🔎 Neural Passage Search

See what changes when a neural ranker replaces, or re-orders, a lexical one.

MS MARCO passages100,000 passages indexed Running on ZeroGPU (on demand)TF-IDF · BM25 · Bi-encoder · Cross-encoder

Query

Systems to compare

Select systems to compare. Rank-movement chips are shown relative to BM25.

TF-IDF BM25 Bi-encoder BM25 → Cross-encoder Hybrid RRF Hybrid RRF → Cross-encoder

Results per system

3 20

Cross-encoder candidate depth (BM25 top-N)

10 200

Highlight query terms

Pick an example above or type a question, then compare how each ranking system answers it.

OFFLINE EVALUATION

No local benchmark results

Run the evaluation commands to populate artifacts/results/comparison_table.csv. No fallback or fabricated metrics are displayed.

Mark passages explicitly as relevant or irrelevant. Feedback is session-local and updates a TF-IDF ranking with Rocchio's formula.

Feedback query

Relevant passages (positive feedback)

Irrelevant passages (negative feedback)

Select a held-out, qrels-judged test query. Live search queries have no inferred labels and are not included in this benchmark analysis.

Query filter

Judged test query

Prepared test queries and qrels are required for this analysis.

System	Idea	Strength	Weakness
TF-IDF	Weight terms by frequency and rarity, cosine similarity	Fast, transparent	No length normalisation, exact words only
BM25	TF-IDF refinement with saturation and length normalisation	Strong lexical baseline	Vocabulary mismatch (synonyms, paraphrase)
Bi-encoder	Embed query and passage independently, compare vectors	Very fast at query time, semantic	Single vector per passage loses detail
Hybrid RRF	Fuse BM25 and bi-encoder ranks with Reciprocal Rank Fusion	Combines complementary lexical and semantic candidates	Adds dense retrieval cost; needs evaluation
BM25 → Cross-encoder	Re-score BM25's top-N by reading query and passage together	Most precise per pair	Slow; capped by BM25's recall
Hybrid RRF → Cross-encoder	Re-score a fused BM25+dense candidate pool	Combines broad recall and interaction scoring	Higher latency and candidate costs

Coloured bar: the score, scaled within that column. Scores from different systems are on different scales, so compare ranks, not raw values.
▲ / ▼ chips: how far a passage moved relative to BM25's ranking.
✨ outside BM25 top-N: the system found a passage BM25's candidate list never contained. For the cross-encoder this can't happen (it only re-orders BM25 candidates), which is exactly why a dense first stage helps.
Highlighted words: query terms, so you can see when a system retrieves without lexical overlap.