Summary: LlamaParse, Unstructured and Docling all turn messy PDFs, slides and scans into text your RAG pipeline can use, but they charge in completely different ways. LlamaParse gives 10,000 credits every month and bills by tier, Unstructured gives 10,000 pages once and then charges $0.015 per page, and Docling is MIT-licensed and runs on your own machine for free. This comparison walks through setup, real per-page math, and the gotchas we found in each project's official docs on 2026-11-07.
Why the parser matters more than the model
When a retrieval-augmented app gives a wrong answer, people usually blame the LLM or the vector database. Very often the real problem happened earlier: a two-column PDF got read across the columns, a table was flattened into a row of numbers, or a scanned page came out as noise. Whatever the parser produces is all the model ever sees. LlamaIndex's own docs make the same point. For one short document, handing the raw PDF to a multimodal model often works. At scale, tables lose structure and token cost grows with every page whether it was useful or not.
The three tools here take different routes to the same goal:
- LlamaParse (from LlamaIndex) is a hosted API and web app. Its Parse product uses generative models for layout-aware parsing, and you choose one of four tiers per request.
- Unstructured comes in two forms: a hosted API/platform (Transform and Pipelines) with a page-based price, and an Apache-2.0 open-source Python library that its own docs describe as a starting point for prototyping.
- Docling is an open-source toolkit started by IBM Research Zurich and now hosted by the LF AI & Data Foundation. It is MIT-licensed, runs locally, and can work in air-gapped environments.
Plans and limits at a glance
All figures below come from each vendor's official pricing or docs pages, checked on 2026-11-07 (US pricing for LlamaParse).
| LlamaParse (hosted) | Unstructured (hosted API) | Docling (open source) | |
|---|---|---|---|
| Free allowance | 10,000 credits per month on Pay-as-you-go; unused monthly credits don't carry over | 10,000 free pages when the account starts (one-time), no card required | Free software, MIT license; you pay for your own hardware |
| Paid price | $1.50 per 1,000 credits (Pay-as-you-go); Pro is $500/month with $1.25 per 1,000 credits; Enterprise custom | $0.015 per page after the free pages; Business plan custom | No usage fee |
| Cost per page | Fast 1, Cost-effective 3, Agentic 10, Agentic Plus 25 credits; layout extraction +3 | Flat per page; .docx pages come from page metadata, and most other formats count as file size ÷ 100 KB | Your compute time only |
| Where it runs | LlamaIndex cloud (SaaS); hybrid/BYOC on Enterprise | SaaS; dedicated instance or in-VPC on Business | Your laptop, server, Docker or air-gapped box |
| Concurrency | 5 concurrent parse jobs (Pay-as-you-go), 20 (Pro), 100 (Enterprise) | Single user and workspace on Pay-As-You-Go; multi-user on Business | Limited by your hardware |
| Notable features | 130+ file types, Markdown/JSON/XLSX output, 48-hour free re-parse cache, version pinning | Partition, chunk, enrich and embed; 40+ connectors; 45+ file types on the API | PDF layout, reading order, tables, formulas; DoclingDocument JSON; CLI, MCP server and docling-serve |
What a page really costs: worked math
Because LlamaParse charges in credits and Unstructured charges by page, the sticker prices are hard to compare directly. Converting both to dollars per page with the published rates gives a clearer picture:
- LlamaParse Cost-effective (3 credits): $0.0045 per page on Pay-as-you-go.
- LlamaParse Agentic (10 credits): $0.015 per page on Pay-as-you-go, which is exactly Unstructured's per-page rate, or $0.0125 on Pro.
- LlamaParse Agentic Plus (25 credits): $0.0375 per page on Pay-as-you-go.
- Unstructured: $0.015 per page after the first 10,000 pages.
The free allowances work differently too. LlamaParse's 10,000 monthly credits come to about 1,000 Agentic pages, about 3,333 Cost-effective pages or 10,000 Fast pages a month, and the allowance resets each month. Unstructured's 10,000 pages are a one-time starting grant. If you only parse a few hundred pages a month for a side project, LlamaParse's monthly reset is the better deal. If you need to backfill one large archive, Unstructured's 10,000 pages arrive all at once.
One more Unstructured detail is easy to miss. For PDFs, PPTX and TIFF a page is a page, slide or image. But for formats without page metadata, the billed "page" count is the file size divided by 100 KB. A 2 MB HTML export, for example, counts as about 20 pages however long it reads.
Getting started with each tool
LlamaParse: web UI first, then the SDK
- Create an account and generate an API key (keys start with
llx-in the docs' examples). Export it asLLAMA_CLOUD_API_KEY. - To try it without code, open cloud.llamaindex.ai/parse, pick a tier under Recommended Settings, upload a file (or use a sample), and click Run Parse.
- For code, install the current SDK with
pip install "llama-cloud>=2.8", upload the file, then callclient.parsing.parse(...)withtier="agentic"andversion="latest", requestingexpand=["markdown"]. - Before production, swap
latestfor the dated version you tested so the output doesn't change when a new model ships. The docs say new tier versions arrive roughly monthly.
Two settings are mandatory on v2: tier and version. Leaving either out returns a validation error, which is the first thing older v1 tutorials will trip over.
Unstructured: hosted API or local library
- For the hosted route, sign up (the free plan needs no card) and call the API with the Unstructured Python SDK or any REST client such as cURL or Postman.
- For local prototyping, the official quickstart uses
uv: install the library withuv add "unstructured[pdf]", adding extras like[pdf,docx]for the formats you need. - Install the system packages the docs list:
libmagic-devfor file-type detection, pluspoppler-utilsandtesseract-ocrfor PDFs and images. - Call
partition_pdf(filename=...). By default it uses theautopartitioning strategy and returns typed document elements with metadata.
Unstructured is unusually honest about the gap between its two editions. The open-source library has no GPU support, no access to its fine-tuned OCR or VLM partitioning, no by-page or by-similarity chunking, and no image extraction. Judge Unstructured by its API, not by a quick local test.
Docling: pip install and convert
- Run
pip install docling(oruv add docling). It supports macOS, Linux and Windows on x86_64 and arm64. - On a CPU-only Linux server, the docs suggest adding
--extra-index-url https://download.pytorch.org/whl/cpuso pip doesn't pull a large GPU build of PyTorch. - In Python, create a
DocumentConverter(), callconvert(source)with a file path or URL, and export withexport_to_markdown(). - From the terminal,
docling <file-or-url>does the same, and--vlm-model granite_doclingswitches to the GraniteDocling vision-language model.
Docling also parses DOCX, PPTX, XLSX, HTML, EPUB, email files, images and even audio through ASR. Recent releases added video files, XBRL financial reports and chart-to-table conversion.
Which one should you pick?
- Choose LlamaParse if your documents are visually hard (dense financial tables, charts, multi-column scans) and you want to switch quality levels per job without running GPUs. The monthly free credits are plenty for evaluation. Use the Cost Optimizer on long mixed documents so simple pages bill at the cheaper tier.
- Choose Unstructured if parsing is one step in a bigger pipeline: pulling from SharePoint, Google Drive or S3, chunking, embedding and writing to a vector store. Its connector list and flat per-page price are the main draw, along with compliance options (SOC 2 Type 2, HIPAA, GDPR, ISO 27001) on the hosted product.
- Choose Docling if documents can't leave your network, you have predictable volume, or you're happy to trade some setup for zero per-page fees. It also plugs into LangChain, LlamaIndex, CrewAI and Haystack.
A practical approach is to take 20 of your own worst documents, run them through Docling locally and through the free hosted tiers of the other two, and compare the Markdown side by side. All three let you do that without paying.
Editor's note
Who it's for: developers building RAG, search or extraction over PDFs and Office files, especially with tables or scans. Not for: someone who only needs to read one PDF. A chat assistant that accepts uploads is simpler.
Gotchas we verified:
- LlamaParse credits reset monthly and don't roll over. Purchased credit packs expire one year after purchase.
- LlamaParse Pro shows two credit figures. The plan card advertises a limited-time 800k one-time bonus for business email domains, while the compare table and billing docs list 400,000 monthly credits. Budget on the 400,000.
- Re-parsing is free for 48 hours because LlamaParse caches results. Turn caching off only when you really need a fresh parse.
- LlamaParse 402 errors can come from storage limits, not only credits: a project at its retained-file or directory limit rejects new retained uploads with a
402. Upload files ephemerally or delete old ones. - The Fast tier runs no AI model, so custom prompts, forms and chart extraction don't apply there.
- Docling
ImportError: libGL.so.1in Docker or headless VMs comes from conflicting OpenCV packages. The FAQ's preferred fix is to uninstall bothopencv-pythonandopencv-python-headless, then reinstall the headless one. Installinglibgl1(Debian) ormesa-libGL(Fedora/RHEL) also works. - Docling offline: PDF processing needs model weights. Download them once and point
PdfPipelineOptions(artifacts_path=...)at that folder for air-gapped use. - Unstructured local install failures are usually missing system libraries (Poppler, Tesseract, libmagic), not Python errors.
FAQ
Is any of these completely free?
Docling is free open-source software under the MIT license. Unstructured's open-source library is Apache-2.0 and free too, but it's limited compared with its API. LlamaParse is a paid service with a 10,000-credit free monthly allowance on Pay-as-you-go. Unstructured's hosted API gives 10,000 free pages once.
Can I run LlamaParse locally?
Not on the self-serve plans. Hybrid or BYOC deployment is an Enterprise option. For a local, VLM-free option, LlamaIndex now offers a separate open-source parser called LiteParse, described as VLM-free, local parsing; the docs recommend it for plain text and simple layouts.
Which output format should I store for RAG?
Markdown is the common denominator: all three export it, and it keeps headings and tables readable for chunking. Keep the richer structured output as well (LlamaParse JSON, Unstructured elements, or Docling's lossless JSON) if you need page numbers or bounding boxes for citations.
Related posts on lightmind24
- Pinecone vs Qdrant vs Weaviate for Your First RAG App: where the parsed chunks go next.
- Firecrawl Tutorial: Scrape, Crawl and Map Sites for AI: the same idea for web pages instead of files.
- AnythingLLM Not Working? Fixes for Docs, Ollama and Embeds: when uploaded documents don't show up in answers.
Test your ugliest ten documents in all three before you commit. The parser you choose will shape your RAG quality more than almost any later tweak.
Last verified: 2026-11-07
Sources: LlamaParse pricing, LlamaParse credit pricing docs, LlamaParse billing, Parse overview, Parse tiers, Parse getting started, Unstructured pricing, Unstructured API overview, Unstructured open source overview, Unstructured quickstart, Docling docs, Docling usage, Docling FAQ, Docling on GitHub.
Image credit: official LlamaParse documentation image (LlamaIndex) plus our own screenshots of the public LlamaParse pricing, Unstructured pricing and Docling documentation pages, captured 2026-11-07. See the official screenshots at the end of this post.
댓글
댓글 쓰기