## Using RAG (Retrieval-Augmented Generation)

Work with large datasets with RAG

# Using RAG with Superexpert.AI

Superexpert.AI supports powerful Retrieval-Augmented Generation (RAG) functionality to enhance your AI agents’ ability to retrieve and utilize information from large files and databases effectively.

## Why Use RAG?

Retrieval-Augmented Generation (RAG) allows your AI agents to access information from extensive documents and databases that exceed the typical context window limits of AI models. Unlike file attachments, which load the entire file into the agent’s context, RAG intelligently retrieves relevant text chunks based on the similarity to the user’s queries.

When you use RAG:

- Files are chunked into smaller, manageable pieces.
- Each chunk is converted into vector embeddings and stored in a database.
- User queries are similarly converted into embeddings and matched against these stored chunks.

**Ideal Use Cases for RAG:**

- Quickly searching product documentation for customer support.
- Providing accurate, context-specific answers from large sets of academic articles or research papers.
- Accessing detailed historical or factual data that would otherwise exceed standard context limitations.

## Walkthrough: Setting Up and Using RAG

In this walkthrough, we’ll upload and use a file named **[titanic.json](/content/docs/retrieval-augmented-generation/titanic.json)**, which contains detailed passenger data from the Titanic.

### Step 1: Create a New Corpus

1. From the main navigation menu, click **Corpora**.
2. If no corpus exists, you will automatically navigate to the **New Corpus** page. Otherwise, click **New Corpus**.
3. Enter a descriptive name and optional description for your corpus (e.g., “Titanic Passenger Data”).

### Step 2: Upload Files to Your Corpus

You have two methods:

#### Method 1: Using the Web Interface

- Click **Choose File** to select your **[titanic.json](/content/docs/retrieval-augmented-generation/titanic.json)** file.
- Specify the **Chunk Size** (recommended: 100 tokens for detailed accuracy).
- Set the **Chunk Overlap** percentage (recommended: 15%).
- Click **Upload**.

_Figure 1: Uploading files with the web interface._

#### Method 2: Using the CLI Tool

- Open your terminal and run:

```
npx @superexpert-ai/rag init
```

- Executing the tool with `init` generates a rag.config.js file. Enter your configuration settings in rag.config.js and execute the following command to chunk and upload your files:

```
npx @superexpert-ai/rag run
```

The CLI tool reliably handles large files, multiple uploads, and both text and PDF files.

_Figure 2: Uploading files with the RAG tool._

> **RAG Tool:** For details on using the RAG tool, visit [@superexpert-ai/rag](https://www.npmjs.com/package/@superexpert-ai/rag).

### Step 3: Test Your Corpus

- Navigate to the **Corpora** page.
- Click the **Test** button next to your corpus.
- Enter a **Limit** for the number of returned matches and a **Similarity Threshold** (higher thresholds yield more precise results).
- Enter test queries such as:
  - “How old was Flynn?”
  - “What cabin was Mrs Nye in?”
  - “How old was Bertha?”
- Click **Search** to view retrieved matches.

_Figure 3: Testing your corpus._

### Step 4: Use RAG in a Chat Interface

- From the main navigation menu, select **Agents**.
- Click **Tasks** next to your chosen Agent.
- Expand the **Context Data** panel.
- Select your corpus and configure **Limit** and **Similarity Threshold** for real-time retrieval during chat.
- Begin a conversation by navigating to **Agents** > clicking **Chat** next to the agent.

> **Similarity Thresholds:** We recommend that you experiment with different Similarity Thresholds. Lower values will yield more matches.

As you chat, your agent dynamically fetches relevant information from your corpus based on your messages, significantly enhancing interaction quality and accuracy.

_Figure 4: Using RAG with your agent._

## Technical Background

Superexpert.AI uses PostgreSQL and `pgvector` to efficiently store and query vector embeddings, ensuring fast and reliable RAG operations.