noisy-rag-bench

A retrieval corpus and QA set for measuring what realistic document noise does to a RAG pipeline, plus the benchmark run over 18 noise conditions.

Retrieval benchmarks run on clean text. Documents inside a bank or a law firm are scans: OCR confusions, running headers, hyphens broken across lines, redacted spans. This is the corpus for measuring that, and the result it was built to expose.

Code and full write-up: https://github.com/lgoyal6/noisy-rag-bench

The finding

At 4.1% character error, recall@5 is 0.98, no lower than the 0.98 measured on clean documents, while answer recoverability has fallen from 1.00 to 0.51. A dashboard tracking retrieval quality reports that system as healthy. Retrieval still finds the right page; the figure on that page is no longer legible.

Canonicalising the text recovers 0.51 to 0.83 at the same condition.

Character error rate is a poor predictor of damage. header x3 carries 25.9% character error, 6.3 times the OCR case above, and costs nothing: recall@5 0.99, answers 1.00. What kind of noise it is matters far more than how much.

Files

file	rows	what it is
questions.jsonl	132	Gold QA pairs. qid, question, answer, gold_chunk_id, family.
corpus.jsonl	132	The chunks to index. id, doc_id, doc_title, section, text. gold_chunk_id joins to id.
noise_bench.json	72	The run: 18 noise conditions across 4 retrievers.

The corpus is synthetic and generated, seeded and reproducible from corpus.py in the repository. No client data, nothing scraped. The shape is the realistic part: near-duplicate boilerplate across many counterparties, so answering requires telling entity A's threshold from entity B's, which is exactly what OCR noise destroys.

The curve (hybrid retriever)

recall@5 is whether the right chunk was retrieved. answer is whether the answer was still extractable from what came back. +canon is after an 11-line canonicalizer.

condition	CER	recall@5	answer	+canon
clean	0.0%	0.98	1.00	1.00
ocr cer~1%	0.8%	0.96	0.89	0.94
ocr cer~2%	1.6%	0.95	0.83	0.95
ocr cer~5%	4.1%	0.98	0.51	0.83
ocr cer~10%	8.7%	0.90	0.42	0.74
whitespace cer~1%	0.7%	0.98	0.97	0.97
whitespace cer~3%	2.5%	0.94	0.89	0.91
whitespace cer~6%	5.0%	0.92	0.82	0.83
hyphen cer~1%	1.1%	0.98	0.95	1.00
hyphen cer~3%	3.0%	0.95	0.85	0.99
hyphen cer~6%	5.8%	0.95	0.73	0.98
header x1	8.3%	0.99	1.00	1.00
header x3	25.9%	0.99	1.00	1.00
redaction 2% words	5.3%	0.98	0.95	0.96
redaction 5% words	12.0%	0.92	0.86	0.87
redaction 10% words	22.7%	0.91	0.76	0.78
scan-realistic	18.5%	0.93	0.64	0.84
scan-degraded	36.9%	0.84	0.38	0.65

The losing axis

A canonicalised BM25 baseline matches or beats the hybrid retriever on recall@5 at 14 of 18 conditions. Both are in noise_bench.json, so the comparison can be redone on any column rather than taken from this sentence.

Scope and limits

• 132 questions over 132 chunks. Small, and single-hop.
• Queries are clean; only the indexed corpus is corrupted.
• Dense retrieval is sentence-transformers/all-MiniLM-L6-v2 (pure NumPy forward pass), CPU only, peak RSS 261 MB serving.
• The corpus is synthetic. It is built to have realistic structure, not to be a sample of real documents, and results on it transfer as a method rather than as numbers.

Licence

MIT. Generated corpus, no third-party document data.