SIR · Semantic Information Retrieval

Retrieval configuration

Collection

Retrieval mode

Top k BM25 pool Fusion α

0.85

CE pool Bi-encoder

Pipeline stages (this query)

No collection loaded

Library

Import and export

Add documents

Drop file or click to upload

PDF, text, HTML, CSV, or JSON

Collection pack

Import .sirpack

Full library

Import zip

Replace local copies from zip

Collections, qrels, pipeline and UI ranking config.

Titles

loaded indexed missing working

Detail

Select a title from the catalog.

Results

Ranked passages with scores, trace fields, and full text inline. Click a hit to open pipeline analytics.

No search yet.

Definitions

BM25 (lexical): Lexical bag-of-words retrieval over the candidate pool using FTS5 inverted index.
Hybrid: Pool-restricted fusion of BM25 and Learning-to-Rank BERT semantic scores with weight α.
CE cascade: Hybrid ranking followed by cross-encoder reranking on the top candidates.
Dense: Semantic-only ranking by Learning-to-Rank BERT scores within the BM25 pool (min-max normalized).
Cross-encoder only: Cross-encoder reranking on the BM25 pool without LTR fusion.
BM25 pool: First-stage lexical retrieval size (e.g. top 100). All dense and CE stages run on this candidate set only.
α (fusion): Hybrid weight on min-max normalized scores: higher α favors BM25; lower α favors LTR-BERT semantic scores.
CE rerank pool: How many fused (or BM25) candidates the cross-encoder rescoring pass considers.
Bi-encoder: LTR-BERT compresses BERT token states to 24-d vectors; max token match scores query against each passage.
Fused score: Combined lexical and semantic score before optional CE reranking.

Configuration

Model	Learning-to-Rank BERT
Mode	Hybrid
Collection	-
Last query	-
Latency	-
Top k	10
BM25 pool	100
α	0.85
CE pool	32
Bi-encoder	LTR-BERT
Cross-encoder	MiniLM

Retrieval pipeline

Flow from corpus ingestion through query execution. Active stages are highlighted for the selected hit.

Run a search to trace the pipeline.

Search analytics

Timing breakdown, score decomposition, and rank movement across the candidate pool.

Run a search to see analytics.

Evaluation

Mark hits in results, save to qrels, then evaluate. Or upload TSV manually.

No qrels loaded.

Queries TSV (qid TAB text) Qrels TSV (qid docid rel)

Search with labeled queries or run batch evaluation.

Statistics