DealBench

dev split. Real numbers from the tuning set, not the held-out test split; retrieval only, no answer layer has run yet.

Failure attribution

Every question is attributed to the first pipeline stage that lost its gold evidence: parse → chunk → recall → rerank → generation → grounding. Click a segment to list its questions.

  • Parse
  • Chunk
  • Recall
  • Rerank
  • Generation
  • Grounding
  • No miss
B1 BM25 over raw tokens, fixed 512-word windows.
B2 BM25 over MorphoDiTa lemmas with diacritics folded, fixed 512-word windows.
D1 Naive RAG - dense bge-m3 over fixed 512-word windows.
H1L Lemmatised BM25 + dense bge-m3 over structure chunks, fused by RRF (local twin of H1).
T1L H1L over table-row chunks and XLSX cell rows (local twin of T1, no reranker).

Select a segment above — for example the naive config's chunk misses.