Adding Retrieval-Augmented Generation (RAG)

While an agent that can perform math is nifty (LLMs are usually not very good at math), LLM-based applications are always more interesting when they work with large amounts of data. In this case, we're going to use a 200-page PDF of the proposed budget of the city of San Francisco for fiscal years 2024-2024 and 2024-2025. It's a great example because it's extremely wordy and full of tables of figures, which present a challenge for humans and LLMs alike.

To learn more about RAG, we recommend this introduction from our Python docs. We'll assume you know the basics:

• Parse your source data into chunks of text.
• Encode that text as numbers, called embeddings.
• Search your embeddings for the most relevant chunks of text.
• Use the relevant chunks along with a query to ask an LLM to generate an answer.

We're going to start with the same agent we built in step 1, but make a few changes. You can find the finished version in the repository.

Installation

npm install llamaindex @llamaindex/openai @llamaindex/huggingface 

New dependencies

We'll be bringing in SimpleDirectoryReader, HuggingFaceEmbedding, VectorStoreIndex, and QueryEngineTool, OpenAIContextAwareAgent from LlamaIndex.TS, as well as the dependencies we previously used.

import { QueryEngineTool, Settings, VectorStoreIndex } from "llamaindex";
import { OpenAI, OpenAIAgent } from "@llamaindex/openai";
import { HuggingFaceEmbedding } from "@llamaindex/huggingface";
import { SimpleDirectoryReader } from "@llamaindex/readers/directory";

Add an embedding model

To encode our text into embeddings, we'll need an embedding model. We could use OpenAI for this but to save on API calls we're going to use a local embedding model from HuggingFace.

Settings.embedModel = new HuggingFaceEmbedding({
  modelType: "BAAI/bge-small-en-v1.5",
  quantized: false,
});

Load data using SimpleDirectoryReader

SimpleDirectoryReader is a flexible tool that can read various file formats. We will point it at our data directory, which contains a single PDF file, and retrieve a set of documents.

const reader = new SimpleDirectoryReader();
const documents = await reader.loadData("../data");

Index our data

We will convert our text into embeddings using the VectorStoreIndex class through the fromDocuments method, which utilizes the embedding model defined earlier in Settings.

const index = await VectorStoreIndex.fromDocuments(documents);

Configure a retriever

Before LlamaIndex can send a query to the LLM, it needs to find the most relevant chunks to send. That's the purpose of a Retriever. We're going to get VectorStoreIndex to act as a retriever for us

const retriever = await index.asRetriever();

Configure how many documents to retrieve

By default LlamaIndex will retrieve just the 2 most relevant chunks of text. This document is complex though, so we'll ask for more context.

retriever.similarityTopK = 10;

Approach 1: Create a Context-Aware Agent

With the retriever ready, you can create a context-aware agent.

const agent = new OpenAIContextAwareAgent({
  contextRetriever: retriever,
});
 
// Example query to the context-aware agent
let response = await agent.chat({
  message: `What's the budget of San Francisco in 2023-2024?`,
});
 
console.log(response);

Expected Output:

The total budget for the City and County of San Francisco for the fiscal year 2023-2024 is $14.6 billion. This represents a $611.8 million, or 4.4 percent, increase over the previous fiscal year's budget. The budget covers various expenditures across different departments and services, including significant allocations to public works, transportation, commerce, public protection, and health services.

Approach 2: Using QueryEngineTool (Alternative Approach)

If you prefer more flexibility and don't mind additional complexity, you can create a QueryEngineTool. This approach allows you to define the query logic, providing a more tailored way to interact with the data, but note that it introduces a delay due to the extra tool call.

const tools = [
  index.queryTool({
    metadata: {
      name: "san_francisco_budget_tool",
      description: `This tool can answer detailed questions about the individual components of the budget of San Francisco in 2023-2024.`,
    },
  }),
];
 
// Create an agent using the tools array
const myAgent = agent({ tools });
 
let toolResponse = await myAgent.run("What's the budget of San Francisco in 2023-2024?");
 
console.log(toolResponse);

Expected Output:

{
  toolCall: {
    id: 'call_iNo6rTK4pOpOBbO8FanfWLI9',
    name: 'san_francisco_budget_tool',
    input: { query: 'total budget' }
  },
  toolResult: {
    tool: QueryEngineTool {
      queryEngine: [RetrieverQueryEngine],
      metadata: [Object]
    },
    input: { query: 'total budget' },
    output: 'The total budget for the City and County of San Francisco for Fiscal Year (FY) 2023-24 is $14.6 billion, which represents a $611.8 million, or 4.4 percent, increase over the FY 2022-23 budget. For FY 2024-25, the total budget is also projected to be $14.6 billion, reflecting a $40.5 million, or 0.3 percent, decrease from the FY 2023-24 proposed budget. This budget includes various expenditures across different departments and services, with significant allocations to public works, transportation, commerce, public protection, and health services.',
    isError: false
  }
}

Once again we see a toolResult. You can see the query the LLM decided to send to the query engine ("total budget"), and the output the engine returned. In response.message you see that the LLM has returned the output from the tool almost verbatim, although it trimmed out the bit about 2024-2025 since we didn't ask about that year.

Comparison of Approaches

The OpenAIContextAwareAgent approach simplifies the setup by allowing you to directly link the retriever to the agent, making it straightforward to access relevant context for your queries. This is ideal for situations where you want easy integration with existing data sources, like a context chat engine.

On the other hand, using the QueryEngineTool offers more flexibility and power. This method allows for customization in how queries are constructed and executed, enabling you to query data from various storages and process them in different ways. However, this added flexibility comes with increased complexity and response time due to the separate tool call and queryEngine generating tool output by LLM that is then passed to the agent.

So now we have an agent that can index complicated documents and answer questions about them. Let's combine our math agent and our RAG agent!