What Unstructured does
Unstructured provides open-source libraries and a platform for partitioning, enriching, chunking, embedding, and routing documents into AI data pipelines.
Unstructured addresses the ingestion layer that precedes search and RAG. Its open-source library can partition many file types in code, while the platform adds connectors, workflow scheduling, hosted processing, enrichment, chunking, embeddings, and destinations. Layout-aware strategies are useful for PDFs, scans, tables, slides, and documents that plain text extraction damages.
The library is open source, while hosted processing and the platform use usage-based or plan-specific limits; dedicated instances, VPC deployment, multi-user isolation, support, and enterprise terms are quoted. OCR, high-resolution processing, vision models, embeddings, source scans, destination writes, retries, storage, and egress can all affect total cost. Buyers should run a representative corpus through the live calculator or contract.
Successful parsing is not the same as faithful understanding. OCR can change numbers, tables can lose row relationships, headers can detach from clauses, and chunking can remove legal or scientific qualifiers. Source connectors may also over-collect files or flatten permissions. Minimize scope, preserve document and page provenance, enforce access before retrieval, quarantine hostile content, compare extractions with originals, and maintain deletion and reprocessing paths.
How Unstructured works
A workflow reads authorized files from an upload, API, or source connector and partitions each document into typed elements such as titles, narrative text, lists, images, and tables. Configured strategies may use OCR or vision models, then clean, enrich, chunk, and optionally embed the content before a destination connector writes it to storage or a vector database. Teams inspect extracted elements, provenance, permissions, and failures before indexing.
How to set up Unstructured
Inventory documents and rights
List formats, scans, languages, owners, licenses, PII, retention, permissions, expected tables and images, and prohibited sources.
Choose library or platform
Compare local open-source processing with hosted or dedicated workflows based on scale, models, connectors, region, security, and operations.
Build a representative extraction set
Include clean files, scans, columns, tables, handwriting, corrupted pages, long documents, and known answers for element-level comparison.
Configure provenance and access
Preserve source IDs, pages, versions, timestamps, and ACL metadata; scope connector credentials and prevent unauthorized chunks from entering retrieval.
Validate and operate
Measure extraction and table accuracy, review failed files, test updates and deletes, monitor spend and drift, and reprocess after parser changes.
Unstructured FAQs
Is Unstructured open source?
Yes. Its core document-processing library is open source; the managed platform, hosted processing, dedicated deployments, and support are commercial.
What does Unstructured output?
It partitions documents into typed elements with metadata, which can then be cleaned, enriched, chunked, embedded, or sent to downstream systems.
Can it process scanned PDFs?
Yes, supported strategies can use OCR and layout-aware or vision processing, but scan quality and complex layouts still require validation.
Does Unstructured preserve source permissions automatically?
Permission handling depends on connectors, metadata, destinations, and application logic. Teams must verify ACL propagation and enforce authorization before retrieval.
Is extracted content ready for production RAG?
Not automatically. Teams should test extraction fidelity, chunk boundaries, tables, provenance, access filters, injection, freshness, and deletion before generation.
Listing reviewed 2026-07-24. Product details and pricing can change; verify important terms on the provider's website.
Related Research AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.