🗞️ Ever wonder how your parser knows where to start reading in a newspaper layout? Reading order isn't as obvious as it looks. Columns, pull quotes, captions, a headline spanning three of them. Read it straight across and the text turns into nonsense. Unstructured works out the exact sequence a human would actually follow, then labels and extracts each piece in that logical flow. See it on your own documents: https://lnkd.in/ebhGexr9
Unstructured
Data Infrastructure and Analytics
San Francisco, CA 30,379 followers
Stop dilly-dallying. Get your data.
About us
Unstructured is the data infrastructure company solving the most critical bottleneck in enterprise AI: making unstructured data accessible to AI applications. Trusted by 87% of the Fortune 1000, we transform the 80–90% of enterprise information trapped in inaccessible formats—PDFs, Word docs, PowerPoints, emails, HTML, and 70+ other file types—into clean, AI-ready data with industry-leading accuracy and performance benchmarks. Companies that try to build and maintain custom data pipelines in-house find it's a significant and ongoing engineering drain. Unstructured replaces that entirely, enabling enterprises to move from experimental workflows to AI applications that execute real business value. Recognized by Forbes AI50, Fast Company's Most Innovative Companies, and CB Insights AI 100, Unstructured is the data foundation that makes enterprise AI work.
- Website
-
http://www.unstructured.io/
External link for Unstructured
- Industry
- Data Infrastructure and Analytics
- Company size
- 51-200 employees
- Headquarters
- San Francisco, CA
- Type
- Privately Held
- Founded
- 2022
- Specialties
- nlp, natural language processer, data, unstructured, LLM, Large Language Model, AI, RAG, Machine Learning, Open Source, API, Preprocessing Pipeline, Machine Learning Pipeline, Data Pipeline, artificial intelligence, and database
Employees at Unstructured
View 111 employees at Unstructured
or
Locations
-
Primary
Get directions
San Francisco, CA, US
Updates
-
Fine-tuning Object Detection for documents is harder than it looks. The short version: Object Detection (OD) is the foundation under every document pipeline, and small inconsistencies in bounding boxes cascade into measurable errors in OCR, reading order, and table structure. Getting it right takes consistent annotations, careful control of data distribution, and a lot of debugging that most teams underestimate. 🫠 If you've ever assumed OD is a solved problem, you'll wanna check this out: https://lnkd.in/eqvxbRJ5
-
We're at #AFANational!! ✈️ If you're attending, keep an eye out for Marshall Leipprandt 👀 He'll be there, ready to chat all things Air and Space Forces. Learn more about how Unstructured helps companies deliver next-generation agentic AI systems using mission-ready data: https://lnkd.in/eg6cPU8M
-
-
"We already use Textract, it works fine, why would we add Unstructured?" It does work fine, until the requirements grow. Then it is scanned PDFs, complex tables, slide decks, multi-cloud, chunking, enrichment, retrieval, and suddenly your document pipeline is an infrastructure project in disguise, owned by people who never signed up to maintain one. The honest math is not the per-page rate. It is the engineering months you spend holding the homegrown version together. We wrote up where DIY quietly gets expensive: https://lnkd.in/eYZTKJTt
-
-
Unstructured reposted this
Just built my first simple RAG (Retrieval-Augmented Generation) system from scratch. What it does: A conversational Q&A system that answers questions grounded in your own documents, with full chat memory, so follow-up questions like "when did that happen?" actually work. Stack: - LangChain for orchestration - OpenAI (text-embedding-3-small + gpt-4o) for embeddings & generation - ChromaDB as the vector store - Character-based chunking (fixed-size splitting) to break documents into retrievable pieces The interesting part: history-aware retrieval: Naive RAG breaks on multi-turn conversations. If you ask "What did Tesla release?" then "When?", a plain vector search on "When?" retrieves garbage. I implemented a query-reformulation step that rewrites follow-up questions into standalone, searchable queries using the conversation history before hitting the vector store. A core pattern in production conversational RAG systems. What's next: - Swapping basic chunking for unstructured.io, smarter, structure-aware chunking with native support for PDFs, Word docs, HTML, etc., instead of blind character splitting. - Wrapping this in a FastAPI backend. - Building a Next.js/React frontend. - Making it production-ready end-to-end. This has been such a great way to actually understand what's happening under the hood of RAG instead of just calling an API. More to come as I build this into a full-stack product! #RAG #LangChain #OpenAI #ChromaDB #AI #MachineLearning #LLM #BuildInPublic
-
Good document parsing starts before the parsing itself. A deeper look at the role object detection plays behind the scenes, and why getting it right matters 👇
Everyone's talking about VLMs parsing documents end to end. But there's a quieter layer underneath that decides whether any of it works: Object Detection. Object Detection is the quiet layer that decides whether document parsing *actually* works. Before a page can be read, something has to segment it into tables, headers, blocks, and formulas. Get those bounding boxes wrong and everything downstream pays for it. VLMs receive oversized or truncated crops, prompts turn brittle, and pipelines bloat with code that exists only to clean up bad segmentation. Every point of Object Detection accuracy you gain shows up directly in OCR, reading order, and table extraction quality. So we went deep on fine-tuning our own OD model, and it was a lot harder than the usual "add data, tweak hyperparameters, train" story suggests. Aligning annotation philosophies across datasets, holding onto prior capabilities while adapting to tighter boxes, keeping bounding boxes consistent enough that downstream cropping stays clean. None of it is plug and play. But here's the upside: the payoff is worth it. OD is a ton of effort, and when you get it right, it lifts everything above it. Tighter boxes, cleaner crops, better tables, more reliable reading order, all from nailing the foundation. So we do the hard part, and you get to focus on what you build with the outputs 💅 Read the full write-up: https://lnkd.in/eqvxbRJ5
-
Some text is built to be misread. Rotated sideways, stamped as a watermark, set in a font nobody has shipped since 1998, or written in a script your OCR never trained on. Traditional OCR reads it wrong, and that wrong text lives in your index forever. Generative OCR catches those cases and routes them to a vision language model, so the hard-to-read parts come out accurate instead of quietly corrupted. Try it on your own docs: https://lnkd.in/ebhGexr9
-
-
Document parsing changed more in the last year than in the five before it. VLMs went from novelty to the default way serious teams pull structure out of hard documents. Our "State of Generative Parsing 2026" whitepaper lays out where the field actually is: what generative parsing does well, where it still breaks, and how to evaluate it without fooling yourself. Read more! 🔗 https://lnkd.in/e73RfGqq
-
