61 lines (47 loc) · 2 KB

Performance

SQLRite is optimized first for embedded retrieval.

Current benchmark snapshot

Deterministic filtered cosine workload:

5k records
120 measured queries
64 dimensions
8 tenants
top_k=10

SQLRite results

Mode	QPS	p95 latency	Recall@10
`brute_force` embedded	`3380.07`	`0.3543 ms`	`1.0`
`hnsw_baseline` embedded	`3530.96`	`0.3327 ms`	`1.0`
`brute_force` HTTP compact	`1807.27`	`0.7538 ms`	`1.0`
`hnsw_baseline` HTTP compact	`1828.17`	`0.7070 ms`	`1.0`

Comparator snapshot on the same workload

Engine	Mode	QPS	Recall@10
SQLRite	`brute_force` embedded	`3380.07`	`1.0`
sqlite-vec	exact	`3163.27`	`1.0`
Qdrant	exact	`2576.75`	`1.0`
pgvector	exact	`1739.64`	`1.0`
SQLRite	`hnsw_baseline` embedded	`3530.96`	`1.0`
Qdrant	HNSW	`2661.91`	`1.0`
pgvector	HNSW	`1924.01`	`0.5740`
LanceDB	IVF_FLAT	`1331.18`	`0.9510`

What these numbers mean

embedded mode is the strongest SQLRite deployment path
compact HTTP reduces transport overhead materially compared with the older full-response path
the product-facing recommendation remains: embed SQLRite when you can, serve it only when you need a process boundary

Tuning knobs

Variable	Purpose
`SQLRITE_VECTOR_STORAGE`	choose `f32`, `f16`, or `int8`
`SQLRITE_SQLITE_MMAP_SIZE`	raise SQLite mmap size
`SQLRITE_SQLITE_CACHE_SIZE_KIB`	raise SQLite cache size
`SQLRITE_ENABLE_ANN_PERSISTENCE`	persist ANN sidecar state
`SQLRITE_ANN_MIN_CANDIDATES`	control ANN floor
`SQLRITE_ANN_MAX_CANDIDATE_MULTIPLIER`	cap ANN expansion

Practical guidance

Goal	Recommendation
fastest local retrieval	embed SQLRite directly
low memory footprint	try `SQLRITE_VECTOR_STORAGE=int8`
faster service path	use `/v1/query-compact`
tighter filtered search	keep metadata filters narrow and explicit