Hydrate vs vstash
conceptarXiv-published local-first hybrid-retrieval substrate (one SQLite file, sqlite-vec + FTS5, adaptive RRF). Not a product — but its negative result on cross-encoder reranking directly informs Hydrate's retrieval roadmap.
What vstash is
| Stack / runtime | Single SQLite file (sqlite-vec ANN + FTS5) |
| Licence | Research (arXiv 2604.15484) |
| Storage | One SQLite file |
| Hooks | n/a |
| Injection | n/a (retrieval substrate) |
| Scope | Document-memory retrieval research |
| Maturity | Paper |
| Price | n/a |
| Version referenced | arXiv 2604.15484 (16 Apr 2026) |
| Source | arxiv.org/abs/2604.15484 |
Where vstash leads
- Adaptive RRF with per-query IDF weighting (+ up to 21.4% NDCG@10 on ArguAna vs fixed)
- Self-supervised embedding refinement via retrieval disagreement
- Production-grade substrate (integrity checks, schema versioning, diagnostics)
Where Hydrate leads
- Full product: capture, injection, team, cross-vendor, orchestration (vstash is retrieval-only)
Benchmarks as published by vstash
These are the vendor's own published figures. Hydrate has not
reproduced them and, where the benchmark differs, has no comparable
number, shown as n/a. Listed for transparency, not as a
head-to-head result.
| Benchmark | vstash | Hydrate | n | Source |
|---|---|---|---|---|
| NDCG@10 (adaptive vs fixed RRF, BEIR) | + up to 21.4% (ArguAna) | n/a | 5 datasets | link |
Verification notes
ROADMAP INPUT, not competitor. KEY: vstash reports cross-encoder reranking did NOT improve NDCG after RRF — weigh against ClawMem's positive rerank claim before flipping any rerank default. Adaptive per-query IDF weighting is a cheap dep-free idea for the c8b sweep. '16 MCP tools' unconfirmed.
Sources
- arXiv paper https://arxiv.org/abs/2604.15484