A blazing fast (currently) in memory vector database.

• Advanced querying using the logistic classifier method described here.
• Binary embeddings
• Metadata filtering (for now only support equality matching)
• Written in less than 1000 lines of pure C
• Built in Python bindings
• Much faster writes than FAISS with HNSW
• Hackable and extendable if you so desire

run these yourself with bench.py

Here the FFVec queries are the advanced version that trains a logistic classifier in the process of doing the query.

Benchmarking with 100,000 documents

FFVec Query Average Time: 1.64ms
FAISS Flat Query Average Time: 10.56ms
FAISS HNSW Query Average Time: 1.55ms
FAISS IVFPQ Query Average Time: 0.18ms


Benchmarking with 500,000 documents

FFVec Query Average Time: 2.98ms
FAISS Flat Query Average Time: 52.11ms
FAISS HNSW Query Average Time: 2.28ms
FAISS IVFPQ Query Average Time: 0.73ms


Benchmarking with 1,000,000 documents

FFVec Query Average Time: 4.96ms
FAISS Flat Query Average Time: 104.27ms
FAISS HNSW Query Average Time: 0.93ms
FAISS IVFPQ Query Average Time: 1.33ms

• [] On disk indexing
• [] More efficient non-vector querying
• [] More non-vector querying options (ne, range, etc)