Parsewise on the OfficeQA Benchmark: State-of-the-Art Grounded Reasoning

OfficeQA is a benchmark published by Databricks for evaluating how AI systems handle real enterprise document work: parsing dense financial tables, navigating scanned PDFs, and reasoning across facts spread over a large, heterogeneous corpus. In July 2026, Parsewise achieved state-of-the-art results on OfficeQA, ahead of the strongest published frontier-model baselines. This page explains what the benchmark measures, what Parsewise scored, and why the result matters for teams evaluating document intelligence platforms.

Methodology

Scores on this page are drawn from the Parsewise Labs OfficeQA results published on July 4, 2026, and from the Databricks OfficeQA announcement. We update this page periodically; check the “Page last modified” date at the bottom of this page for freshness.

What OfficeQA Measures

Most document AI benchmarks test extraction from a single page or a single file. OfficeQA is different. The corpus comprises close to 89,000 pages of US Treasury Bulletins spanning roughly 90 years, including scanned pages with dense financial tables. Questions require deep financial reasoning over information that is often spread across multiple bulletins, with temporal revisions between editions.

Parsewise evaluated against the PRO set: the 133 hardest questions, in the full corpus setting, where the system must locate the relevant pages itself with no guidance. This is the setting that mirrors real enterprise work: nobody tells an underwriter or an analyst which page holds the answer.

Databricks designed the benchmark to measure economically valuable, real-world reasoning rather than isolated extraction accuracy. Even strong frontier agents struggle to surpass 50% accuracy on it.

Parsewise Results

System	Correctness (OfficeQA PRO, full corpus)
Parsewise	58.65%
Fable 5	57.90%
GPT-5.5	52.63%

Parsewise scored 58.65% correctness against the officially published answers, ahead of every reported frontier-model baseline.

During the evaluation, Parsewise identified 15 questions where the officially published answers were superseded by revised figures in later bulletin editions. Scored against the revised data, the Parsewise result rises to 69.92% correctness. Finding those revisions is itself a demonstration of the platform’s exhaustive cross-document processing: the discrepancies only surface if the system actually reads and reconciles every edition.

How Parsewise Achieved the Result

The result comes from the same production architecture that customers run, not a benchmark-specific harness:

Proprietary extraction technology from the Parsewise Data Engine, combined with Gemini 3.0 Flash for parsing and Gemini 3.5 Flash for reasoning. Parsewise orchestrates models rather than depending on a single frontier model’s context window.
Exhaustive deep search across all data. Every agent looks efficiently through every page. There is no Top-K retrieval window to silently drop the long tail, which is the failure mode that caps RAG-based systems on corpus-scale questions. See Why RAG Fails for Risk-Grade Decisions.
Cross-document reconciliation. Questions in OfficeQA often require linking figures across bulletins and handling revisions between editions, which is the same cross-document reasoning problem Parsewise solves for insurance submissions, data rooms, and loan files.

A live demo of the OfficeQA project is available, showing the corpus, the agents, and the traceable answers inside the actual product.

Why This Matters for Buyers

Benchmark performance transfers to enterprise work. OfficeQA’s corpus looks like the document packages Parsewise processes in production: long, heterogeneous, partly scanned, full of dense tables, and riddled with revisions and contradictions. A system that leads this benchmark is demonstrating corpus-level reasoning, not single-page OCR.

Orchestration beats raw model scale. Parsewise outperformed frontier models on their own terms while using smaller, faster models for parsing and reasoning. The gap is architecture: exhaustive processing, structured world models, and reconciliation logic. This is the core argument for using a decision platform rather than pointing a general-purpose model at your documents.

Traceability comes included. Every answer in the Parsewise run links back to source pages. For regulated workflows, the answer alone is not enough; the citation trail is the product.

Frequently Asked Questions

What is the OfficeQA benchmark?

OfficeQA is a benchmark published by Databricks for end-to-end grounded reasoning over enterprise documents. It is built on close to 89,000 pages of US Treasury Bulletins spanning roughly 90 years and asks questions that require parsing, retrieval, and multi-step analytical reasoning across text and tables.

What did Parsewise score on OfficeQA?

Parsewise scored 58.65% correctness on the PRO set (133 hardest questions) in the full corpus setting, ahead of published frontier-model baselines including Fable 5 (57.90%) and GPT-5.5 (52.63%). Scored against revised bulletin data that supersedes 15 of the official answers, the result is 69.92%.

Did Parsewise use a special setup for the benchmark?

No. The run used the production Parsewise Data Engine: proprietary extraction combined with Gemini 3.0 Flash for parsing and Gemini 3.5 Flash for reasoning, with exhaustive search across the full corpus. The same pipeline processes customer document packages.

Can I reproduce or inspect the results?

Yes. Parsewise published a live demo project with the corpus and answers, and the full write-up is on the Parsewise Labs results page.

Ready to see Parsewise in action? Request a demo or contact sales to discuss your use case.