Vector Stores: FAISS, Chroma

1. What is a Vector Store?

It is a specialized database. It stores:

1. Vector (Index)
2. Metadata (Filter)
3. Content (Payload)

We use it because SQL databases are too slow for vector math.

2. Indexing vs Searching

If you have 1 million vectors, comparing your query to every single one (Brute Force) is slow. Vector stores use ANN (Approximate Nearest Neighbor) indexing. They group similar vectors into clusters. When you search, they only check the relevant cluster. Trade-off: You gain Speed (100x), you lose Recall (1-5% errors).

3. FAISS vs Chroma

FAISS (Facebook AI Similarity Search)

• Pros: Incredibly fast. Runs on GPU. The rigorous standard.
• Cons: Hard to use. No metadata filtering natively (without wrappers). In-memory only (usually).

ChromaDB

• Pros: Developer friendly. "Just works." Handles metadata/storage/embedding for you.
• Cons: Newer, less optimized for massive scale (yet).

4. Usage in LangChain

They share the same interface.

from langchain_community.vectorstores import Chroma

# Add
db = Chroma.from_documents(docs, embedding_model)

# Search
results = db.similarity_search("How do I reset password?")

5. Summary

For < 100k docs, use Chroma/FAISS locally. For > 1M docs, use Pinecone/Weaviate/Milvus. The logic is the same: Index vectors to find neighbors fast.

Key Intuition: "It's a map of meaning, not a list of keywords."