LlamaParse

0.6.94 · deprecated · verified Sat Feb 28

GenAI-native cloud document parser by LlamaIndex for RAG-optimized output. Parses PDFs, PPTX, DOCX, XLSX, HTML and more into markdown, text, or structured JSON with accurate table extraction and multimodal support. Cloud API service — requires an API key from cloud.llamaindex.ai. NOT a local/offline tool. CRITICAL: The llama-parse package (and its successor llama-cloud-services) are DEPRECATED as of early 2026. The replacement is 'llama-cloud' (pip install llama-cloud), which targets LlamaParse API v2. The old packages are maintained until May 1, 2026 only.

Warnings

breaking llama-parse and llama-cloud-services are DEPRECATED. Both packages will receive no new features and are maintained only until May 1, 2026. New LlamaParse API v2 features are only available in the 'llama-cloud' package.
Fix: Migrate to: pip install llama-cloud. New import: from llama_cloud.services.parse import LlamaParse. Review the v1→v2 migration guide at developers.llamaindex.ai.
breaking LlamaParse API v2 changed target_pages from 0-based indexing to 1-based indexing. Code using target_pages='0,1,2' (v1) must be updated to target_pages='1,2,3' (v2). Silent wrong results if not updated.
Fix: Add 1 to all target_pages values when migrating from v1 to v2.
breaking v2 API removed save_images and take_screenshot boolean flags. Replaced by images_to_save parameter. In v1, save_images defaulted to True; in v2, images are NOT saved by default.
Fix: Explicitly set images_to_save in v2 if you need images extracted. Do not assume v1 image defaults carry over.
breaking parsing_instruction parameter (v1) is deprecated. Replaced by system_prompt + user_prompt combination in v2. Old parsing_instruction values are silently ignored in v2.
Fix: Rewrite parsing_instruction content as system_prompt and/or user_prompt in v2 configuration.
gotcha LlamaParse is a cloud API — it is not a local parser. All documents are sent to LlamaIndex servers. Not suitable for sensitive/private documents without a VPC or on-prem enterprise agreement.
Fix: For on-prem or data-sensitive use cases, contact LlamaIndex for enterprise/VPC options. There is no self-hosted OSS equivalent.
gotcha nest_asyncio.apply() is required in Jupyter notebooks and environments that already have a running event loop (e.g., FastAPI startup). Without it, calling sync methods like load_data() raises 'This event loop is already running'.
Fix: Add import nest_asyncio; nest_asyncio.apply() at the top of any notebook or async-host environment. New llama-cloud SDK has improved sync/async handling.
gotcha llama-parser (note: singular, no 'e') is a completely different, unmaintained package on PyPI released in 2024. pip install llama-parser installs the wrong package. The correct package names are llama-parse (deprecated) or llama-cloud (current).
Fix: Always install: pip install llama-cloud (new) or pip install llama-parse (old/deprecated). Never llama-parser.
gotcha Parsing tiers in v2 are: fast, cost_effective, agentic, agentic_plus. Using any other string (e.g. old v1 mode names) returns: 'Unsupported tier: must be one of: fast, cost_effective, agentic, agentic_plus'.
Fix: Use only the four valid tier strings. 'fast' is text-only spatial extraction. 'agentic_plus' is highest fidelity for complex layouts.

Install

pip install llama-cloud NEW — recommended for all new projects (API v2)
pip install llama-parse DEPRECATED — maintained until May 1, 2026 only (API v1)
pip install llama-cloud-services DEPRECATED — intermediate package, also replaced by llama-cloud

Imports

LlamaParse (new — llama-cloud)
```
from llama_cloud.services.parse import LlamaParse
```
New package is llama-cloud (pip install llama-cloud). Old import from llama_parse still works until May 2026 but points to deprecated API v1.
LlamaParse (old — llama-parse / llama-cloud-services)
```
from llama_parse import LlamaParse
```
Deprecated. No underscore variant. Package name is llama-parse, import is llama_parse.

Quickstart

Cloud API — requires internet and valid API key. Free tier available. For notebooks, call nest_asyncio.apply() before using sync methods or use aload_data() for async.

# NEW API (llama-cloud, v2) — recommended
# pip install llama-cloud
import os
from llama_cloud.services.parse import LlamaParse

parser = LlamaParse(
    api_key=os.environ["LLAMA_CLOUD_API_KEY"],
    tier="cost_effective",  # fast | cost_effective | agentic | agentic_plus
    result_type="markdown",
)

documents = parser.load_data("./my_file.pdf")
print(documents[0].text[:500])

# ---
# OLD API (llama-parse, v1) — deprecated, works until May 2026
# pip install llama-parse
import nest_asyncio
nest_asyncio.apply()  # required in notebooks

from llama_parse import LlamaParse

parser = LlamaParse(
    api_key=os.environ["LLAMA_CLOUD_API_KEY"],
    result_type="markdown",
    num_workers=4,
    verbose=True,
)

documents = parser.load_data("./my_file.pdf")
documents_batch = parser.load_data(["./file1.pdf", "./file2.pdf"])
documents_async = await parser.aload_data("./my_file.pdf")

view raw JSON →