Releases·Tool Release
sqlite-sparse: run inference-free sparse retrieval inside SQLite with hybrid search in plain SQL
A developer released sqlite-sparse, a tool that runs a learned sparse retrieval model inside SQLite with no model invocation at query time. The encoder emits weighted vocabulary terms per document rather than dense vectors, so the index behaves like a keyword index whose terms a transformer selected. This matters to practitioners building retrieval without a serving stack for embeddings, since queries become simple token-weight lookups.
3 engagement1 sourceThu, Sep 10, 2026, 06:33 PM
AI Overview
- sqlite-sparse runs learned sparse retrieval inside SQLite; queries need no model, only a fixed per-token weight table stored in the database
- Benchmarks show the mini model trades ~7% retrieval quality for 99% lower query latency, 99% lower cold start, and 95% less query-path RAM versus dense search
- SPLADE models are BERT with the MLM head attached; since llama.cpp drops it in conversion, sqlite-sparse copies it to a sidecar file and applies it in C on ggml at insert time
- Postings and query weights live as rows in the database; queries scatter-add weights into per-document scores for top results
- Three OpenSearch inference-free models ship as aliases (mini, base, multilingual), downloading on first use, with scripts to convert custom encoders
- Pairing with sqlite-vec enables hybrid dense-plus-sparse RAG retrieval in one .db file, served via plain SQL with no GPU or model download
Related
Developments·Paper
Researchers introduce scrydb for hybrid search in SQLite
A new Python library called scrydb enables lexical, semantic, and hybrid search within SQLite. It uses FTS5 for lexical search and sqlite-vec for semantic search, with reranking and fusion capabilities. The library provides a lightweight solution for information retrieval and agentic search tasks, evaluated on benchmark datasets.
A new Python library called scrydb enables lexical, semantic, and hybrid search within SQLite. It uses FTS5 for lexical search and sqlite-vec for semantic search, with reranking and fusion capabilities. The library provides a lightweight solution for information retrieval and agentic search tasks, evaluated on benchmark datasets.
0 engagement1 sourcearxivTue, Aug 25, 2026, 04:42 AM
Products·launched
raggio: open source hybrid vector DB with 4-bit quantization
raggio is an open-source hybrid vector database that uses TurboQuant 4-bit quantization to shrink vector indexes by about 8x, enabling large document collections on small hardware. It offloads collections to disk, so storage isn't limited by RAM, and is designed as a plug-and-play alternative to commercial hosted services or hand-rolled FAISS. It targets companies and individuals needing a self-hosted, resource-efficient vector database for production workloads.
raggio is an open-source hybrid vector database that uses TurboQuant 4-bit quantization to shrink vector indexes by about 8x, enabling large document collections on small hardware. It offloads collections to disk, so storage isn't limited by RAM, and is designed as a plug-and-play alternative to commercial hosted services or hand-rolled FAISS. It targets companies and individuals needing a self-hosted, resource-efficient vector database for production workloads.
1 engagement1 sourceredditMon, Aug 24, 2026, 03:10 PM
Developments·Paper
UEmbed: decoder-only model unifies sparse and dense multimodal embeddings
Researchers introduced UEmbed, a decoder-only multimodal embedding model that produces both sparse lexical and dense representations in a single causal forward pass, extending learned sparse retrieval to multimodal settings without auxiliary cross-modal modules. This could simplify retrieval pipelines and improve efficiency for web search and RAG systems.
Researchers introduced UEmbed, a decoder-only multimodal embedding model that produces both sparse lexical and dense representations in a single causal forward pass, extending learned sparse retrieval to multimodal settings without auxiliary cross-modal modules. This could simplify retrieval pipelines and improve efficiency for web search and RAG systems.
0 engagement1 sourcearxivMon, Aug 3, 2026, 05:54 PM
Products·launched
Local RAG baseline using SQLite FTS5, no embeddings or vector DB
JAS RAG is a simple local RAG system that indexes Markdown files into SQLite, uses FTS5 for retrieval, and formats chunks for LLM context. It provides a baseline for debugging and evaluation without requiring embeddings, vector databases, or external APIs.
JAS RAG is a simple local RAG system that indexes Markdown files into SQLite, uses FTS5 for retrieval, and formats chunks for LLM context. It provides a baseline for debugging and evaluation without requiring embeddings, vector databases, or external APIs.
1 engagement1 sourceredditFri, Jul 10, 2026, 03:55 PM
Releases·Tool Release
Horosvec: a pure Go ANN vector index on SQLite for CPU-based RAG
Horosvec is an embedded approximate nearest neighbor vector index written entirely in Go, storing its index in a single SQLite file. It targets CPU-only RAG workloads, claiming 26 million entry searches in milliseconds without CGO dependencies.
Horosvec is an embedded approximate nearest neighbor vector index written entirely in Go, storing its index in a single SQLite file. It targets CPU-only RAG workloads, claiming 26 million entry searches in milliseconds without CGO dependencies.
4 engagement1 sourceredditFri, Jul 10, 2026, 07:28 AM