ParseBench is a benchmark for evaluating how well document parsing tools convert PDFs into structured output that AI agents can reliably act on. It tests whether parsed output preserves the structure and meaning needed for autonomous decisions — not just whether it looks similar to a reference text.

The benchmark covers ~2,000 human-verified pages from real enterprise documents (insurance, finance, government), organized around five capability dimensions, each targeting a failure mode that breaks production agent workflows.

Prerequisites: Create a .env file with the API key for the parsing tool you want to evaluate (see Configuration for details).

# Install uv sync --extra runners # Quick test run (small dataset, 3 files per category — good for trying things out) uv run parse-bench run llamaparse_agentic --test # Full benchmark run (replace llamaparse_agentic with any pipeline name, see "Available Pipelines" below) uv run parse-bench run llamaparse_agentic # View interactive reports in your browser uv run parse-bench serve llamaparse_agentic

A pipeline is a document parsing tool or configuration that you want to evaluate. There are 90+ pipelines available -- see docs/pipelines.md for the full list, or run uv run parse-bench pipelines.

Hosted on HuggingFace: llamaindex/ParseBench

The dataset is stratified into five capability dimensions, each with its own ground-truth format and evaluation metric:

Content Faithfulness and Semantic Formatting share the same 507 underlying text documents, evaluated with different rule sets. Totals reflect unique pages and documents. Tables uses a continuous metric (no discrete rules).

What each dimension tests and why it matters for agents:

• Tables — Structural fidelity of merged cells and hierarchical headers. A misaligned header means the agent reads the wrong column when looking up a value.
• Charts — Exact data point extraction with correct series and axis labels from bar, line, pie, and compound charts. Most parsers return raw text instead of structured data, leaving agents unable to extract precise values.
• Content Faithfulness — Omissions, hallucinations, and reading-order violations. If the agent's context is incomplete or contains fabricated content, every downstream decision is compromised.
• Semantic Formatting — Preservation of formatting that carries meaning: strikethrough (marks superseded content), superscript/subscript (footnotes, formulas), bold (defined terms, key values), and title hierarchy. A strikethrough price is not the current price.
• Visual Grounding — Tracing every extracted element back to its source location on the page. Required for auditability in regulated workflows where every value must be traceable.

The dataset is automatically downloaded when you run a pipeline. To manage it manually:

# Download the full dataset uv run parse-bench download # Download a small test dataset (3 files per category, good for trying things out) uv run parse-bench download --test # Check whether the dataset has been downloaded and show summary statistics uv run parse-bench status

The run command runs inference (calls the parsing tool), evaluates the results against ground truth, and generates reports:

# Evaluate a parsing tool on all five dimensions uv run parse-bench run <pipeline_name> # Evaluate on a single dimension only (e.g., chart, table, layout, text_content, text_formatting) uv run parse-bench run <pipeline_name> --group chart # Skip calling the parsing tool — just re-evaluate existing results uv run parse-bench run <pipeline_name> --skip_inference # Control how many pages are processed in parallel uv run parse-bench run <pipeline_name> --max_concurrent 10 # Run on the small test dataset only (3 files per category, good for trying things out) uv run parse-bench run <pipeline_name> --test

When running all dimensions, the benchmark produces:

• Per-dimension detailed HTML reports with drill-down per test case
• An aggregation dashboard showing all dimensions side-by-side
• A leaderboard comparing all evaluated tools in the output directory
• CSV, Markdown, and JSON exports per dimension

# View reports in your browser (needed because browsers block PDF rendering from file:// URLs) uv run parse-bench serve <pipeline_name> # Compare two parsing tools side-by-side uv run parse-bench compare <pipeline_a> <pipeline_b> # Generate a leaderboard across all evaluated tools uv run parse-bench leaderboard # Leaderboard for specific tools only uv run parse-bench leaderboard llamaparse_agentic llamaparse_cost_effective

# Run inference only (call the parsing tool, don't evaluate) uv run parse-bench inference run <pipeline_name> # Run evaluation only (on existing inference results) uv run parse-bench evaluation run --output_dir ./output/<pipeline_name> # Generate detailed HTML report from evaluation results uv run parse-bench analysis generate_report --evaluation_dir ./output/<pipeline_name> # Regenerate the aggregation dashboard uv run parse-bench analysis generate_dashboard --evaluation_dir ./output/<pipeline_name>

/integrate-pipeline <name> <API docs or SDK link>

Each pipeline calls a specific parsing tool's API. You only need the API key for the tool you want to evaluate — add it to a .env file at the project root:

# Only add the keys you need. For example, to evaluate LlamaParse: LLAMA_CLOUD_API_KEY=... # To evaluate OpenAI-based pipelines: OPENAI_API_KEY=... # To evaluate Anthropic-based pipelines: ANTHROPIC_API_KEY=... # To evaluate Google-based pipelines: GOOGLE_API_KEY=...

ParseBench does not use LLM-as-a-judge — all evaluation is deterministic and rule-based. API keys are only used to call the parsing tool being evaluated.

Advanced subcommands: inference, evaluation, analysis, pipeline, data

output/ ├── _leaderboard.html # Cross-pipeline leaderboard └── <pipeline_name>/ ├── chart/ │ ├── *.result.json # Inference results │ ├── _evaluation_report.json # Evaluation summary │ ├── _evaluation_report_detailed.html # Interactive detailed report │ ├── _evaluation_results.csv # Per-example CSV │ └── _evaluation_report.md # Markdown summary ├── layout/ (same structure) ├── table/ (same structure) ├── text_content/ (same structure) ├── text_formatting/ (same structure) ├── _evaluation_report_dashboard.html # Aggregation dashboard └── _metadata.json # Run metadata 

src/parse_bench/ ├── cli.py # Fire CLI entry point ├── pipeline/cli.py # End-to-end pipeline orchestration ├── data/ │ ├── download.py # HuggingFace dataset download │ └── cli.py # Data management CLI ├── inference/ │ ├── runner.py # Batch inference with concurrency │ ├── pipelines/ # Pipeline registry (parse, extract, layout) │ └── providers/ # Provider implementations per product type ├── evaluation/ │ ├── runner.py # Parallel evaluation │ ├── evaluators/ # Product-specific evaluators (parse, extract, layout) │ ├── metrics/ # Metric implementations (TEDS, GriTS, rules, IoU) │ └── reports/ # CSV, HTML, markdown export ├── analysis/ │ ├── aggregation_report.py # Multi-category dashboard │ ├── detailed_report.py # Interactive per-category HTML report │ ├── comparison.py # Pipeline comparison │ └── comparison_report.py # Comparison HTML report ├── test_cases/ │ ├── loader.py # Load test cases (JSONL or sidecar .test.json) │ └── schema.py # TestCase types (Parse, Extract, LayoutDetection) └── schemas/ ├── pipeline_io.py # InferenceRequest, InferenceResult ├── evaluation.py # EvaluationResult, EvaluationSummary └── product.py # ProductType enum (PARSE, EXTRACT, LAYOUT_DETECTION) 

@misc{zhang2026parsebench, title={ParseBench: A Document Parsing Benchmark for AI Agents}, author={Boyang Zhang and Sebastián G. Acosta and Preston Carlson and Sacha Bron and Pierre-Loïc Doulcet and Simon Suo}, year={2026}, eprint={2604.08538}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2604.08538}, }

• Paper: arXiv:2604.08538
• HuggingFace Dataset: llamaindex/ParseBench
• Code: run-llama/ParseBench