Chapter 4.1 - Vector Embeddings (FAISS)

[!info] Building an HNSW FAISS index.

Once documents are chunked, we pass them through an Embedding Model (like HuggingFace all-MiniLM-L6-v2). This converts the text into a dense array of floats (vectors). To search these vectors instantly (even with millions of chunks), we use FAISS (Facebook AI Similarity Search). Specifically, we use the IndexHNSWFlat index type.

HNSW (Hierarchical Navigable Small World)

HNSW builds a multi-layered graph of your vectors. Instead of comparing a user's query vector against every single document (which is O(N) and extremely slow), HNSW traverses the graph hierarchically to find the nearest neighbors in O(log N) time!

import faiss

# Create the HNSW index
embedding_dim = 384
index = faiss.IndexHNSWFlat(embedding_dim, 32)
index.add(document_embeddings)