LlamaParse: Document Parsing for LLM and Agent Pipelines

LlamaIndex, a private company founded in 2022 and headquartered in San Francisco, CA, ships two complementary document parsing products. LlamaParse is a cloud service for complex layout-aware extraction targeting production document intelligence. LiteParse is an open-source local CLI released in March 2026 for agent-native pipelines where speed, privacy, or zero cost are the constraint. The two products occupy opposite ends of the same parsing spectrum, designed to route users to the right tier rather than compete with each other.

In June 2026, LlamaIndex announced a $19M Series A led by Norwest Venture Partners with participation from Greylock Partners, bringing total capital raised to $27.5M. Named enterprise customers include Rakuten, Carlyle, Salesforce, and KPMG. The company achieved SOC 2 Type 2 certification, with GDPR-compliant EU data residency in early access. Employee count and revenue are not publicly disclosed.

Products

LlamaParse v2 replaced manual model configuration with four outcome-focused tiers: Fast (1 credit/page), Cost Effective (3 credits/page at 0.38¢), Agentic (10 credits/page), and Agentic Plus (45 credits/page at 1.25¢). Agentic Plus pricing dropped 50% from the prior version. Version pinning in YYYY-MM-DD format lets production workflows lock parsing behavior across model upgrades, addressing a recurring complaint about silent output changes when vendors update underlying models.

The LlamaParse API v2 reorganizes configuration into three structured objects: input_options, output_options, and processing_options. New llama-cloud SDK packages for Python and TypeScript replace the older llama-cloud-services packages. API v1 continues to be supported during migration.

ParseBench: performance in context

In June 2026, LlamaIndex published ParseBench (arXiv:2604.08538), a benchmark of 2,078 human-verified pages from 1,211 enterprise documents across insurance, finance, and government domains. It evaluates 90+ pipelines across five dimensions: table accuracy, chart data extraction, content faithfulness, semantic formatting, and visual grounding.

LlamaParse Agentic scored 84.88 overall, ranking first. For context, per Miles K.'s independent summary: Google Gemini 3 Flash scored 71.0, Reducto Agentic scored 67.8, Azure Document Intelligence achieved the best non-LlamaParse visual grounding score (73.8) but near-zero chart performance (1.6%), and AWS Textract scored 47.9 overall. On cost, LlamaParse Agentic runs at 1.2 cents per page versus Gemini 3.1 Pro at 8.49 cents and GPT-5.4 at 2.90 cents.

Semantic formatting scores (strikethrough, superscripts, bold, title hierarchy) ranged from 1.0% (Docling) to 85.2% (LlamaParse Agentic). Most IDP vendors and benchmarks ignore this dimension entirely. For legal documents and financial filings where strikethrough text indicates amendments and superscripts indicate footnote references, stripping this formatting changes document meaning.

One credibility caveat applies: ParseBench was designed, curated, and scored by the LlamaIndex team evaluating their own product. The benchmark code and dataset are publicly available on GitHub and HuggingFace for independent replication. The arXiv paper acknowledges this conflict and invites external validation. Until independent researchers replicate results, the 84.88 figure should be read as vendor-associated data, not independent certification.

LiteParse: local parsing for agents

LiteParse is TypeScript-native, built on PDF.js and Tesseract.js, with zero Python dependencies. It runs entirely on local CPU, outputs spatial text with preserved indentation, page screenshots, and JSON metadata with bounding boxes. Language bindings cover Node.js/TypeScript (napi-rs), Python (PyO3), Rust (cargo), and browser/WASM (wasm-bindgen).

The routing logic is explicit in the LiteParse documentation: for complex documents with dense tables, multi-column layouts, charts, or scanned PDFs, LlamaParse delivers significantly better results. LiteParse is the right choice for well-structured files where cloud latency or cost is the constraint.

As Jia Chen at softmaxdata.com notes: "The bottleneck in many agentic workflows isn't parsing quality, it's parsing latency. An agent that waits 10 seconds for a cloud API to return can't iterate quickly. An agent that gets rough-but-fast text in 100 milliseconds can scan dozens of documents, identify the relevant pages, and then selectively invest in deeper analysis."

LiteParse includes complexity detection to flag files that would benefit from LlamaParse's heavier processing before committing to full local parsing. MindStudio's March 2026 analysis notes that layout-aware spatial parsing is what separates useful parsed output from noise in privacy-sensitive deployments, and LiteParse's self-hosted model directly addresses compliance requirements in healthcare, legal, finance, and government sectors.

One output format discrepancy is worth flagging: the official LlamaIndex blog describes LiteParse as preserving spatial layout as ASCII-formatted text, while the GitHub repository describes Markdown output with reconstructed structure. Both modes likely exist. Evaluators should verify current behavior against the documentation before committing to a pipeline design.

Granular bounding boxes

LlamaParse introduced granular bounding boxes in beta across all paid tiers, enabling line-level, word-level, and cell-level coordinate tracking for extracted content. The feature activates via a single granular_bboxes parameter at job creation. Only explicitly visible text receives coordinates; inferred values, AI summaries, and reconstructed content do not. Word-level coordinates enable high-precision PII redaction without manual page-blocking. The Agentic Plus tier runs additional verification rounds for attribution-sensitive workflows.

Use cases with evidence

The LlamaIndex team documented a loan underwriting pipeline built at a New York City developer workshop: LlamaParse parsed pay stubs and brokerage statements, extracted structured data into Pydantic models, and ran cross-document analysis using async Python, SQLite, and FastAPI. This is a vendor-documented outcome, not an independent case study.

Davor Bonaci, CTO at DataStax, stated: "By incorporating LlamaIndex into RAGStack, we are providing enterprise developers with a comprehensive Gen AI stack that simplifies the complexities of RAG implementation."

Competitive context

For teams evaluating alternatives, Unstract offers a no-code large language model (LLM) platform with hallucination mitigation for production-grade extraction. Hyperscience and Instabase target similar enterprise document automation segments at higher price points. ABBYY and Tungsten Automation dominate regulated-industry on-premise deployments, the segment LiteParse now enters from the open-source side.

Independent practitioners describe a three-tier production architecture: LiteParse for fast local extraction, a capable on-premise model (such as GLM-OCR from Tsinghua KEG Lab) as a middle tier for teams with GPU infrastructure, and LlamaParse for complex layouts where accuracy cannot be compromised. As Dr. Prahlad G. Menon writes at themenonlab.blog: "Start with LiteParse. If your documents have complex layouts or low text-layer quality, reach for GLM-OCR before paying for cloud. Reserve LlamaParse for the cases where accuracy genuinely can't be compromised and data residency isn't a constraint."

Known limitations and roadmap

The LlamaIndex team has publicly committed to whole-document parsing improvements for tables split across pages and title hierarchy across page breaks, addressing a structural limitation of page-by-page processing architectures. Neither has a confirmed release date.

Charts remain the most polarizing dimension in ParseBench: only four providers exceed 50% accuracy, while most specialized parsers score below 6%. LlamaParse Agentic scored 78.1% on chart extraction versus AWS Textract at 6.0% and Azure Document Intelligence at 1.6%, reflecting its vision-language model integration. Traditional optical character recognition (OCR) pipelines do not attempt to convert chart visuals into structured data at all.

Technical specifications

Resources

• LlamaParse developer documentation
• Python SDK on PyPI
• API sandbox
• LiteParse announcement
• LiteParse npm package
• LiteParse GitHub
• LiteParse documentation
• ParseBench paper
• API v1 to v2 migration guide