karl
caf6a2f82e
Retrieval regression corpus
This synthetic seed corpus contains 20 documents and 24 predetermined queries: exact filenames, error/invoice identifiers, remembered concepts and English/German cross-language retrieval. Each query labels one expected file. Supplier-contract termination and meeting cancellation deliberately share words but differ in intent.
Report top-1 accuracy, recall@10, per-query rank and latency. Keep failures in reports; do not rewrite expected answers or remove difficult queries to make a model pass. Evaluate lexical, semantic and hybrid modes on the same corpus.
This corpus is a regression fixture, not a representative user dataset. Production quality needs a larger independently labelled corpus, varied document lengths, PDF/OCR/media cases, ambiguous/no-answer queries and measured filter correctness.