Building a RAG Pipeline

Complete RAG Pipeline

RAG combines semantic retrieval with LLM generation to answer questions grounded in your own documents.

# Complete RAG implementation
import anthropic
from sentence_transformers import SentenceTransformer
import chromadb

embedder   = SentenceTransformer('all-MiniLM-L6-v2')
chroma     = chromadb.Client()
collection = chroma.create_collection('knowledge')
client     = anthropic.Anthropic()

# --- INDEXING (run once) ---
def index_docs(docs: list[dict]):
    texts  = [d['text'] for d in docs]
    embeds = embedder.encode(texts).tolist()
    ids    = [d['id'] for d in docs]
    collection.add(documents=texts, embeddings=embeds, ids=ids)

# --- QUERY (on each user question) ---
def ask(question: str) -> str:
    # 1. Embed question
    q_embed = embedder.encode([question]).tolist()[0]
    # 2. Retrieve top-3 relevant chunks
    results = collection.query(query_embeddings=[q_embed], n_results=3)
    context = '
---
'.join(results['documents'][0])
    # 3. Generate grounded answer
    prompt = 'Context:
' + context + '

Question: ' + question
    response = client.messages.create(
        model='claude-opus-4-5', max_tokens=500,
        system='Answer ONLY from context. Say you don not know if absent.',
        messages=[{'role':'user','content':prompt}]
    )
    return response.content[0].text