AI Document Search Benchmarks Explained: What the Numbers Mean

When an AI document tool claims "99% accuracy," the number is almost meaningless unless you know three things: which documents it was tested on, who wrote the test, and what the tool does when it doesn't know the answer. A benchmark is the answer to all three — a fixed set of questions and documents, ideally published by an outside party, that anyone can re-run to check the score. This article explains the two public benchmarks that matter most for professional-services document search (CUAD for legal contracts and FinanceBench for financial filings), what a good result actually looks like, and why a lower score on a hard benchmark is often more trustworthy than a perfect score on an easy one.

If you run a small law or accounting firm, this is the difference between buying a tool that quietly fabricates a figure in an audit workpaper and one that tells you when the answer isn't in your documents. The numbers are worth understanding.

What is an AI benchmark, in plain English?#

A benchmark is a standardised exam for software. Instead of a vendor demoing on documents they chose and questions they wrote, a benchmark fixes both in advance: here are the documents, here are the questions, here are the correct answers. The tool is scored the same way every time, so two different products — or two versions of the same product — can be compared on equal footing.

The critical word is public. A benchmark a vendor builds in-house has an obvious problem: they can tune the product to the test. A public benchmark, published by academics or an independent lab, can't be gamed the same way, because the vendor didn't write the rubric. As we cover in how to tell if an AI document tool is actually accurate, the single best question to put to any vendor is "what public benchmark did you run, and can I reproduce the result?" If the answer is a marketing percentage with no source, treat it as unverified.

Which benchmarks matter for legal and financial documents?#

Two stand out because they're built from exactly the kind of documents professional-services firms actually handle.

CUAD (Contract Understanding Atticus Dataset) is the standard academic benchmark for legal-contract understanding, released by The Atticus Project and published at NeurIPS 2021 (Hendrycks et al., The Atticus Project). It's a set of commercial contracts annotated for dozens of clause types — and it's deliberately adversarial. Many questions ask about clause categories that simply aren't present in the contract being queried, which means the correct behaviour is often to abstain rather than to answer.

FinanceBench is the equivalent for financial filings, released by Patronus AI and published at NeurIPS 2023 (Patronus AI, arXiv:2311.11944). The open-source release ships 150 real question-and-answer pairs drawn from 84 distinct SEC filings (10-Ks and 10-Qs) across 32 publicly-traded companies. The Patronus paper tested 16 model configurations — including GPT-4-Turbo, Llama-2 and Claude-2 — and found clear hallucination weaknesses across all of them. That last detail is why it's a useful benchmark: it's hard enough to expose the failure that matters.

Here's how the two compare as tests:

| | CUAD | FinanceBench | |---|---|---| | Domain | Legal contracts | SEC financial filings | | Published by | The Atticus Project (NeurIPS 2021) | Patronus AI (NeurIPS 2023) | | What it stresses | Clause retrieval + correct abstention | Numeric retrieval + hallucination | | Why it's hard | Asks about clauses often absent | Real filings, adversarial numeric Qs | | Correct "I don't know" | Frequently the right answer | Right answer when data is missing |

What does a good benchmark result actually look like?#

Not what most people expect. Because these benchmarks are hard on purpose, a lower pass rate paired with zero fabrication is a better outcome than a high pass rate that includes confident wrong answers.

To make that concrete, here are the results from SureCiteAI's own most recent published run (2026-04-27), which put the same retrieval pipeline that serves live customer queries through five suites plus the full 150-case FinanceBench set — 297 cases in total (SureCiteAI BENCHMARKS.md):

| Suite | Pass rate | Retrieval hit | Citation hallucinations | |---|---|---|---| | Accounting (SEC 10-Ks) | 35/35 (100%) | 23/23 | 0/35 | | Healthcare (openFDA) | 35/35 (100%) | 27/27 | 0/35 | | Legal (CUAD) | 17/35 (49%) | 19/27 | 0/35 | | FinanceBench (Patronus) | 95/150 (63%) | 144/150 (96%) | 0/150 | | Aggregate | 221/297 (74%) | 239/253 (94%) | 0/297 |

Read that legal row carefully, because it's the most instructive. A 49% pass rate on CUAD looks weak next to the 100% on the internal suites — and it's supposed to. CUAD is adversarial, and the "misses" are overwhelmingly cases where the tool correctly refused to answer a question whose answer wasn't in the contract. The honest low number on a hard public benchmark is more credible than a suspiciously perfect one on a tame internal test. That philosophy — "honestly low on hard public benchmarks rather than artificially high on tame internal ones" — is the whole point of publishing the score.

The number that carries the real weight is the last column: zero citation hallucinations across all 297 cases. For a citation-first tool, that's the load-bearing guarantee. A citation hallucination is when the AI cites a source document that wasn't actually in the material it retrieved — the most damaging failure mode for anyone whose work depends on traceable sources.

Why does abstention matter more than a high score?#

Because in legal and accounting work, a confident wrong answer is worse than no answer at all. This isn't a hypothetical concern. A Stanford HAI study of leading legal-research AI tools found that even purpose-built, RAG-based legal tools hallucinated on a meaningful share of queries — reported at roughly 17% for one major tool and 33% for another — while general-purpose models hallucinated on the majority of legal questions (Stanford HAI). The errors weren't just invented cases; they included mischaracterising real ones and citing authority that didn't apply.

That's why a benchmark that rewards abstention is measuring the right thing. Imagine a 12-lawyer firm running due diligence on a stack of vendor contracts. A tool that answers every question — including the ones the contracts don't address — will eventually hand an associate a fabricated indemnity clause that reads perfectly and survives review because it sounds right. A tool that abstains on those questions is slower to impress in a demo and far safer in production. The benchmark that penalises the first tool and rewards the second is the one you want.

We go deeper on this in why an AI should refuse to answer, and on the mechanics of source-linking in why AI document search needs citations.

What questions should you ask a vendor about their benchmarks?#

Use the benchmark literacy above as a checklist. Four questions separate a serious tool from a polished demo:

1. "Which public benchmark did you run?" A real answer names a specific external dataset — CUAD, FinanceBench, or similar — not an internal accuracy figure.
2. "Can I reproduce the number?" Reproducibility means a fixed set of cases, a fixed seed, and a script you could in principle re-run. If the result can't be reproduced, it can't be verified.
3. "What's your citation hallucination rate?" This should be zero, or as close as the tool can honestly hold it. Anything else means the tool sometimes cites sources it never actually read.
4. "What does it do when the answer isn't in my documents?" The right answer is "it tells you." A tool that always produces something is optimised to impress, not to be correct.

A vendor who welcomes these questions is telling you something. A vendor who deflects to a single marketing percentage is telling you something too.

The bottom line on AI document search benchmarks#

Benchmarks exist so you don't have to take a vendor's word for it. The ones that matter — CUAD for contracts, FinanceBench for filings — are hard on purpose, published by outsiders, and reproducible. On tests like these, the result to trust isn't a perfect score; it's a believable score paired with zero fabricated citations and correct abstention when the documents come up short. A tool that scores 49% on an adversarial legal benchmark by refusing to guess is doing its job. A tool that scores 99% on documents it was tuned against is doing a magic trick.

If you'd like to see the full methodology and the numbered runs behind the figures above — five suites, the RAGAS triad implemented in TypeScript, and calibration reported per-suite — it's all public and reproducible. Start a free SureCiteAI workspace and test it against your own documents, or read how to choose an AI document search platform for the wider buying checklist.

Frequently asked questions#

What is the best benchmark for AI document search?#

There's no single "best" one — the right benchmark depends on your documents. For legal-contract work, CUAD (The Atticus Project, NeurIPS 2021) is the standard academic test. For financial filings, FinanceBench (Patronus AI, NeurIPS 2023) is the most direct public comparison. What matters more than the specific benchmark is that it's public, reproducible, and hard enough to expose hallucination and over-confidence rather than hiding them.

Why would a tool score only 49% and still be trustworthy?#

Because the benchmark is adversarial. CUAD deliberately asks about clause types that often aren't in the contract, so the correct behaviour is frequently to abstain. A tool that "fails" those cases by refusing to fabricate an answer is behaving correctly. Paired with zero citation hallucinations, a modest score on a hard test is more credible than a perfect score on an easy, vendor-chosen one.

What is a citation hallucination?#

It's when an AI cites a source document — a filename, a page, a clause — that wasn't actually in the material it retrieved to answer your question. For any tool whose value rests on traceable answers, this is the most damaging failure, because the citation looks authoritative while pointing at nothing real. A well-built citation-first system holds this rate to zero by verifying every cited source against what was actually retrieved.

Do legal and financial AI tools really hallucinate?#

Yes, and more than their marketing suggests. A Stanford HAI study of leading legal-research tools found meaningful hallucination rates even in purpose-built RAG systems (around 17–33% for major legal tools, and higher for general-purpose models), including fabricated and mischaracterised cases. That's precisely why benchmarks that reward abstention, and citation verification that blocks invented sources, matter for professional work.

How can I verify a vendor's benchmark claim myself?#

Ask for the specifics: the name of the public dataset, the number of cases, the citation hallucination rate, and whether the run is reproducible with a fixed seed. If a vendor publishes numbered runs and the scripts behind them, you can in principle re-run the evaluation. If all you get is a percentage with no source, treat it as an unverified marketing claim.