TF-IDF
Weights terms by frequency and rarity, then scores matching words.
Good at fast, transparent exact-term matching
Watch for vocabulary mismatch and no length normalisation
🔎 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 CPU (ZeroGPU bypassed)TF-IDF · BM25 · Bi-encoder · Cross-encoder
INTERACTIVE RETRIEVAL LAB
Ask a question. Compare how each system finds answers.
Choose an example or enter your own query. Live results are ranked, not judged; use the Benchmarks tab for the saved offline quality evaluation.
3 20
10 200
Ask anything
Pick an example above or type a question, then compare how each ranking system answers it.
OFFLINE EVALUATION
Local MS MARCO passage test evaluation — saved results
Heads-up: these metrics were produced by local/Colab evaluation on the MS MARCO passage test split (750 queries; latency sampled on 200 queries). They are not measurements of this Space's currently bundled corpus or model checkpoints. The table is shown as a reference, not as a claim about live Space quality or latency.
Best MRR@10
0.489
Bi-encoder
Best Recall@100
0.995
Hybrid RRF
Lowest median latency
8.2 ms
Bi-encoder
MRR@10 lift vs. BM25
+0.158
Bi-encoder
Evaluation protocol and reproducibility details
- Dataset/split: msmarco-passage / test
- Queries: 750 test queries; latency sample: 200 queries
- Candidate/rerank depth: 100 / 50
- RRF constant: k=60
- Query fingerprint: 39a9494c56f2e867ec3f54cb4b8b28b77bed82de79ca56896089f3c3feaebbb1
- Qrels fingerprint: bd272e0089d080b97beb265f1cdb586178b17b84ddaae7afe23d57084f3242de
Metrics are macro-averaged over the recorded test queries. Paired confidence intervals resample the same query IDs for both systems. Latency is a separate query sample, not a per-query quality measurement. These files were copied from the completed local evaluation; the live Space uses different bundled corpus/model artifacts.
| System | MRR@10 | NDCG@10 | P@5 | Recall@10 | Recall@100 | MAP@10 | Median latency (ms/query) |
|---|---|---|---|---|---|---|---|
| TF-IDF | 0.2943 | 0.3854 | 0.1091 | 0.691 | 0.9073 | 0.2874 | 45.96 |
| BM25 | 0.3316 | 0.423 | 0.1192 | 0.7269 | 0.9087 | 0.3247 | 151.9 |
| Bi-encoder | 0.4892 | 0.5788 | 0.1528 | 0.872 | 0.982 | 0.4831 | 8.244 |
| BM25 → Cross-encoder | 0.397 | 0.4772 | 0.1251 | 0.7444 | 0.9087 | 0.3902 | 272.2 |
| Hybrid RRF | 0.4822 | 0.5687 | 0.1533 | 0.8531 | 0.9947 | 0.4746 | 149.7 |
| Hybrid RRF → Cross-encoder | 0.3402 | 0.4237 | 0.1109 | 0.7086 | 0.9947 | 0.3319 | 209.1 |
| BM25 → Cross-encoder (fixed candidates) | 0.4312 | 0.5072 | 0.1368 | 0.762 | 0.9087 | 0.4231 | 199.9 |
Best value per column highlighted. Results and run provenance are in artifacts/local-evaluation. Recall@100 is from saved full-depth local ranked runs, not top-10 outputs.
Ranking quality at a glance
Higher is better. MRR@10 and NDCG@10 focus on top-ranked relevance; Recall@100 measures candidate coverage.
Median latency per query
Sorted fastest to slowest · measured on 200 local test queries. Lower is faster; speed ratios use BM25 as the reference.
Quality vs. speed
Saved local MS MARCO passage test evaluation. Favor the upper-left: higher MRR with lower query latency.
Paired MRR@10 difference vs. BM25
Point is the mean difference; whisker is the paired bootstrap 95% confidence interval. Intervals wholly right of zero favor the method; crossing zero means the estimated difference is uncertain.
Per-query MRR@10 distribution
Center line is median; box is the interquartile range; whiskers span the minimum and maximum query scores.
How often each system beats BM25
Paired test queries compared by MRR@10. Counts at right are improved / tied / lower; each query is compared to itself.
PERSONALIZE A SEARCH
Teach the ranker what matters for this query
Mark a few retrieved passages as relevant or irrelevant. Rocchio feedback adjusts the TF-IDF query for this update only; it does not train the model or change benchmark results.
1 Retrieve candidates2 Label examples3 Review rank changes
Start with a queryRetrieve up to 20 TF-IDF candidates to begin labeling examples.
Your updated ranking will appear hereCandidate order changes after you apply positive or negative feedback.
QUERY-LEVEL DIAGNOSTICS
See where retrieval succeeds or misses
Filter prepared, judged test queries by length, then compare how the Space's current retrieval systems handle a selected query. Live searches are not assigned inferred relevance labels.
Filter test queries
Select a judged query
READY TO EXPLORE
750 held-out queries with qrels
Choose a query to compare candidate coverage and top-ranked passages.
How to read the retrieval comparison
Each method sees the same query and passage collection, but finds and orders candidates differently. Compare the returned passages and rank changes; relevance is not inferred for a live search.
01Find candidatesLexical, semantic, or both
→
02Fuse or re-rankOptional deeper scoring
→
03Compare top resultsInspect ranks and evidence
Retrieval methods
BM25
A tuned lexical baseline with term saturation and document-length normalisation.
Good at strong exact-term retrieval
Watch for synonyms and paraphrases that use different words
Bi-encoder
Encodes the query and passages separately, then compares their vectors.
Good at matching meaning across different wording
Watch for detail lost in a single vector representation
Hybrid RRF
Combines BM25 and bi-encoder rankings with Reciprocal Rank Fusion.
Good at combining lexical and semantic candidates
Watch for extra retrieval cost; fusion still needs evaluation
BM25 → Cross-encoder
Reads each query–passage pair together and re-orders BM25's candidate list.
Good at precise pairwise relevance scoring
Watch for latency and BM25's candidate recall ceiling
Hybrid RRF → Cross-encoder
Re-scores a candidate pool assembled from both lexical and semantic retrieval.
Good at broader candidate coverage plus pairwise scoring
Watch for additional compute and candidate-pool sensitivity
What the cards tell you
- Rank number is the method's position for that passage.
- ▲ / ▼ chips show movement relative to BM25; ✨ means the passage was outside BM25's candidate depth.
- Score bars are scaled within each method. Raw scores across methods are not directly comparable.
- Highlighted terms reveal exact query-word overlap; no highlight does not mean a passage is irrelevant.
- Latency is an observed time for this query on the current Space hardware, not a benchmark guarantee.
How quality is measured
MRR@10
Rewards placing the first relevant passage near the top ten.
NDCG@10
Rewards highly ranked relevant passages, with a discount for lower positions.
Recall@100
Share of judged relevant passages retrieved in the top 100 candidates.
These require judged queries and matching qrels. The saved local benchmark is explicitly labeled on its tab.