Stage 1
Fast
Bi-encoder candidates
| Rank | Document | Category | Cosine score |
|---|
Interactive search laboratory
Compare retrieval and reranking
Models not loaded
First-run download:
the browser downloads quantized ONNX model files from Hugging Face.
This can take a minute depending on your connection. Files are cached
by the browser for faster later searches.
Executed query
Labelled sample evaluation
Query-level ranking quality
Stage 2
Precise
Cross-encoder reranking
| Final | Move | Document | Cross score |
|---|
Architecture
Why use two ranking stages?
Independent embeddings
The bi-encoder embeds documents once and embeds each query independently. This makes candidate search efficient and reusable.
Joint pair scoring
The cross-encoder reads the query and document together, enabling richer token-level interaction and more precise relevance scoring.
Practical tradeoff
The slower model evaluates only the top candidate set, preserving interactive latency while improving the final order.
Metric guide
Verified offline evaluation
How ranking quality is measured
Live values for the currently selected query are displayed in the results section above. The cards below explain each measure, while the verified BEIR section reports aggregate performance across hundreds of held-out benchmark queries.
Recall@K
Measures whether known relevant documents entered the candidate set. A reranker cannot recover a document that was never retrieved.
MRR@10
Rewards systems that place the first relevant result as early as possible in the first ten positions.
nDCG@10
Uses graded relevance and penalizes useful documents that appear lower in the ranked list.
Latency
Separates model loading, document indexing, query embedding, retrieval, reranking, and total search time.
Verified BEIR benchmark
Aggregate ranking evidence
These fixed values were generated offline with the Python evaluation pipeline. They are separate from the live query-specific metrics shown above.
—
queries
—
documents
—
bootstrap samples
Loading verified benchmark results…
Responsible use
Human review remains essential
This project is an educational portfolio demonstration. Rankings may be incomplete, biased, irrelevant, or misleading. Cross-encoder outputs are relevance estimates—not factual guarantees or calibrated probabilities.
Do not upload confidential, proprietary, copyrighted, sensitive, or personally identifiable content. Do not use this demo as the sole basis for hiring, rejection, compensation, promotion, immigration, legal, or employment decisions.