MonaVec: A Training-Free Embedded Vector Search Kernel for Edge and Offline AI Systems
2026-06-19T08:52:19Z•2dbe84d88be6905c603413ad530bac49a5edf06885727347cb95cb641f020540
.mvec4-bit-quantizationCLIPChaCha20Lloyd-MaxRandomized-Hadamard-TransformRustSAFE-CascadeSIMDToken-Factorychart-QAcold-startcost-adaptive-routing','semantic-caching','calibration','P-CHR-Adeterminisme-commerceembedded-retrievalmultimodal-retrievaloffline-agentson-device-RAGquantizationrecommendation-modelssoft-tokenstraining-freevector-searchvision-language-models
What happened
This arXiv CS/IR batch highlights advances in efficient, on-device and scalable retrieval, multimodal systems, and LLM-agent context management. Key paper: MonaVec — a training-free, data‑oblivious embedded vector-search kernel (single .mvec file) that uses a Randomized Hadamard Transform + precomputed Lloyd‑Max tables to quantize to 4 bits, a ChaCha20 rotation seed for build‑determinism, and a Rust implementation with SIMD; reported 0.960 Recall@10 (27 MB) on semantic embeddings. Other notable works: VCG — a CLIP‑based multimodal retrieval system for e‑commerce video cold‑start (50% uplift in
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- arxiv_cs_ir
- Record identifier
- 2dbe84d88be6905c603413ad530bac49a5edf06885727347cb95cb641f020540
- Enrichment time
- 2026-06-19T08:52:19Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.