Brian Raymond

Every document AI demo I’ve seen starts with a clean PDF. That’s a problem, because almost no one’s real data looks like that. The files that break production: the claim faxed twice, the form with notes in the margin, the table that spills onto page three. Those are the pages where an agent quietly reads the wrong column or guesses at a blank, and no one notices until a customer does. Our industry has gotten too comfortable judging parsers on documents picked to make them look good. So we’re offering a different test. Send us up to 25 of your toughest pages plus what your current parser returns. Within 24 hours, a named engineer sends back Transform’s output side by side. Free, and you keep it either way. unstructured.io/hot-mess-tes…

We’re excited to launch Unstructured Transform, our document parser built for the docs agents struggle with most. Frontier models are really good at reading a page now. But the documents our customers care about aren’t clean pages. They’re credit agreements with three levels of merged headers, insurance claims that have been scanned and re-scanned, lab panels, bills of materials, and 10-Ks with tables that spill across pages. And when an agent is the one reading them, small mistakes get expensive fast (one number lands in the wrong column and a payment goes out wrong, with nobody in the loop to catch it). So we tested Transform against GPT-6 Astra, Claude Opus 5.5 and Gemini 3.8 Flash on 224 pages of real business docs. Here’s how it stacks up: #1 on tables: 0.896 (next best: 0.845) Lowest hallucination rate: 5.0% (frontier models: 7.4 to 7.6%) $0.015 a page, about 1/7 the cost of GPT-6 Astra We’d love to see what you’re building with it! unstructured.io/blog/unstruc… #DocumentAI #AgenticAI #DocumentParsing #EnterpriseAI #Unstructured

The question I hear from engineering teams has changed, and I think it's the most important shift in the stack right now. In 2024 it was "why would I need the extra tenth of a percent?" Their parser did a pretty good job transforming a file into JSON or markdown, the retriever found the chunk, and the generated response looked fine. Good enough was good enough. In 2026 nobody asks me that. They ask whether we can tell them exactly where on the page a value came from and how sure we are, because an agent is going to act on it without a human in the loop. The math behind that is simple. With multi-step agentic systems, errors at the context generation phase (when raw data is converted into JSON or markdown) compound downstream. We ran an experiment to demonstrate this point: 200 scanned invoices through the same 25-step invoice-payment workflow, using three differing parsing engines. Field accuracy looked close. Whole invoices were not close at all. GPT 5.6 Sol only parsed 1% of invoices correctly, but most of these errors were invisible if you aren't an agent. I wrote a short ebook on what changed over the past couple of years and what bar you should hold for your preprocessing engine, ours included. RAG needed mediocre data. Agents need extraordinary data. unstructured.io/blog/rag-nee…

Field accuracy is the number everyone reports when they talk about parsing forms. It's also the one most likely to mislead you. We ran 40 grant applications through Unstructured Transform and GPT-5.6 Sol. On field accuracy they looked almost identical: 99.5% against 98.7%. You'd call it a tie. But field accuracy counts individual fields. A form can get 99% of its fields right and still have one wrong value sitting in it, which means the whole form still needs a human to catch it. So we measured what actually decides your human workload: how many whole forms came back with zero errors. Transform returned 34 of 40 clean while the frontier model got about half. The more accurate system turned out to be the cheaper one too, at roughly 7.5x lower cost per page. A wrong value that still looks plausible slips past review and ends up in a customer record or an automated decision, where it's far harder to catch and far more expensive to fix than it would have been on the page. Read more in this week's blog: unstructured.io/blog/how-to-…

Last week we wrote about why better frontier models haven't made document processing go away. So this week we tested it, on forms. If you're evaluating a parser, field accuracy is the first number most people look at, and it's also the most misleading. We ran 40 grant applications through Unstructured Transform and GPT-5.6 Sol. 99.5% vs 98.7%. Basically a tie. But count the forms where every value is right and it's 34 out of 40 for Unstructured vs about 20 for GPT-5.6 Sol. At 10,000 forms, that's roughly 15% for a person to check vs ~50% while costing 7.5x more per page. The mistakes are small, like a name with one letter wrong or a number in the wrong column. One wrong field means the whole form is wrong, but these are easy to miss because the answer still “looks” right. So when you test any vendor, score whole documents, not fields, and check what each one costs you per page. We wrote about this here: unstructured.io/blog/how-to-…

Greatest crossover in the history of AI Infra!!! @rak_garg

Today we’re announcing that Unstructured is now natively embedded inside Teradata Vantage. This isn’t an integration or a connector, our platform is running as a native service inside one of the most trusted enterprise data platforms in the world. Teradata customers run some of the most critical workloads — financial services, healthcare, defense, government. These are organizations where data sovereignty isn’t a nice-to-have, it’s a requirement. Now they can transform 70+ file types of unstructured content into AI-ready data, directly inside Vantage. No external pipelines. No additional infrastructure. We founded Unstructured with a simple conviction: the biggest unlock for enterprise AI is better data. This is what that looks like at scale.

2025 was our biggest year yet for enterprise deployments into customer VPCs. Every single one taught us something about what actually breaks in production. Here are five of the most impactful lessons our Field Engineering team learned this past year: 1. Environment volatility isn't an edge case—it's the default. Proxy servers silently intercept traffic. Custom certificate authorities break TLS handshakes. API calls vanish into black holes. We stopped treating these as surprises and built pre-deployment validation tools that catch them before the first container even spins up. 2. Configuration shouldn’t require a "tear down." In the early days, changing a single parameter meant rebuilding the stack. Working in diverse customer environments taught us that adaptability can’t be an afterthought. Now, we build runtime configurability into the platform from day one, allowing us to adjust critical settings on the fly without downtime. 3. Answers are needed in minutes, not hours. Debugging in a customer's VPC with limited access is a high-pressure exercise. To solve this, we built AI-powered log analyzers that surface root causes instantly. What used to take half a day of manual digging now takes minutes. 4. Documentation is its own product. Our runbooks aren't written in a vacuum; they are forged from support tickets and troubleshooting sessions. By sharing these "living documents" with customers, we empower their teams to handle routine issues independently, reserving our engineering syncs for genuinely new challenges. 5. On-prem isn't always the answer. This year, we saw a shift: many teams (especially those already utilizing Snowflake or Splunk) wanted enterprise-grade security without the overhead of managing the infrastructure. In response, we built Dedicated Instances (DI). DIs provide a fully managed, isolated environment with complete network separation, offering the security of on-prem with none of the environmental volatility. Deploying into customer environments is an exercise in humility. You can't anticipate every variable, but you can build systems that adapt to the reality of enterprise environments. Reflecting on the past three years, it’s clear why the demand for on prem deployment is skyrocketing. The GenAI ecosystem grew up around incredible open-source tools—Unstructured, Weaviate, Chroma, LlamaIndex, LangChain, Crew AI, etc—that are easy to run on prem. However, as builders graduate these prototypes into production, they hit the "Enterprise Wall": friction with internal infra teams, security audits, and complex networking. At Unstructured, we’ve realized that success doesn’t come from a great product alone; it comes from being the partner that helps builders navigate those wickets and move to production at speed.

Couldn't be more excited to share v1 of our industry benchmarks. New models and preprocessing strategies are announced almost every week, and up until today, it's been almost impossible to draw conclusions with any confidence. What we found: - Different model pipelines excel at different tasks: Some configurations optimize for table extraction, others for content fidelity, and others for structural understanding - Trade-offs exist: A pipeline with the highest table accuracy might have slightly different hallucination characteristics than one optimized for element alignment - Consistent leadership: Across the diversity of configurations, Unstructured pipelines consistently outperform competitors on key metrics - Informed choice: You can select the pipeline that best matches your specific document types and downstream requirements Early next month we will be open sourcing our labeled dataset so that everyone can independently evaluate the performance of their preprocessing pipelines. No cherry picking, overfitting to specific docs, or obfuscation. With Unstructured you can trust us for maximum transparency for which models we're using, how we're using them, and how they're performing.

Incredibly excited for all the new goodness we'll be rolling out this month. Starting things off, we've decided to massively simplify the user experience. No feature gating, multiple pricing tiers, etc. If you sign up through our website you can expect: 1) Access to every feature, connector, model, etc. We're not holding anything back. 2) No complexity around pricing. Go nuts and run wild DAGs with multiple VLM/LLMs. Create custom schemas, custom metadata, etc. One flat cost of $.03/page. Try doing anything similar with hyperscalers and you're looking at >$.10/page. That's with janky connectors. Trying doing the same thing with an IDP startup and you won't be able to do...they don't provide GenAI-ready data 3) We continuously test every connector, every API, every pipeline and deliver world class SLAs and uptime numbers. When we say we're "enterprise ready," it's not just a bumper sticker. We're committed to building tools that developers love. If you have any feedback at all, don't hesitate to drop me a note [email protected]. I'd love to hear from you.

Excited to announce our OEM partnership with @IBM. Together, we’re helping enterprises move from messy, unstructured content to trusted, AI-ready data at scale. watsonx.data + Unstructured = reliable AI. Special thanks to @ritika_gunnar, Ed Calvesbert, Emily Fontaine, and Stevan Slusher or helping to make this a reality!! You can read more here: ibm.com/new/announcements/ib…

DIY data processing solutions work — until their maintenance costs eclipse their value. We’re proud that thousands of teams run production systems on Unstructured OSS. That’s why we open-sourced it in the first place. Over the years, we’ve seen a clear pattern emerge. Teams start strong with DIY pipelines built on OSS: prototypes succeed, workflows get automated, and RAG pipelines begin powering real applications. But as usage grows — more documents, more complexity, more enterprise requirements — maintaining these systems quietly become many full-time jobs. Teams find themselves spending cycles on orchestration, retries, scaling, and troubleshooting, rather than building the features that differentiate their product. From helping hundreds of customers navigate this transition, one lesson stands out: knowing when to consider managed infrastructure isn’t about whether OSS works — it’s about recognizing when your engineering effort could be better spent elsewhere. Managed infrastructure allows teams to focus on delivering value while still benefiting from the base capabilities OSS provides, with the added advantage of continuous improvements, advanced parsing, extensive metadata enrichment features, and more, all handled by a team of experts. In our upcoming webinar, we’ll share these signals and lessons in more detail, helping you decide when it’s time to move from OSS to the Unstructured ETL+ platform: 📷 When to Move from Open Source to the Unstructured ETL+ Platform unstructured.io/events/when-…