Unstructured

Editor’s note: I evaluate Unstructured’s documented capabilities, pricing, adoption signals, development momentum, openness, and exit costs alongside published benchmarks and practitioner feedback. This review is not a hands-on performance test.

Quick verdict: I recommend Unstructured for teams that need one pipeline turning 60+ file types from enterprise sources into RAG-ready, element-typed output. The trade-off: independent tests found it the slowest of three parsers in one and the weakest on complex tables in two, so test table-heavy work on the free tier first. In this Unstructured.io review, I’ll compare its pricing, document workflows, and alternatives, and explain how I weigh the vendor’s benchmark against two independent tests with conflicting results.

Key Takeaways 🔍

• Live pricing, September 2026: 10,000 free pages, then $0.015 per page, all features included; older articles print different figures
• Apache-2.0 open-source core: 15,515 GitHub stars, 60+ file formats, last pushed September 27, 2026
• Newest product: Foundation, an early-access MCP plugin giving coding agents enterprise file access
• Con: independent benchmarks put it slowest of three parsers in one test (51 to 141 seconds per document) and third of three on tables in another
• Con: the vendor’s own pages variously cite 30, 35, or 71 connectors

Pros and Cons

How Much Does Unstructured Cost?

Unstructured’s pricing page lists three tiers as of September 2026, and both self-serve tiers include every feature:

• Free ($0): 10,000 pages, no card required, all features included
• Pay-As-You-Go ($0.015/page): kicks in after the free 10,000 pages; all features; no monthly cap and no per-strategy price split published
• Business (custom): dedicated instance, VPC or multi-tenant SaaS deployment, multi-user accounts, dedicated technical support

Tier	Price	Deployment	Support
Free	$0 for 10,000 pages	Hosted API	No dedicated support tier listed
Pay-As-You-Go	$0.015 per page after 10,000	Hosted API	No dedicated support tier listed
Business	Custom	Dedicated instance, VPC, or multi-tenant SaaS	Dedicated technical support

An August 2026 comparison article printed different figures: 15,000 free pages per month, $0.03 per page, and a $3,000 monthly cap. The live vendor page in September 2026 shows 10,000 free pages, $0.015 per page, and no cap. Either pricing changed between those dates or the article was wrong, so treat unstructured.io/pricing as the source of truth and re-check it before you commit.

The open-source library is the real $0 column: Apache-2.0 licensed and self-hosted, with the fast and hi_res strategies included. The vlm strategy is exposed only in the paid API. The same partition logic moves from the library to the hosted API or the Business platform without a rewrite, so starting free costs you no rework later.

Is Unstructured Good Value for Money?

• Twelve times LlamaParse’s fast tier, roughly level with its agentic modes: LlamaParse’s fast tier runs around $0.00125 per page, and its agentic modes cost 10x to 45x that in credits; Panoply’s LlamaParse review maps that menu
• The volume math is blunt: 100,000 pages costs $1,350 on Pay-As-You-Go after the free 10,000, against about $125 at LlamaParse’s fast-tier rate
• Docling is free but self-run: zero license cost; you supply the compute and the ops
• The real cost variable is strategy choice: hi_res took 51 to 141 seconds per document in an independent benchmark, which is GPU or CPU time on your hardware if you self-host and wait time in your pipeline if you use the API
• $0.015 buys every feature: nothing parsing-related sits behind a higher tier; Business adds deployment options, multi-user accounts, and support

I’d choose Pay-As-You-Go for a RAG team ingesting under 100,000 pages a month after the free pages run out. I’d consider Business when a dedicated instance, VPC, multi-user accounts, or dedicated support justify the custom price.

Reviewer’s Notes 📝

I’d use the free 10,000 pages on your own document mix, with both fast and hi_res, before relying on any benchmark discussed here. Ten thousand pages is enough to parse your ugliest PDFs several times over and check the table output by hand.

Panoply reviewer

Getting Started With Unstructured

I’d follow the workflow in Unstructured’s documentation: install the library, partition one document, inspect the elements, then expand to the full pipeline.

Partitioning a Document

Setup is a pip install of the unstructured package, and the core of the product is one function. partition() auto-detects the file type; for PDFs, the documented function is partition_pdf from unstructured.partition.pdf. On a clean, text-native PDF, partition_pdf(filename=”report.pdf”, strategy=”fast”, extract_images=False) is the documented quick path.

What I like about the output is its structure: a list of typed elements (Title, NarrativeText, Table, ListItem) rather than a wall of text, each carrying metadata about where it came from. Run a scanned or mixed-layout file with strategy=”hi_res” and infer_table_structure=True and tables are preserved as structured HTML inside the element metadata instead of being flattened into prose.

Choosing a Strategy

Three strategies behave like three different tools:

• fast: pdfminer-based, seconds per document, built for simple digital PDFs
• hi_res: Detectron2/Chipper layout models, slower, better on tables and figures, GPU-optional
• vlm: vision-language models for image-heavy documents and screenshots, exposed only in the paid API

For financial filings, one practitioner’s published configuration is hi_res with table inference on, chunked by title so tables never mix with narrative text; its author self-reported 40% fewer hallucinated KPIs versus a plain pdfminer pipeline, a figure I have not independently verified. There is also a documented Docker path: pull unstructured-api:latest, run it on port 8000, and POST files with parameters like chunking_strategy=by_title and combine_under_n_chars=500.

The ETL Frame

The ingest CLI is where Unstructured stops looking like a parser and starts looking like a data pipeline. unstructured-ingest connects sources (S3, SharePoint, Google Drive, Slack, Airtable) to destinations (vector databases and warehouses), with flags for parallelism and reprocessing. The documented Airtable example goes from a personal access token to partitioned JSON in one command. I see it as an ingestion pipeline for LLM stacks, with more scope than a standalone PDF parser.

Reviewer’s Notes 📝

I would pay close attention to the strategy flag when balancing speed against fidelity. My reading is that differences in strategy may help explain why published Unstructured benchmarks disagree so widely. Before you trust any number in this review or anywhere else, run fast and hi_res on your own corpus and compare the output.

Panoply reviewer

The Benchmark Contradiction, Printed Whole

Unstructured’s own benchmark page says it wins. Two independent tests put it last: on speed in one, on tables in the other.

The vendor-published case: unstructured.io/benchmarks tests Unstructured against Reducto, LlamaParse, Docling, Snowflake, Databricks, NVIDIA Nemotron-Parse, and off-the-shelf frontier models on a 1,000-plus-page dataset of scanned invoices, complex layouts, nested tables, and handwritten notes. Unstructured’s best configuration, the VLM partitioner paired with GPT-5-mini, scored 0.883 on adjusted CCT, a 0.036 hallucinated-token rate, and roughly 0.82 on table cell accuracy, against competitor top scores of 0.625 to 0.835. The caveat: Unstructured chose the dataset, the metrics, and its own best configuration to feature. These are vendor-claimed wins, not independently reproduced.

One independent benchmark, run on five corporate sustainability PDFs, measured Unstructured at 75% cell accuracy on a three-level hierarchical table where Docling hit 97.9%, and called its table-of-contents output severely deficient (the title came back, no entries, no page numbers). The same test clocked it slowest of the three tools: 51.06 seconds for a one-page document, stretching to 141.02 seconds at 50 pages, against Docling’s 6.28 to 65.12 seconds and LlamaParse’s roughly 6 seconds regardless of length. Its authors did rate Unstructured’s text extraction and simple-table handling adequate.

A second independent test, on multi-column scientific papers, financial statements, and business memos, ranked Unstructured third of three on table completeness, accuracy, and structure, calling it “more like a generic document scanner than a table-first parser.”

My reading is that the tests may be measuring different configurations. The vendor benchmarked its paid VLM configuration; the independent tests most likely ran the default or hi_res paths, and did not always disclose the strategy. My takeaway: out-of-the-box table extraction lags Docling and LlamaParse, while the paid VLM path may not. That question is exactly what the free 10,000 pages exist to answer on your own documents.

Engineers on Hacker News add a third view: cloud OCR platforms (AWS Textract, Azure, Google) still lead for true structured extraction of forms and tables, and tools like Unstructured shine at segmenting text for RAG rather than defining document structure.

Foundation: The MCP Play

Unstructured’s newest product is not a parser at all. Foundation, now in early access, is an MCP (Model Context Protocol) plugin that lets coding agents (the vendor names Claude, Cursor, and Codex) read 65+ file types across SharePoint, Google Drive, Slack, S3, and databases, returning targeted context instead of flooding the model’s window with whole files.

The headline claims are vendor-stated, and I would not treat them as independently established results: 14x higher retrieval accuracy, 85% fewer agent turns, 10x lower costs. No independent test of Foundation is publicly documented yet, so treat those numbers as a pitch, not a result.

Unstructured does not market branded tiers like a Lite or an Enterprise edition. It presents one platform reached through API, UI, and now MCP, deployable bare metal, SaaS, or in-VPC. The scaling figure attached to that platform, over 15 million pages per hour per workflow, is likewise vendor-stated.

Strategically, Foundation points the same ETL engine at the agent market, and if your team runs Claude or Cursor against enterprise document stores, it is the one piece of Unstructured with no direct rival in this review. Because it is still in early access, I’d evaluate it before making it a standard part of an enterprise workflow.

Adoption, Momentum, and the Counting Problem

The Apache-2.0 repository is actively maintained: 15,515 GitHub stars, 1,349 forks, and 319 open issues, with the last push on September 27, 2026, two days before the September 29 research date. PyPI version 0.27.10 shipped the same day and supports Python 3.11 through 3.13.

The vendor’s own spec sheets disagree on counts. Its docs cite 35+ connectors, its product page cites 30+ sources plus 30+ destinations, and one of its own articles claims over 71 pre-built connectors; file-type counts drift between 60+ and 65+. My best explanation is that the biggest number bundles destinations and LLM and vector-database integrations in with source connectors. I wouldn’t rule it out over inconsistent counts, but I’d verify the specific connector you need before committing.

A GitHub issue (#2557) reporting that both fast and hi_res ran slower than expected on larger files was closed back in May 2024, yet the 2025 independent benchmark above measured the same slowness. That benchmark’s authors named speed and complex-table parsing as the two areas Unstructured most needs to improve. I’d plan around hi_res latency rather than assume a future update will remove it.

How Does Unstructured Compare to Competitors?

Every rival in this category wins somewhere specific, and two of them beat Unstructured outright in the independent tests above.

• Docling had the best overall accuracy and table-structure preservation in both independent benchmarks: 97.9% table-cell accuracy and a 100% sentence match in one test. It is free, has IBM Research lineage, and runs at a moderate, predictable speed that scales linearly with page count. Its noted gap is flat Markdown heading levels, which lose parent-child section hierarchy. I’d pick it for out-of-the-box table fidelity at zero license cost if you’re willing to run it yourself.
• LlamaParse was the fastest tool tested, roughly 6 seconds per document regardless of length, with the best row and column structure preservation in one benchmark and a fast tier around $0.00125 per page. Its failure mode is systematic misplacement in complex tables (100% of the data extracted, 0% placed correctly in one test), and it also stumbles on multi-column layouts, word merging, currency symbols, and footnotes. Its credit pricing jumps 10x to 45x across modes; Panoply’s LlamaParse review covers that menu.
• Cloud OCR (AWS Textract, Azure Document Intelligence, Google) is the practitioner pick for true structured extraction. Engineers on Hacker News report the cloud platforms outperform both Unstructured and LlamaParse on forms and complex tables, and position Unstructured as a RAG segmentation tool first, not a structure-definition tool.
• Reducto appears only as a comparator (in its Agentic mode) in Unstructured’s own benchmark. No independent benchmark of it is publicly documented, so if it is on your shortlist, run it on the same documents you put through Unstructured’s free tier.

My decision rule: many file types plus enterprise connectors plus RAG-ready elements means Unstructured; maximum table fidelity out of the box means Docling; speed and a LlamaIndex-native flow means LlamaParse; forms and high-stakes structure mean cloud OCR.

How I Reviewed Unstructured

My assessment draws on Unstructured’s documentation and pricing, its GitHub and PyPI activity, published benchmarks, and practitioner discussions. I weigh its file-format coverage, connector support, deployment options, and exit costs against Docling, LlamaParse, and cloud OCR services. The performance figures cited here come from vendor and third-party tests, not tests I ran myself.

I give more weight to the documented workflow and specific benchmark limitations than to aggregate star ratings. I distinguish vendor claims from independent findings and recommend checking performance on your own documents. Prices current as of September 2026.

Unstructured Review: Should You Feed Your RAG Stack With Unstructured?

I recommend Unstructured as the ingestion backbone for teams pulling many file types out of enterprise sources and into vector stores. If your pipeline touches PDFs, Slack exports, SharePoint libraries, and S3 buckets in the same week, the 60+ formats, the 30+ connectors, and element-typed output that keeps tables separate from narrative text are the practical argument. The Apache-2.0 core is real, actively maintained, and a legitimate exit path. Foundation extends the same engine to coding agents, though as an early-access product I treat it as a bonus rather than a reason to buy.

I’d test before buying if table extraction is your main requirement. Two independent benchmarks put out-of-the-box table extraction behind Docling, and hi_res took 51 to 141 seconds per document in one of them, a pace that will hurt latency-sensitive pipelines. For single-format PDF-to-table jobs, I’d start with Docling or a cloud OCR service based on the independent evidence discussed above.

My recommended next step costs nothing: run the free 10,000 pages against your ugliest documents, fast versus hi_res, and check the extracted tables by hand. The benchmark that matters is yours.

FAQ

Is Unstructured free?

Partly. The open-source library is Apache-2.0 licensed and free to self-host, with the fast and hi_res strategies included. The hosted API adds 10,000 free pages with all features and no card required, then bills $0.015 per page. The vlm strategy is only available through the paid API.

Is Unstructured accurate for tables?

It depends on whose test you trust and which strategy you run. The vendor’s own benchmark, on a dataset and metrics it chose, reports roughly 0.82 table cell accuracy for its best paid VLM configuration. Two independent benchmarks tell a different story: 75% cell accuracy on a complex hierarchical table versus Docling’s 97.9%, and a last-of-three table ranking. Test both strategies on your own tables.

What is the difference between Unstructured and LlamaParse?

Unstructured is a document ETL platform: connectors, partitioning, chunking, enrichment, and destinations in one pipeline. LlamaParse is a parse-first API with credit-priced quality tiers, built for the LlamaIndex ecosystem. LlamaParse was faster in independent testing (roughly 6 seconds per document), and its fast tier costs around $0.00125 per page; Unstructured covers far more file types and enterprise sources. Panoply’s LlamaParse review covers its mode pricing.

What is Unstructured Foundation?

Foundation is an early-access MCP plugin that lets coding agents such as Claude, Cursor, and Codex read 65+ file types across SharePoint, Google Drive, Slack, S3, and databases. The vendor claims 14x retrieval accuracy, 85% fewer agent turns, and 10x lower costs; none of those figures is independently verified.

Why do published Unstructured prices differ?

Either pricing changed or older articles were wrong. An August 2026 comparison printed 15,000 free pages at $0.03 per page with a $3,000 monthly cap; the live vendor page in September 2026 shows 10,000 free pages at $0.015 per page with no stated cap. Check unstructured.io/pricing at decision time.