Unstructured

Unstructured (Unstructured Technologies, unstructured.io) is a San Francisco company that turns messy enterprise documents (PDFs, Office files, HTML, emails, images) into clean, structured JSON "elements" for LLM and RAG pipelines. It rose to prominence on the back of its open-source Python ETL library (~15k GitHub stars, 1.3k forks), which became a near-default preprocessing layer in LangChain/LlamaIndex stacks. The commercial side is a hosted serverless Platform API plus a no-code workflow UI, layering chunking, embedding generation, table/image enrichment, generative (VLM) OCR, and 30+ source/destination connectors (Slack, Snowflake, S3, Zendesk) on top of the open-source core. The company raised a $40M Series B in March 2024 (led by Menlo Ventures, with NVIDIA's NVentures, Databricks Ventures, and IBM Ventures), bringing total funding to ~$65M at a reported ~$230M valuation.

Where it wins: breadth and pipeline ergonomics. It handles 25-64+ file types through one partition() entry point with selectable strategies (fast / hi_res / ocr_only / auto), and its newer VLM-based pipelines post strong numbers on its own published benchmark, leading on content fidelity (adjusted CCT 0.880 vs LlamaParse 0.835, Reducto 0.812) with very low hallucination/token-addition rates. The connector-plus-chunking story is genuinely differentiated for teams building ingestion at scale rather than just calling an OCR endpoint. Pricing is simple: 15,000 free pages, then a flat $0.03/page pay-as-you-go (an evolution from the earlier Fast $1/1k, Hi-Res $10/1k two-tier model), with custom Business/VPC/self-hosted tiers.

Where it loses: practitioners repeatedly note that the open-source library can be heavy and computationally expensive on long documents requiring OCR, and that hyperscaler document services (Google/Azure/AWS) and specialized parsers can beat it on hard layouts, tables, and non-Latin scripts. Sentiment is mixed-to-positive: people like it for "easy" text and flexible chunking/categorization, but some migrate to LlamaParse or cloud OCR for image-heavy and complex-table documents. It is best understood as an ETL/orchestration layer for document-to-LLM pipelines rather than a pure best-in-class OCR engine, though its VLM refinement options narrow that gap.