FAISS

FAISS is an in-process similarity search library — the index lives in memory inside your Python process, with no server and no persistence unless you save it explicitly.

from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

vector_store = FAISS.from_documents(documents, embeddings)
vector_store.save_local("faiss_index")

loaded = FAISS.load_local(
    "faiss_index",
    embeddings,
    allow_dangerous_deserialization=True,
)

Save and load

save_local/load_local serialize the index to a folder on disk; there's no running service to point at, which makes FAISS well suited to batch jobs and notebooks but awkward for a multi-process web app sharing one index. The allow_dangerous_deserialization flag isn't decoration — see below.

Pitfalls

load_local unpickles a file. A FAISS index saved by an untrusted party is a code-execution vector, not just data — only load indexes you built or that came from a trusted source. FAISS also has no built-in metadata filtering comparable to Chroma or pgvector; filtering by metadata means post-filtering results in your own code, which costs you recall unless you over-fetch first.

See also

• Overview — the shared VectorStore interface.
• Comparison — FAISS against the alternatives.