Binary-quant search in production (our funnel, shipped) + SIMD kernels

Engines that ship the binary-scan + rerank design we converged on — useful to validate our numbers and steal serving/recall-ratio choices from.

• Weaviate (Go): weaviate.io. "flat + BQ" index = brute-force binary-quantized scan + full-precision rerank — literally our binary funnel in production. Also has product quantization and a BQ-over-HNSW mode. Closest external analog; learn their rerank ratio and when they pick flat-BQ vs graph.
• Qdrant (Rust): github.com/qdrant/qdrant. Binary + scalar + product quantization, with oversampling + rerank (the funnel pattern), and the best filtered-search design (filterable HNSW + payload index). Same language as us — most readable reference codebase.
• Milvus / Knowhere (C++): index library wrapping FAISS/HNSW; segment lifecycle (growing→sealed) for updates — the dynamic-index angle.
• usearch / SimSIMD (unum): github.com/unum-cloud/usearch. Single-header SIMD distance kernels across ISAs incl. hamming/popcount — the vectorized-distance craft reference (relevant to [[history 050]]).
• Lance / LanceDB: columnar on-disk vector format + IVF_PQ; the disk-native storage angle (relevant to [[history 045]]).

Why tracked

Weaviate-BQ and Qdrant-BQ are the production proof that our 1-bit-funnel design is the right one; they're where to check our recall/QPS against shipped systems and to learn filtered-search + dynamic-segment patterns if we ever expand scope. usearch is the kernel-craft reference.

Related

• Our funnel: [[history 009]]. Serving: carousel [[history 041]]. SIMD: [[history 050]].