From documents
to defensible evidence.
Search a collection, inspect exactly where each result comes from, and compare transparent ranking methods. No invented answers. Just attributable evidence.
Evidence search
Start with a question. Review source-level evidence before drawing conclusions.
Ranked evidence 0 results
Document passages with traceable IDs and lexical relevance.
Evidence is not an answer. Rankings measure term relevance, not factual truth or entailment. Check the complete original document and any conflicting sources before making a claim.
Evaluation Lab.
Measure what the retrieval engine actually finds. Compare algorithms against explicit relevance labels, with per-query diagnostics and inspectable methodology.
Algorithm comparison
Retrieval quality by method
Higher is better. Macro averages from the same question set.
Query-level diagnostics
Identify failed retrievals and inspect expected versus returned documents.
| QUESTION | LANG | EXPECTED | FIRST HIT | RECALL@K | STATUS |
|---|
Benchmark transparency: this demonstration uses synthetic documents with hand-authored relevance labels. These numbers are not claims about performance on real datasets. The system does not perform cross-language semantic retrieval.
Corpus Workspace.
Bring your own documents and labelled questions into a private, client-side research environment. Nothing is sent to a backend by this application.
Import documents
Accepts JSON, CSV, TXT and Markdown. CSV should contain id,title,body headers.
Import relevance labels
Upload JSON questions, each with query and relevant_ids for a defensible benchmark.
Collection quality
Languages
Categories
Integrity checks
Quality controls for incomplete or inconsistent evaluation inputs.
Indexed documents
Document IDs, provenance fields and corpus coverage.
Audit & Export.
Turn an exploratory search into a repeatable evidence record. Record methods, parameters, ranked passages, known limitations and quality metrics.
Your research. An inspectable trail.
Exports are generated locally, immediately, as JSON files you can inspect, archive or reproduce.
Retrieval report
Query, method, ranked document identifiers, score, matching terms and verbatim evidence excerpts.
Full evaluation report
All ranking methods, corpus integrity profile, labeled questions, metrics, per-question retrieval and caveats.
Research methodology
Index and normalizeUnicode tokens with accent folding and stopword filtering. Titles receive 2× weight.
Retrieve and rankBM25 k1=1.5 / b=0.75, cosine TF-IDF or reciprocal rank fusion k=60.
EvaluatePrecision@k, Recall@k, reciprocal first-relevant rank and binary nDCG@k.
Examine evidencePassages are exact excerpts of the corpus; relevance does not imply factual support.