Tutorial: Build a RAG Agent

Learn how to give your agent long-term memory and access to private documentation using Vector Stores.

What we'll build#

We will build a "Knowledge Base Agent" that can answer questions about a fictional product manual. We'll use:

OpenAI Embeddings for vectorizing text.
In-Memory Vector Store for simplicity (swappable for Pinecone/PgVector).
RAG Tool to retrieve context.

Setup Knowledge Base

First, we need a way to store and retrieve documents. We'll define a simple VectorStore interface.

typescript

import { Tool } from '@akios/core'
import { z } from 'zod'

// Mock vector store (in production, use Pinecone/Weaviate)
const knowledgeBase = [
  { id: 1, text: "The ACME 2000 has a battery life of 12 hours.", vector: [...] },
  { id: 2, text: "To reset the device, hold the power button for 5s.", vector: [...] },
]

async function search(query: string) {
  // Simulate semantic search
  console.log(`Searching for: ${query}`)
  return knowledgeBase.map(k => k.text).join('\n')
}

Create Retrieval Tool

We wrap our search function in a Tool so the agent knows how and when to use it.

typescript

const retrievalTool = new Tool({
  name: 'search_manual',
  description: 'Search the product manual for technical specifications and troubleshooting.',
  schema: z.object({
    query: z.string().describe('The semantic search query')
  }),
  handler: async ({ query }) => {
    const results = await search(query)
    return results || "No relevant information found."
  }
})

Initialize Agent

Now we give the tool to the agent. Notice the system prompt encourages using the tool.

typescript

const agent = new Agent({
  name: 'SupportBot',
  model: 'gpt-4-turbo',
  tools: [retrievalTool],
  systemPrompt: `
    You are a technical support assistant for ACME Corp.
    ALWAYS search the manual before answering technical questions.
    If the answer isn't in the manual, say "I don't know".
  `
})

Test It

Run the agent with a question that requires external knowledge.

typescript

const response = await agent.run("How do I perform a hard reset?")
console.log(response.text)

> Searching for: perform a hard reset
> Tool Output: "To reset the device, hold the power button for 5s."
> Agent: You can perform a hard reset by holding the power button for 5 seconds.

Pro Tip

For large datasets, use the @akios/rag package (coming soon) which handles chunking, embedding, and re-ranking automatically.