our first result is somewhat relevant to the question.

What if we wanted to search for results from a specific time period?

const searchResults = await vectorStore.similaritySearch(
  "videos on RAG published in 2023"
);
console.log(searchResults[0].metadata.title);
console.log(searchResults[0].metadata.publish_year);
console.log(searchResults[0].pageContent.slice(0, 500));

OpenGPTs
2024
hardcoded that it will always do a retrieval step here the assistant decides whether to do a retrieval step or not sometimes this is good sometimes this is bad sometimes it you don't need to do a retrieval step when I said hi it didn't need to call it tool um but other times you know the the llm might mess up and not realize that it needs to do a retrieval step and so the rag bot will always do a retrieval step so it's more focused there because this is also a simpler architecture so it's always

Our first result is from 2024, and not very relevant to the input. Since we’re just searching against document contents, there’s no way for the results to be filtered on any document attributes.

This is just one failure mode that can arise. Let’s now take a look at how a basic form of query analysis can fix it!

Query analysis

To handle these failure modes we’ll do some query structuring. This will involve defining a query schema that contains some date filters and use a function-calling model to convert a user question into a structured queries.

Query schema

In this case we’ll have explicit min and max attributes for publication date so that it can be filtered on.

import { z } from "zod";

const searchSchema = z
  .object({
    query: z
      .string()
      .describe("Similarity search query applied to video transcripts."),
    publish_year: z.number().optional().describe("Year of video publication."),
  })
  .describe(
    "Search over a database of tutorial videos about a software library."
  );

Query generation

To convert user questions to structured queries we’ll make use of OpenAI’s function-calling API. Specifically we’ll use the new ChatModel.withStructuredOutput() constructor to handle passing the schema to the model and parsing the output.

import { ChatPromptTemplate } from "@langchain/core/prompts";
import { ChatOpenAI } from "@langchain/openai";
import {
  RunnablePassthrough,
  RunnableSequence,
} from "@langchain/core/runnables";

const system = `You are an expert at converting user questions into database queries.
You have access to a database of tutorial videos about a software library for building LLM-powered applications.
Given a question, return a list of database queries optimized to retrieve the most relevant results.

If there are acronyms or words you are not familiar with, do not try to rephrase them.`;
const prompt = ChatPromptTemplate.fromMessages([
  ["system", system],
  ["human", "{question}"],
]);
const llm = new ChatOpenAI({
  model: "gpt-3.5-turbo-0125",
  temperature: 0,
});
const structuredLLM = llm.withStructuredOutput(searchSchema, {
  name: "search",
});

const queryAnalyzer = RunnableSequence.from([
  {
    question: new RunnablePassthrough(),
  },
  prompt,
  structuredLLM,
]);

Let’s see what queries our analyzer generates for the questions we searched earlier:

console.log(await queryAnalyzer.invoke("How do I build a rag agent"));

{ query: "build a rag agent" }

console.log(await queryAnalyzer.invoke("videos on RAG published in 2023"));

{ query: "RAG", publish_year: 2023 }

Retrieval with query analysis

Our query analysis looks pretty good; now let’s try using our generated queries to actually perform retrieval.

Note: in our example, we specified tool_choice: "Search". This will force the LLM to call one - and only one - function, meaning that we will always have one optimized query to look up. Note that this is not always the case - see other guides for how to deal with situations when no - or multiple - optimized queries are returned.

import { DocumentInterface } from "@langchain/core/documents";

const retrieval = async (input: {
  query: string;
  publish_year?: number;
}): Promise<DocumentInterface[]> => {
  let _filter: Record<string, any> = {};
  if (input.publish_year) {
    // This syntax is specific to Chroma
    // the vector database we are using.
    _filter = {
      publish_year: {
        $eq: input.publish_year,
      },
    };
  }

  return vectorStore.similaritySearch(input.query, undefined, _filter);
};

import { RunnableLambda } from "@langchain/core/runnables";

const retrievalChain = queryAnalyzer.pipe(
  new RunnableLambda({
    func: async (input) =>
      retrieval(input as unknown as { query: string; publish_year?: number }),
  })
);

We can now run this chain on the problematic input from before, and see that it yields only results from that year!

const results = await retrievalChain.invoke("RAG tutorial published in 2023");

console.log(
  results.map((doc) => ({
    title: doc.metadata.title,
    year: doc.metadata.publish_date,
  }))
);

[
  {
    title: "Getting Started with Multi-Modal LLMs",
    year: "2023-12-20T08:00:00.000Z"
  },
  {
    title: "LangServe and LangChain Templates Webinar",
    year: "2023-11-02T07:00:00.000Z"
  },
  {
    title: "Getting Started with Multi-Modal LLMs",
    year: "2023-12-20T08:00:00.000Z"
  },
  {
    title: "Building a Research Assistant from Scratch",
    year: "2023-11-16T08:00:00.000Z"
  }
]