LlamaParse

LlamaParse is a GenAI-native document parsing client from LlamaIndex. It sends unstructured files - PDF, PPTX, DOCX, XLSX, HTML - to the LlamaCloud parsing API and returns clean text or Markdown, accurately extracting tables, visual elements, and awkward layouts for downstream LLM use in RAG and agent pipelines.

The Python package wraps the LlamaCloud API behind a simple LlamaParse client with sync, async, and batch loaders, plus a llama-parse command-line tool. It integrates directly with LlamaIndex readers and supports custom parsing instructions, multimodal extraction, and multi-worker throughput. Note: LlamaParse is part of the LlamaCloud Services repository, which is deprecated in favor of the newer llama-cloud packages, but remains widely used.

What You Get

• A LlamaParse client with sync load_data, aload_data, and batch parsing across multiple files
• Text and Markdown output modes, plus raw JSON for structured downstream processing
• Table recognition that preserves tabular data as text or semi-structured output
• Multimodal parsing that extracts images and diagrams using vision models
• Custom parsing instructions and a llama-parse CLI for one-off file conversion
• Native integration with LlamaIndex document readers and RAG pipelines

Common Use Cases

• Parsing PDFs with dense tables into Markdown for a RAG knowledge base
• Batch-converting slide decks and spreadsheets into LLM-ready text
• Feeding cleanly structured document text into an agent’s tool context
• Extracting figures and diagrams from technical documents via multimodal parsing

Under The Hood

Architecture - The llama-parse package is now a thin compatibility layer over llama-cloud-services: base.py simply re-exports LlamaParse, ResultType, ParsingMode, and the job-route constants from llama_cloud_services.parse.base, while cli/main.py provides the llama-parse command. The real work is remote - the client uploads a file to the LlamaCloud parsing service, polls a job status route, and fetches the structured result, so the library is orchestration around an asynchronous parsing API rather than a local parser.

Tech Stack - Pure Python packaged with Hatch, declaring a single runtime dependency on llama-cloud-services>=0.6.94. It lives in the run-llama/llama_cloud_services monorepo (which also hosts TypeScript packages under ts/) and exposes a console-script entry point for the CLI.

Code Quality - The wrapper is deliberately minimal and explicit about its public surface via __all__, delegating parsing, retries, and job management to the shared llama_cloud_services core that carries the monorepo’s tests/unit_tests suites. The repository is under active use but the package itself is formally deprecated in favor of llama-cloud.

API Design - The LlamaParse client is ergonomic: construct it with an API key and result_type, then call load_data, aload_data, or batch-load multiple files with num_workers. Results come back as LlamaIndex Document objects for drop-in RAG use, and options like language, custom parsing instructions, and Markdown/text/JSON output modes keep the common cases to a few lines.