Skip to main content

Hybrid Search & Feature Benchmarks

Alpha Parameter​

The alpha parameter controls the balance between dense (semantic) and sparse (keyword) retrieval.

ModeAlphaStrategyMeanMedianRange
Dense-only1.0dense19.2ms17ms16-26ms
Sparse-only0.0sparse8.8ms9ms7-11ms
Hybrid (historical default)0.7hybrid19.6ms16ms16-27ms

These latency measurements predate the calibrated 0.5 default and retain 0.7 only as the configuration used for that historical run.

Choosing alpha
  • alpha=0.0 — Fastest. Pure BM25 keyword matching. Best when you know exact terms.
  • alpha=0.5 — Current default. Equal weighting reduces dense drift on lexical queries.
  • alpha=0.7 — Historical benchmark configuration with a stronger semantic bias.
  • alpha=1.0 — Pure semantic. Best for conceptual queries where exact keywords may not appear.

Reranking​

Cross-encoder reranking via fastembed TextCrossEncoder with Xenova/ms-marco-MiniLM-L-6-v2.

MetricValue
StatusPASS
ModelXenova/ms-marco-MiniLM-L-6-v2
Mean latency (3 runs)1,469.67ms
Median latency (3 runs)1,468ms
Range1,435-1,506ms
Top resultbank_of_america_2024.pdf ranked #1

Evidence Result rendering (top 3)​

Query: "bank of america financial results"

RankDocument VersionContent evidence item
1bank_of_america_2024.pdf (unversioned)bank_of_america_2024.pdf:7
2bank_of_america_2024.pdf (unversioned)bank_of_america_2024.pdf:8
3bank_of_america_2024.pdf (unversioned)bank_of_america_2024.pdf:9
Reranking adds measured overhead

Use --rerank selectively for precision-critical queries. In this historical run, reranking averaged 1,469.67ms while hybrid search at alpha=0.7 averaged 19.6ms.

Metadata Filters​

Document-level exact filtering narrowed Evidence Results to a specific document. Filters constrain the candidate set; they do not affect rank.

just search "financial results" -- --scope Content \
--filter document_id=bank_of_america_2024.pdf
MetricValue
StatusPASS
Evidence Results returned10
Historical matching chunks observed12
Query time30ms

The current caller contract reports totals and pagination in Document Versions. The historical chunk count is retained only as benchmark provenance.

Pagination​

TestStatusDetails
Page 1 (limit=5, offset=0)PASS5 results, has_more=true, total=10, query_time=19ms
Page 2 (limit=5, offset=5)PASS5 results, has_more=true, total=20, query_time=17ms
just search "bank" -- --scope Content --limit 5 --offset 0
just search "bank" -- --scope Content --limit 5 --offset 5