llamaparse（llamaparse 文档）

LlamaParse

Parse documents (PDFs, images, spreadsheets, presentations — 130+ formats) into LLM-ready text, markdown, and structured data using the LlamaParse API.

Prerequisites

• Python package: llama-cloud>=1.0 (pip install llama-cloud)
• API key: Set LLAMA_CLOUD_API_KEY environment variable. Get one at https://cloud.llamaindex.ai

Verify setup:

pip install llama-cloud>=1.0
export LLAMA_CLOUD_API_KEY=llx-...

Quick Start

from llama_cloud import AsyncLlamaCloud
import asyncio

async def parse_document(file_path: str):
    client = AsyncLlamaCloud()  # Uses LLAMA_CLOUD_API_KEY env var
    file = await client.files.create(file=file_path, purpose="parse")
    result = await client.parsing.parse(
        file_id=file.id,
        tier="agentic",
        version="latest",
        expand=["markdown", "text"],
    )
    return result

result = asyncio.run(parse_document("document.pdf"))
print(result.markdown.pages[0].markdown)

Core Concepts

Tiers (required — choose one)

Always specify both tier and version. Use version="latest" for dev, or a date string like "2026-01-08" for production reproducibility.

Output Views (expand parameter)

Request one or more in the expand list:

• markdown — Structured markdown with headings, lists, tables. Best for RAG/LLM pipelines.
• text — Clean flattened text per page. Good for search/retrieval.
• items — Structured tree of page elements (headers, paragraphs, tables, figures) with bounding boxes. Use for layout-aware processing.
• metadata — Document metadata.
• images_content_metadata — Image/screenshot metadata with presigned URLs.

Access results: result.markdown.pages[i].markdown, result.text.pages[i].text, result.items.pages[i].items

Output Options

Control markdown rendering:

output_options={
    "markdown": {
        "tables": {
            "output_tables_as_markdown": True,  # or False for HTML tables
        },
    },
    "images_to_save": ["screenshot"],  # Save page screenshots
}

Processing Options

processing_options={
    "ignore": {"ignore_diagonal_text": True},
    "ocr_parameters": {"languages": ["en"]},  # OCR language hints
    "specialized_chart_parsing": "agentic_plus",  # Extract charts as structured data
}

Custom Prompts (Agentic Parsing Instructions)

Guide the parser like an LLM — useful for extracting specific data or transforming output:

from llama_cloud.types.parsing_create_params import (
    ProcessingOptions, ProcessingOptionsAutoModeConfiguration,
    ProcessingOptionsAutoModeConfigurationParsingConf
)

result = await client.parsing.parse(
    file_id=file.id,
    tier="agentic",
    version="latest",
    expand=["markdown"],
    processing_options=ProcessingOptions(
        auto_mode_configuration=[ProcessingOptionsAutoModeConfiguration(
            parsing_conf=ProcessingOptionsAutoModeConfigurationParsingConf(
                custom_prompt="Extract only prices and totals from this receipt."
            )
        )]
    ),
)

Common Workflows

Parse a single document

Use scripts/parse_document.py:

python scripts/parse_document.py document.pdf --tier agentic --output markdown,text

Batch parse a folder

Use scripts/batch_parse.py:

python scripts/batch_parse.py ./documents/ --tier agentic --max-concurrent 5

Extract tables from a document

Request items in expand, then filter for table items:

for page in result.items.pages:
    for item in page.items:
        if hasattr(item, 'rows'):  # Table item
            print(f"Table on page {page.page_number}: {len(item.rows)} rows")
            # item.csv, item.html, item.md available

Extract chart data

Enable specialized chart parsing, then pull table rows from the chart page:

result = await client.parsing.parse(
    file_id=file.id,
    tier="agentic_plus",
    version="latest",
    processing_options={"specialized_chart_parsing": "agentic_plus"},
    expand=["items"],
)

Download page screenshots

import httpx, re

result = await client.parsing.parse(
    file_id=file.id, tier="agentic", version="latest",
    output_options={"images_to_save": ["screenshot"]},
    expand=["images_content_metadata"],
)

for img in result.images_content_metadata.images:
    if img.presigned_url and re.match(r"^page_\d+\.jpg$", img.filename):
        async with httpx.AsyncClient() as http:
            resp = await http.get(img.presigned_url)
            with open(img.filename, "wb") as f:
                f.write(resp.content)

API Reference

For complete API details, see references/api-reference.md.

External Service & Security

This skill uses the LlamaParse API (https://cloud.llamaindex.ai), a cloud document parsing service by LlamaIndex.

• API key required: You must set the LLAMA_CLOUD_API_KEY environment variable. Get a key at https://cloud.llamaindex.ai.
• Data sent externally: Documents are uploaded to the LlamaParse API for server-side parsing. Parsed results are returned to your local machine.
• No other network calls: The scripts only communicate with api.cloud.llamaindex.ai. Screenshot downloads use presigned URLs from the same service.
• Scripts are reference utilities: scripts/parse_document.py and scripts/batch_parse.py are helper scripts meant to be run manually by the user. They are not executed automatically by the skill.

Tips

• Request only the expand views you need — more views = larger response + higher latency.
• Use agentic_plus tier with specialized_chart_parsing for documents with charts/graphs.
• For production, pin a specific version date instead of "latest".
• Use semaphore-based concurrency for batch parsing to respect rate limits.
• The items view provides bounding boxes (b_box) for each element — useful for spatial analysis.