Hybrid Search

Before you run

Activate the venv from Project Setup. Install BM25 support:

pip install rank_bm25

Keep OPENAI_API_KEY in .env from OpenAI Account Setup. Chroma still calls OpenAI when it builds vectors.

Vector search and BM25

The demo uses the same three HTML tag lines as the last lesson. Query href attribute URL contains the word href, which only appears in the <a> line. Vector search alone can still return the wrong lines first.

Build the hybrid retriever

BM25Retriever.from_texts needs no API key. Wrap it and the Chroma retriever in EnsembleRetriever.

from langchain.retrievers import EnsembleRetriever
from langchain_community.retrievers import BM25Retriever
from langchain_community.vectorstores import Chroma
from langchain_openai import OpenAIEmbeddings

texts = [
    "The <a> tag creates a hyperlink. Set the href attribute to the URL.",
    "The <title> tag sets the browser tab title.",
    "The <h1> tag marks the main heading on a page.",
]

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vectorstore = Chroma.from_texts(texts=texts, embedding=embeddings)

vector_retriever = vectorstore.as_retriever(search_kwargs={"k": 2})

bm25_retriever = BM25Retriever.from_texts(texts)
bm25_retriever.k = 2

hybrid_retriever = EnsembleRetriever(
    retrievers=[bm25_retriever, vector_retriever],
    weights=[0.5, 0.5],
)

docs = hybrid_retriever.invoke("href attribute URL")
for doc in docs:
    print(doc.page_content)

weights=[0.5, 0.5] splits the vote evenly. Bump the first number if exact word matches matter more.

Use in RAG

In Retrieval-Augmented Generation (RAG), swap vectorstore.as_retriever() for hybrid_retriever. The chain code stays the same.

Run the demo

Download the script, unzip if needed, then run:

python hybrid_search_demo.py

If it fails

• ModuleNotFoundError: rank_bm25 — run pip install rank_bm25.
• AuthenticationError — check OPENAI_API_KEY in .env.
• Both printed lists match — change QUERY to a string with a rare word from one chunk, like href.

Docs: LangChain ensemble retriever.

What's Next

BM25 and Chroma are wired up. Next: the same pattern with PostgreSQL.