sqlite-sparse runs learned sparse retrieval inside SQLite with no model at query time
Much-Ingenuity5415 · reddit · 2026-09-12
Developer arbazsiddiqui released sqlite-sparse, an open-source project embedding learned sparse retrieval (SPLADE-style models) into a SQLite file with zero model calls at query time.
- How it works: The sparse encoder returns weighted vocabulary terms instead of dense vectors. Following OpenSearch's inference-free approach, the model runs only on documents at insert time; queries just look up per-token weights. SPLADE models are BERT with the MLM head attached — llama.cpp drops that head on conversion, so the project copies it to a sidecar file and applies it in C on ggml at insert time.
- Benchmarks: vs dense search in the same SQLite file, the mini model gives up 7% retrieval quality while cutting query latency by 99%, cold start by 99%, and query-path RAM by 95%.
- Availability: Aliases for mini/base/multilingual OpenSearch models auto-download on first use; BYO sparse encoders supported via conversion scripts. pip install sqlite-sparse.
Like sqlite-vec did for embeddings, it enables hybrid RAG retrieval in a single .db file — semantic search through plain SQL, no GPU or model downloads.
More from Research
- Nature MI paper unifies neural superposition and sparse interpretable codes in one framework — GretaTuckute · 2026-09-12
- Astribot's SmoothRL: online RL learns only from actions the robot actually executes — jiqizhixin · 2026-09-12
- Building adebench: scoring what your agent's memory actually delivers, 93.7/100 — Soft-Lie-434 · 2026-09-12
- ECCV debates: are explicit 3D representations obsolete as video generative models take over? — qixing_huang · 2026-09-12
- OpenAI claims a Millennium Prize problem, and mathematicians are uneasy — The Verge AI · 2026-09-12
- Fruit fly brain sim wired to a 1B LLM: 139,255 neurons now chat with you — TinfoilTricorn · 2026-09-12