Fetching from the wire…
Skills2026-06-24 · source-backed
Keep a fast retriever (BM25 or a dense index) to pull 50-200 candidates, then rerank only those with ColBERT's MaxSim, where each query token matches its best document token. ColBERTv2's quantization shrinks per-token vectors from 256 bytes to ~36 bytes (2-bit) or 20 bytes (1-bit), making token-level storage practical on commodity hardware. The production middle ground between BM25 speed and cross-encoder accuracy (source).
Each link below shares sources, entities, or timing with this story.
Shared entities / Shared topic / Earlier coverage / Tension
Both cover BM25, Keep; overlapping topics (cross-encoder, each, token); earlier BM25 coverage from 2026-03-22.
Shared entity: BM25 / Shared topic / What happened next
Both cover BM25; overlapping topics (bm25, dense, document, token); picks up the BM25 thread on 2026-07-31.
Shared entity: BM25 / Shared topic / Earlier coverage / Tension
Both cover BM25; overlapping topics (bm25, candidat, dense); earlier BM25 coverage from 2026-06-08.
Shared entity: BM25 / Shared topic / What happened next / Tension
Both cover BM25; overlapping topics (bm25, token); picks up the BM25 thread on 2026-07-28.
Shared entities / What happened next
Both cover BM25, Keep; picks up the BM25 thread on 2026-07-24.
Shared entity: BM25 / Shared topic / Earlier coverage
Both cover BM25; overlapping topics (between, bm25, token); earlier BM25 coverage from 2026-06-08.
Shared topic / Tension
Overlapping topics (accuracy, candidat, cross-encoder, each, fast); pushes against this story (but).
Shared entity: BM25 / Shared topic / What happened next
Both cover BM25; overlapping topics (bm25, token); picks up the BM25 thread on 2026-08-07.