xDev Asia

Build RAG pipeline for agents: document loading, chunking strategies, embedding models, vector store (ChromaDB, Qdrant). Semantic search vs keyword search. Hybrid retrieval.

Build AI Agents: From Zero to Production

Part 3: RAG & Memory — Give Agent memory

xdev.asia

Introduction

Agents are stronger when they have their own knowledge base. RAG (Retrieval-Augmented Generation) allows agents to search for information from documents, databases, or knowledge bases before responding — reducing hallucination and increasing accuracy.

1. RAG Pipeline Overview

Documents → Chunking → Embedding → Vector Store → Retrieval → LLM → Answer

2. Implementation with ChromaDB

import chromadb
from openai import OpenAI

client = OpenAI()
chroma = chromadb.PersistentClient(path="./agent_knowledge")
collection = chroma.get_or_create_collection("docs")

def add_documents(texts, metadatas=None):
    embeddings = get_embeddings(texts)
    collection.add(
        documents=texts,
        embeddings=embeddings,
        ids=[f"doc_{i}" for i in range(len(texts))],
        metadatas=metadatas,
    )

def search_knowledge(query, n_results=5):
    query_embedding = get_embeddings([query])[0]
    results = collection.query(
        query_embeddings=[query_embedding],
        n_results=n_results,
    )
    return results["documents"][0]

3. RAG as Agent Tool

@registry.register("search_knowledge", "Tìm kiếm trong knowledge base", {...})
def search_knowledge_tool(query: str) -> str:
    results = search_knowledge(query, n_results=3)
    return "\n---\n".join(results)

Summary

RAG = allows agents to access private knowledge base
Chunking strategy affects retrieval quality
ChromaDB and Qdrant are the two most popular DB vectors
Hybrid search (semantic + keyword) gives the best results

Exercises