[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12680,
    "text": ">> I think we can get started. Hey"
  },
  {
    "startMs": 13760,
    "text": "everyone, I'm Jerry, co-founder and CEO"
  },
  {
    "startMs": 15520,
    "text": "of LlamaIndex, and today I'm excited to"
  },
  {
    "startMs": 18080,
    "text": "uh give a talk called building the"
  },
  {
    "startMs": 19440,
    "text": "document context layer for AI agents. Um"
  },
  {
    "startMs": 22320,
    "text": "really uh big shoutout to the AI"
  },
  {
    "startMs": 24400,
    "text": "Engineer World Fair for hosting. Um and"
  },
  {
    "startMs": 26560,
    "text": "if you've seen some of my earlier talks"
  },
  {
    "startMs": 28160,
    "text": "from the previous uh AI Engineer"
  },
  {
    "startMs": 30000,
    "text": "conferences, uh we've kind of traced"
  },
  {
    "startMs": 31640,
    "text": "through a lot of the evolution of how,"
  },
  {
    "startMs": 33280,
    "text": "you know, um agent advances have"
  },
  {
    "startMs": 35000,
    "text": "correlated with uh you know, how you"
  },
  {
    "startMs": 36640,
    "text": "inject context into evolving"
  },
  {
    "startMs": 38400,
    "text": "applications. Um and so today, in 2026,"
  },
  {
    "startMs": 40960,
    "text": "we'll kind of talk about three main"
  },
  {
    "startMs": 42320,
    "text": "topics. One is, you know, what is RAG in"
  },
  {
    "startMs": 44920,
    "text": "2026? Um two, uh basically that"
  },
  {
    "startMs": 47840,
    "text": "basically decomposes into an agent"
  },
  {
    "startMs": 49480,
    "text": "harness plus a context layer. So kind of"
  },
  {
    "startMs": 51480,
    "text": "going into the modern document context"
  },
  {
    "startMs": 53040,
    "text": "layer for agents, um and giving you a"
  },
  {
    "startMs": 54840,
    "text": "little bit of sense of kind of some of"
  },
  {
    "startMs": 56200,
    "text": "the core capabilities today, um"
  },
  {
    "startMs": 57760,
    "text": "especially uh unlocking, you know, the"
  },
  {
    "startMs": 59640,
    "text": "vast trove of document-based data today,"
  },
  {
    "startMs": 61720,
    "text": "and giving you a glimpse of what's next."
  },
  {
    "startMs": 64160,
    "text": "All right. RAG in 2026. But before that,"
  },
  {
    "startMs": 66640,
    "text": "just a little bit about the company, and"
  },
  {
    "startMs": 67800,
    "text": "I'll kind of get on with the main talk."
  },
  {
    "startMs": 69480,
    "text": "Um you might have seen us as a RAG"
  },
  {
    "startMs": 70760,
    "text": "framework. We started in 2023, got"
  },
  {
    "startMs": 72560,
    "text": "pretty popular, um kind of created a lot"
  },
  {
    "startMs": 74680,
    "text": "of techniques around like advanced RAG,"
  },
  {
    "startMs": 76200,
    "text": "that type of stuff. Um today, we're"
  },
  {
    "startMs": 78000,
    "text": "basically the main document"
  },
  {
    "startMs": 79120,
    "text": "infrastructure for AI agents. Um we"
  },
  {
    "startMs": 81280,
    "text": "deliver the best platform for agents to"
  },
  {
    "startMs": 83000,
    "text": "actually read and operate over"
  },
  {
    "startMs": 84400,
    "text": "documents, and we see ourselves as"
  },
  {
    "startMs": 86120,
    "text": "unlocking basically the vast troves of"
  },
  {
    "startMs": 87920,
    "text": "document context even as agents and"
  },
  {
    "startMs": 89840,
    "text": "models themselves get better."
  },
  {
    "startMs": 93000,
    "text": "All right. So, what does RAG actually"
  },
  {
    "startMs": 95240,
    "text": "mean in 2026, and how are agents"
  },
  {
    "startMs": 97680,
    "text": "actually inhaling and addressing context"
  },
  {
    "startMs": 99480,
    "text": "today?"
  },
  {
    "startMs": 100680,
    "text": "This is a snapshot from um actually 3"
  },
  {
    "startMs": 103160,
    "text": "years ago, um when I first gave a talk"
  },
  {
    "startMs": 105040,
    "text": "on this, and naive RAG was basically"
  },
  {
    "startMs": 106800,
    "text": "just building a simple chatbot over your"
  },
  {
    "startMs": 108520,
    "text": "private corpus of data. If you flash"
  },
  {
    "startMs": 110360,
    "text": "back all the way to January of like"
  },
  {
    "startMs": 111840,
    "text": "2023, um you know, this technique"
  },
  {
    "startMs": 114120,
    "text": "basically consisted of you have some"
  },
  {
    "startMs": 116160,
    "text": "corpus of documents, um you chunk it up,"
  },
  {
    "startMs": 118480,
    "text": "embed it, and put it into a vector"
  },
  {
    "startMs": 120040,
    "text": "database. You do some naive top K"
  },
  {
    "startMs": 122360,
    "text": "retrieval, and then you generate some"
  },
  {
    "startMs": 123840,
    "text": "stuff with an LLM. All the steps are"
  },
  {
    "startMs": 125840,
    "text": "fixed. You use kind of like a fixed set"
  },
  {
    "startMs": 127600,
    "text": "of techniques, and then with that you"
  },
  {
    "startMs": 129119,
    "text": "actually get some basic application"
  },
  {
    "startMs": 130880,
    "text": "results by just being able to chat over"
  },
  {
    "startMs": 132680,
    "text": "a corpus of documents."
  },
  {
    "startMs": 134560,
    "text": "Of course, the agent landscape has"
  },
  {
    "startMs": 136480,
    "text": "changed quite a bit and quite"
  },
  {
    "startMs": 138080,
    "text": "dramatically since then."
  },
  {
    "startMs": 139760,
    "text": "One is agent loops and tool use have"
  },
  {
    "startMs": 142200,
    "text": "gotten a lot a lot better, especially as"
  },
  {
    "startMs": 144400,
    "text": "both the models and agent harnesses have"
  },
  {
    "startMs": 146360,
    "text": "increased. And there's basically a"
  },
  {
    "startMs": 148200,
    "text": "cleaner separation now between agent"
  },
  {
    "startMs": 150120,
    "text": "reasoning and how they actually interact"
  },
  {
    "startMs": 151760,
    "text": "with context. If you look at what I call"
  },
  {
    "startMs": 154080,
    "text": "kind of the modern generalized agents,"
  },
  {
    "startMs": 156080,
    "text": "which includes, you know, all your"
  },
  {
    "startMs": 157280,
    "text": "favorite applications and tools out"
  },
  {
    "startMs": 158720,
    "text": "there from Claude Code, Claude Code"
  },
  {
    "startMs": 160160,
    "text": "Work, Open Claw, Codex, and a few"
  },
  {
    "startMs": 162520,
    "text": "others,"
  },
  {
    "startMs": 163720,
    "text": "you know, the retrieval complexity has"
  },
  {
    "startMs": 165480,
    "text": "started to get baked into the agent"
  },
  {
    "startMs": 167400,
    "text": "layer. So, instead of coming up with a"
  },
  {
    "startMs": 169280,
    "text": "variety of hacks to really work around"
  },
  {
    "startMs": 171160,
    "text": "the limitations of naive top K"
  },
  {
    "startMs": 172560,
    "text": "retrieval, you can start getting the"
  },
  {
    "startMs": 174760,
    "text": "agent to really reason about, for"
  },
  {
    "startMs": 176240,
    "text": "instance, the best key best keyword to"
  },
  {
    "startMs": 177800,
    "text": "search for, best like, you know, search"
  },
  {
    "startMs": 179720,
    "text": "term to actually get back good results."
  },
  {
    "startMs": 181800,
    "text": "Even if the retrieval tools are basic,"
  },
  {
    "startMs": 183840,
    "text": "the agent can input the right queries to"
  },
  {
    "startMs": 185720,
    "text": "basically loop upon itself and help"
  },
  {
    "startMs": 187920,
    "text": "solve the task at hand."
  },
  {
    "startMs": 189560,
    "text": "Number two, context is moving up the"
  },
  {
    "startMs": 191560,
    "text": "stack. There's been a lot of"
  },
  {
    "startMs": 193080,
    "text": "conversations, I think, for the first"
  },
  {
    "startMs": 194959,
    "text": "two and a half years of Gen AI of how do"
  },
  {
    "startMs": 197080,
    "text": "you actually manage the agent context"
  },
  {
    "startMs": 198920,
    "text": "window, make sure it doesn't overflow,"
  },
  {
    "startMs": 200519,
    "text": "that type of thing. But, I think as, you"
  },
  {
    "startMs": 202880,
    "text": "know, a lot of the evolving techniques,"
  },
  {
    "startMs": 204959,
    "text": "compaction, long contexts have evolved,"
  },
  {
    "startMs": 207800,
    "text": "I think more and more of the"
  },
  {
    "startMs": 208880,
    "text": "conversation is actually how do you just"
  },
  {
    "startMs": 210519,
    "text": "hook up the right MCP servers and skills"
  },
  {
    "startMs": 212920,
    "text": "and tasks to the agent to enable it to"
  },
  {
    "startMs": 215120,
    "text": "do various types of tasks."
  },
  {
    "startMs": 217040,
    "text": "And so, that also applies to agent"
  },
  {
    "startMs": 218840,
    "text": "orchestration, right? Coding agents have"
  },
  {
    "startMs": 220600,
    "text": "gotten better."
  },
  {
    "startMs": 222239,
    "text": "Abstractions of actually defining what"
  },
  {
    "startMs": 224160,
    "text": "types of tasks and programs you want to"
  },
  {
    "startMs": 225880,
    "text": "build have moved a little bit upwards"
  },
  {
    "startMs": 227519,
    "text": "towards English as opposed to through"
  },
  {
    "startMs": 229120,
    "text": "code. And so, maybe in 2023 to 2025, you"
  },
  {
    "startMs": 232920,
    "text": "define programs and you still build"
  },
  {
    "startMs": 234440,
    "text": "stuff via like importing Python, using"
  },
  {
    "startMs": 236320,
    "text": "TypeScript through code. It's pretty"
  },
  {
    "startMs": 238280,
    "text": "clear these days more and more people"
  },
  {
    "startMs": 240120,
    "text": "are just building stuff using English,"
  },
  {
    "startMs": 242600,
    "text": "whether you're a software engineer or"
  },
  {
    "startMs": 244720,
    "text": "you're a non-technical function like"
  },
  {
    "startMs": 246760,
    "text": "go-to-market marketing. And you're"
  },
  {
    "startMs": 248560,
    "text": "defining runbooks through English and"
  },
  {
    "startMs": 250320,
    "text": "kind of like defining the right goals"
  },
  {
    "startMs": 252120,
    "text": "and making sure the AI is aligned on the"
  },
  {
    "startMs": 254320,
    "text": "right task. And we see that carrying"
  },
  {
    "startMs": 256160,
    "text": "over going forward as well."
  },
  {
    "startMs": 259359,
    "text": "So, in terms of just like general"
  },
  {
    "startMs": 261079,
    "text": "evolving knowledge work patterns, and"
  },
  {
    "startMs": 262640,
    "text": "this kind of forms a foundation for what"
  },
  {
    "startMs": 264520,
    "text": "we how we think about the context layer."
  },
  {
    "startMs": 266520,
    "text": "You know,"
  },
  {
    "startMs": 267480,
    "text": "in the past maybe you could use AI to"
  },
  {
    "startMs": 269919,
    "text": "just like ask simple questions, get it"
  },
  {
    "startMs": 271480,
    "text": "resolved."
  },
  {
    "startMs": 272640,
    "text": "Today, you're starting to be able to"
  },
  {
    "startMs": 274000,
    "text": "define and solve more end-to-end like"
  },
  {
    "startMs": 276480,
    "text": "tasks just through English and also use"
  },
  {
    "startMs": 279080,
    "text": "English to start"
  },
  {
    "startMs": 280440,
    "text": "compiling repeatable programs to"
  },
  {
    "startMs": 283000,
    "text": "actually execute something at scale."
  },
  {
    "startMs": 285440,
    "text": "Obviously, there's been a lot of"
  },
  {
    "startMs": 286960,
    "text": "discussion on how this will evolve even"
  },
  {
    "startMs": 288760,
    "text": "more, and we do see AI agents as"
  },
  {
    "startMs": 290919,
    "text": "behaving even more autonomously through"
  },
  {
    "startMs": 293120,
    "text": "solving long horizon tasks, looping upon"
  },
  {
    "startMs": 295640,
    "text": "themselves to be able to achieve a goal."
  },
  {
    "startMs": 297400,
    "text": "So, the future will move towards a state"
  },
  {
    "startMs": 299080,
    "text": "where you actually might not even have"
  },
  {
    "startMs": 300960,
    "text": "to define the task in English, but"
  },
  {
    "startMs": 302760,
    "text": "actually more of the goal and a scoring"
  },
  {
    "startMs": 304360,
    "text": "rubric, and then the agent will use all"
  },
  {
    "startMs": 306480,
    "text": "available context available to it to"
  },
  {
    "startMs": 308240,
    "text": "actually solve the task."
  },
  {
    "startMs": 310840,
    "text": "So, with that overall framing, let's"
  },
  {
    "startMs": 313240,
    "text": "talk about what kind of what is a pretty"
  },
  {
    "startMs": 315280,
    "text": "core focus for us as a company, which is"
  },
  {
    "startMs": 317400,
    "text": "basically helping to unlock unstructured"
  },
  {
    "startMs": 319560,
    "text": "context to basically feed into these"
  },
  {
    "startMs": 321360,
    "text": "agents even as they get really, really"
  },
  {
    "startMs": 323120,
    "text": "good both in terms of the core model"
  },
  {
    "startMs": 324960,
    "text": "capabilities as well as the agent"
  },
  {
    "startMs": 326600,
    "text": "harness. And we'll focus a lot on our"
  },
  {
    "startMs": 328960,
    "text": "core focus area, which is document"
  },
  {
    "startMs": 331000,
    "text": "context, but of course there's plenty of"
  },
  {
    "startMs": 332480,
    "text": "other sources of context as well. In"
  },
  {
    "startMs": 334680,
    "text": "fact, we do think context really is"
  },
  {
    "startMs": 337360,
    "text": "everything."
  },
  {
    "startMs": 338480,
    "text": "In the end, you could have a infinitely"
  },
  {
    "startMs": 340480,
    "text": "smart agents, but your ability to"
  },
  {
    "startMs": 342800,
    "text": "actually get value out of this"
  },
  {
    "startMs": 344840,
    "text": "infinitely AI agent um, able to do more"
  },
  {
    "startMs": 347480,
    "text": "and more stuff end to end, is actually,"
  },
  {
    "startMs": 349720,
    "text": "you know, giving it the right things to"
  },
  {
    "startMs": 351000,
    "text": "do. Um, and the right things to do"
  },
  {
    "startMs": 352560,
    "text": "include the actual task, uh, the goal,"
  },
  {
    "startMs": 355640,
    "text": "um, but also access to the vast troves"
  },
  {
    "startMs": 358120,
    "text": "of organizational context available to"
  },
  {
    "startMs": 360320,
    "text": "it. Whether that is web search, web"
  },
  {
    "startMs": 362720,
    "text": "context, whether that is connectors"
  },
  {
    "startMs": 365040,
    "text": "through, you know, like tools and skills"
  },
  {
    "startMs": 366840,
    "text": "and MCP servers, whether it's connectors"
  },
  {
    "startMs": 369160,
    "text": "through Snowflake or Databricks"
  },
  {
    "startMs": 370360,
    "text": "warehouses, um, or for us, whether it's"
  },
  {
    "startMs": 372840,
    "text": "access to the vast trove, like the 90%"
  },
  {
    "startMs": 375320,
    "text": "of documents that are stored within"
  },
  {
    "startMs": 376800,
    "text": "SharePoint, Box, Dropbox, S3, um, you"
  },
  {
    "startMs": 380000,
    "text": "know, all the unstructured, uh, data"
  },
  {
    "startMs": 381919,
    "text": "that's locked up within document"
  },
  {
    "startMs": 383120,
    "text": "containers. Um, and so for us, you know,"
  },
  {
    "startMs": 385919,
    "text": "we care a lot about how do you actually"
  },
  {
    "startMs": 387680,
    "text": "build the right tools and unlock the"
  },
  {
    "startMs": 389360,
    "text": "right context here."
  },
  {
    "startMs": 391720,
    "text": "We see documents as universal containers"
  },
  {
    "startMs": 394080,
    "text": "for unstructured context. Uh, there's"
  },
  {
    "startMs": 395919,
    "text": "over 10 trillion plus pages of human"
  },
  {
    "startMs": 397720,
    "text": "native knowledge locked up within PDFs,"
  },
  {
    "startMs": 399640,
    "text": "PowerPoints, Word documents, and Excel"
  },
  {
    "startMs": 401320,
    "text": "sheets. And at the same time, agents are"
  },
  {
    "startMs": 403840,
    "text": "also starting to generate exponentially"
  },
  {
    "startMs": 405560,
    "text": "more data in terms of, you know, more"
  },
  {
    "startMs": 407120,
    "text": "agent native formats in terms of"
  },
  {
    "startMs": 408560,
    "text": "markdown and HTML."
  },
  {
    "startMs": 412320,
    "text": "For us, building an agent native"
  },
  {
    "startMs": 414120,
    "text": "document platform contains three main"
  },
  {
    "startMs": 416480,
    "text": "pieces. Uh, it contains the document"
  },
  {
    "startMs": 418640,
    "text": "parsing layer, actually digitalizing,"
  },
  {
    "startMs": 421240,
    "text": "um, the vast troves of PDFs,"
  },
  {
    "startMs": 422800,
    "text": "PowerPoints, and Word documents into"
  },
  {
    "startMs": 425120,
    "text": "kind of accurate token efficient"
  },
  {
    "startMs": 427040,
    "text": "context, um, and feeding it, you know,"
  },
  {
    "startMs": 429480,
    "text": "either markdown, metadata, and other"
  },
  {
    "startMs": 431400,
    "text": "forms that actually enable agents to do"
  },
  {
    "startMs": 433240,
    "text": "stuff over your documents."
  },
  {
    "startMs": 435160,
    "text": "The next is actually what we call the"
  },
  {
    "startMs": 436919,
    "text": "semantic and storage layer, um, which is"
  },
  {
    "startMs": 439160,
    "text": "kind of like this concept of document"
  },
  {
    "startMs": 441280,
    "text": "management for humans and agents."
  },
  {
    "startMs": 442880,
    "text": "Instead of humans opening up Microsoft"
  },
  {
    "startMs": 444600,
    "text": "Word, using your file system or file"
  },
  {
    "startMs": 446440,
    "text": "storage, how do you create some sort of"
  },
  {
    "startMs": 448400,
    "text": "agent interface that actually captures,"
  },
  {
    "startMs": 450400,
    "text": "stores all your documents, and enables"
  },
  {
    "startMs": 452600,
    "text": "agents to actually manage and operate"
  },
  {
    "startMs": 454400,
    "text": "over them in various ways. Um, and then"
  },
  {
    "startMs": 457040,
    "text": "there's still a room for actually kind"
  },
  {
    "startMs": 459080,
    "text": "of repeatable document workflows is"
  },
  {
    "startMs": 461040,
    "text": "agent layer they can actually encode"
  },
  {
    "startMs": 462880,
    "text": "within the"
  },
  {
    "startMs": 464040,
    "text": "uh"
  },
  {
    "startMs": 464800,
    "text": "within like this type of platform, too."
  },
  {
    "startMs": 466840,
    "text": "If there is a repeatable workflow, like"
  },
  {
    "startMs": 468800,
    "text": "invoice processing, KYC, or claims,"
  },
  {
    "startMs": 471320,
    "text": "instead of always offloading it to a"
  },
  {
    "startMs": 473040,
    "text": "generalized agent, how do you actually"
  },
  {
    "startMs": 474880,
    "text": "develop some sort of specialized"
  },
  {
    "startMs": 476280,
    "text": "workflow for it and really carefully"
  },
  {
    "startMs": 478320,
    "text": "tune cost and accuracy to make sure that"
  },
  {
    "startMs": 480240,
    "text": "you achieve your goal."
  },
  {
    "startMs": 483520,
    "text": "So, when we talk about like the concepts"
  },
  {
    "startMs": 485520,
    "text": "today in in terms of these three layers,"
  },
  {
    "startMs": 487960,
    "text": "uh we'll talk about this concept of"
  },
  {
    "startMs": 489280,
    "text": "document OCR, which is one of the core"
  },
  {
    "startMs": 491000,
    "text": "concepts of this track, um to document"
  },
  {
    "startMs": 493320,
    "text": "extraction, document search, and"
  },
  {
    "startMs": 494919,
    "text": "document workflows. There's still a"
  },
  {
    "startMs": 496720,
    "text": "variety of topics yet left unsolved from"
  },
  {
    "startMs": 498919,
    "text": "agent native document formats, document"
  },
  {
    "startMs": 500880,
    "text": "versioning, document editing, you know,"
  },
  {
    "startMs": 502880,
    "text": "hill climbing as a service, and just"
  },
  {
    "startMs": 504680,
    "text": "like a wide set of uh kind of remaining"
  },
  {
    "startMs": 507000,
    "text": "concepts to actually create a"
  },
  {
    "startMs": 508800,
    "text": "comprehensive piece of software where"
  },
  {
    "startMs": 510560,
    "text": "agents and humans can collaborate on"
  },
  {
    "startMs": 512360,
    "text": "documents and actually do work over"
  },
  {
    "startMs": 513919,
    "text": "them."
  },
  {
    "startMs": 515960,
    "text": "So, we'll start with the first section"
  },
  {
    "startMs": 517560,
    "text": "um and maybe start kind of with uh the"
  },
  {
    "startMs": 519520,
    "text": "document OCR RP."
  },
  {
    "startMs": 521440,
    "text": "Um first off, uh you know, document OCR"
  },
  {
    "startMs": 524280,
    "text": "is hard. Um some of you might have seen"
  },
  {
    "startMs": 526360,
    "text": "a blog post that we put out a few months"
  },
  {
    "startMs": 528040,
    "text": "ago related to this topic. But, the"
  },
  {
    "startMs": 530160,
    "text": "reason this problem even exists is an"
  },
  {
    "startMs": 532440,
    "text": "agent cannot actually take the like the"
  },
  {
    "startMs": 534800,
    "text": "raw PDF like file binary and make sense"
  },
  {
    "startMs": 537440,
    "text": "of it. Um that's because the like the"
  },
  {
    "startMs": 539800,
    "text": "way PDFs are actually stored as a format"
  },
  {
    "startMs": 542320,
    "text": "is it's rendered for kind of like"
  },
  {
    "startMs": 543760,
    "text": "display purposes um and not really for"
  },
  {
    "startMs": 546360,
    "text": "uh you know, machine consumption in an"
  },
  {
    "startMs": 548200,
    "text": "interpretable manner."
  },
  {
    "startMs": 549800,
    "text": "Um"
  },
  {
    "startMs": 550960,
    "text": "they're designed for printing. So,"
  },
  {
    "startMs": 552320,
    "text": "basically text are represented as almost"
  },
  {
    "startMs": 554600,
    "text": "like individual glyphs with coordinates."
  },
  {
    "startMs": 556920,
    "text": "Um tables are not represented as tables,"
  },
  {
    "startMs": 559520,
    "text": "they're represented as like typically"
  },
  {
    "startMs": 561360,
    "text": "line segments uh drawn with various"
  },
  {
    "startMs": 563360,
    "text": "types of borders, and also text drawn at"
  },
  {
    "startMs": 565640,
    "text": "certain cell positions. And so, if"
  },
  {
    "startMs": 568320,
    "text": "you're kind of an agent that's trying to"
  },
  {
    "startMs": 569800,
    "text": "make sense of this like document, you're"
  },
  {
    "startMs": 571680,
    "text": "going to have a really hard time"
  },
  {
    "startMs": 573000,
    "text": "actually trying to reason about what"
  },
  {
    "startMs": 574760,
    "text": "character and what shapes like map to"
  },
  {
    "startMs": 576680,
    "text": "what. Um so the whole point of document"
  },
  {
    "startMs": 578880,
    "text": "OCR, um, you know, it's been around for"
  },
  {
    "startMs": 580600,
    "text": "like 20 plus years, is to really try to"
  },
  {
    "startMs": 583040,
    "text": "create some sort of, uh, digitalized"
  },
  {
    "startMs": 585200,
    "text": "well-interpretable, uh, representation"
  },
  {
    "startMs": 587560,
    "text": "um, that's both interpretable to humans,"
  },
  {
    "startMs": 589360,
    "text": "um, as well as AI agents."
  },
  {
    "startMs": 591360,
    "text": "Also, reading order itself, if you have"
  },
  {
    "startMs": 592880,
    "text": "a multi-column layout, there is no, uh,"
  },
  {
    "startMs": 595360,
    "text": "guarantee that the way it's represented"
  },
  {
    "startMs": 597080,
    "text": "in the PDF, um, actually corresponds to"
  },
  {
    "startMs": 599640,
    "text": "the typical ways that humans would read"
  },
  {
    "startMs": 601280,
    "text": "it. Cuz again, it's basically just an"
  },
  {
    "startMs": 602880,
    "text": "arbitrary sequence of characters drawn"
  },
  {
    "startMs": 604680,
    "text": "with uh, coordinate positions."
  },
  {
    "startMs": 607880,
    "text": "Related to this, you know, even Word"
  },
  {
    "startMs": 609480,
    "text": "doc, uh, parsing is hard. Um, they're a"
  },
  {
    "startMs": 611360,
    "text": "little bit more structured than PDFs,"
  },
  {
    "startMs": 612960,
    "text": "but they're in kind of like this, uh,"
  },
  {
    "startMs": 614640,
    "text": "custom bespoke XML format, and this"
  },
  {
    "startMs": 616560,
    "text": "applies to PowerPoints as well. Um, it"
  },
  {
    "startMs": 619120,
    "text": "contains more structural information,"
  },
  {
    "startMs": 620920,
    "text": "but still there's a ton of like fluff."
  },
  {
    "startMs": 622800,
    "text": "Like you don't actually need to ingest"
  },
  {
    "startMs": 624360,
    "text": "all the tags to actually have the agent"
  },
  {
    "startMs": 626400,
    "text": "make sense of the Word document. It"
  },
  {
    "startMs": 628040,
    "text": "still needs to infer a bunch of"
  },
  {
    "startMs": 629720,
    "text": "structure from it. It needs like the"
  },
  {
    "startMs": 631640,
    "text": "right, uh, kind of to lift the right"
  },
  {
    "startMs": 633680,
    "text": "metadata around like formatting,"
  },
  {
    "startMs": 635280,
    "text": "semantics, that type of stuff. But also"
  },
  {
    "startMs": 637520,
    "text": "being able to ignore the tags and"
  },
  {
    "startMs": 638960,
    "text": "actually be able to render the document"
  },
  {
    "startMs": 640480,
    "text": "so that the agent can see the overall"
  },
  {
    "startMs": 642000,
    "text": "structure of the page. Um, it's still"
  },
  {
    "startMs": 644200,
    "text": "generally hard problem, and there's a"
  },
  {
    "startMs": 645880,
    "text": "lot of uplift you can get by actually"
  },
  {
    "startMs": 647640,
    "text": "parsing it into a more interpretable"
  },
  {
    "startMs": 649240,
    "text": "format instead of just using the native,"
  },
  {
    "startMs": 651360,
    "text": "uh, OXML and feeding that to an agent."
  },
  {
    "startMs": 655240,
    "text": "Um, if you're familiar with document"
  },
  {
    "startMs": 656920,
    "text": "understanding, um, you know, it's been"
  },
  {
    "startMs": 658720,
    "text": "around for quite a bit of time. Um,"
  },
  {
    "startMs": 660760,
    "text": "there's a lot of these like heuristic"
  },
  {
    "startMs": 662240,
    "text": "and pipeline-based approaches, which,"
  },
  {
    "startMs": 664240,
    "text": "uh, focused on kind of more, uh, I guess"
  },
  {
    "startMs": 666640,
    "text": "like human-driven, hand-handwritten like"
  },
  {
    "startMs": 668760,
    "text": "techniques to analyze like kind of"
  },
  {
    "startMs": 670760,
    "text": "various pieces of text, group them into"
  },
  {
    "startMs": 672680,
    "text": "clusters, and identify tables,"
  },
  {
    "startMs": 674640,
    "text": "paragraphs, and be able to kind of like,"
  },
  {
    "startMs": 676800,
    "text": "uh, generate some sort of output"
  },
  {
    "startMs": 678360,
    "text": "representation. Um, a lot of basic"
  },
  {
    "startMs": 680360,
    "text": "techniques if you use open-source"
  },
  {
    "startMs": 681880,
    "text": "libraries like PyPDF, PyMuPDF, um, and"
  },
  {
    "startMs": 684600,
    "text": "of course like some of the,"
  },
  {
    "startMs": 686000,
    "text": "uh, more recent approaches, um,"
  },
  {
    "startMs": 687800,
    "text": "basically use this type of approach."
  },
  {
    "startMs": 689760,
    "text": "There's also, of course, using a VLM to"
  },
  {
    "startMs": 691960,
    "text": "one-shot a document um, into text, uh,"
  },
  {
    "startMs": 694840,
    "text": "that's what we call like a vision-based"
  },
  {
    "startMs": 696240,
    "text": "approach."
  },
  {
    "startMs": 697440,
    "text": "This works decently well. Obviously, you"
  },
  {
    "startMs": 699760,
    "text": "know, I think there's kind of a lot of"
  },
  {
    "startMs": 701000,
    "text": "uplift you can get by being able to read"
  },
  {
    "startMs": 702560,
    "text": "the visual structure, but there's a lot"
  },
  {
    "startMs": 705360,
    "text": "of sub-optimal pieces about it. It can"
  },
  {
    "startMs": 708080,
    "text": "hallucinate on text-only pages. It costs"
  },
  {
    "startMs": 710800,
    "text": "a ton of money,"
  },
  {
    "startMs": 712080,
    "text": "and also it still lacks a lot of the"
  },
  {
    "startMs": 713840,
    "text": "semantics and grounding that you"
  },
  {
    "startMs": 715400,
    "text": "typically expect with some sort of"
  },
  {
    "startMs": 716920,
    "text": "document processing tool."
  },
  {
    "startMs": 718720,
    "text": "And so for us, we really think about"
  },
  {
    "startMs": 720640,
    "text": "combining both the pipeline-based"
  },
  {
    "startMs": 722760,
    "text": "approaches of deeply understanding the"
  },
  {
    "startMs": 724560,
    "text": "file containers and binaries with the"
  },
  {
    "startMs": 726880,
    "text": "vision-based approaches to help"
  },
  {
    "startMs": 728200,
    "text": "generate, you know, kind of"
  },
  {
    "startMs": 730120,
    "text": "a hybrid approach that we think is at"
  },
  {
    "startMs": 731760,
    "text": "the Pareto frontier of cost and"
  },
  {
    "startMs": 732960,
    "text": "accuracy."
  },
  {
    "startMs": 735920,
    "text": "A quick note on this, I'll probably just"
  },
  {
    "startMs": 738560,
    "text": "like skim the high level, but the Pareto"
  },
  {
    "startMs": 740680,
    "text": "frontier for document OCR"
  },
  {
    "startMs": 743000,
    "text": "will always be like much more accurate"
  },
  {
    "startMs": 744800,
    "text": "and cheap compared to the Pareto"
  },
  {
    "startMs": 746160,
    "text": "frontier for, you know, wherever the"
  },
  {
    "startMs": 747960,
    "text": "frontier models are in terms of document"
  },
  {
    "startMs": 749480,
    "text": "understanding. It's because it's a very"
  },
  {
    "startMs": 751440,
    "text": "specific data type, and there's always"
  },
  {
    "startMs": 753880,
    "text": "ways to kind of like distill the latest"
  },
  {
    "startMs": 755560,
    "text": "visual understanding capabilities from,"
  },
  {
    "startMs": 757240,
    "text": "you know, Gemini, GPT,"
  },
  {
    "startMs": 759200,
    "text": "Opus into kind of a carefully tailored"
  },
  {
    "startMs": 762360,
    "text": "workflow that's able to process your"
  },
  {
    "startMs": 763839,
    "text": "documents at scale and in a highly"
  },
  {
    "startMs": 765800,
    "text": "accurate manner."
  },
  {
    "startMs": 767080,
    "text": "To some extent, that's exactly what we"
  },
  {
    "startMs": 768720,
    "text": "do. You know, we both optimize"
  },
  {
    "startMs": 770800,
    "text": "underlying like PDF engines plus like"
  },
  {
    "startMs": 773160,
    "text": "Word, PowerPoint, and others."
  },
  {
    "startMs": 775000,
    "text": "We have an agentic harness that's like"
  },
  {
    "startMs": 776760,
    "text": "carefully tuned for auto routing between"
  },
  {
    "startMs": 779160,
    "text": "cheaper specialized models to frontier"
  },
  {
    "startMs": 781000,
    "text": "models,"
  },
  {
    "startMs": 782080,
    "text": "and also kind of specialized fine-tuned"
  },
  {
    "startMs": 784200,
    "text": "document VLMs that are parameter"
  },
  {
    "startMs": 786360,
    "text": "efficient and focus on specific classes"
  },
  {
    "startMs": 788400,
    "text": "of documents,"
  },
  {
    "startMs": 789560,
    "text": "elements like tables, charts, and"
  },
  {
    "startMs": 791200,
    "text": "others."
  },
  {
    "startMs": 794680,
    "text": "That's our commercial service called"
  },
  {
    "startMs": 796480,
    "text": "LlamaParse, which I'll talk about in a"
  },
  {
    "startMs": 798200,
    "text": "bit. But I think in general, you know,"
  },
  {
    "startMs": 800880,
    "text": "we're extremely committed to advancing"
  },
  {
    "startMs": 802800,
    "text": "the frontier of just document"
  },
  {
    "startMs": 804120,
    "text": "understanding cuz basically if you're"
  },
  {
    "startMs": 805960,
    "text": "within an enterprise organization and"
  },
  {
    "startMs": 807720,
    "text": "you have a massive long tail of"
  },
  {
    "startMs": 809040,
    "text": "documents across financial services,"
  },
  {
    "startMs": 811400,
    "text": "insurance, manufacturing, legal,"
  },
  {
    "startMs": 813800,
    "text": "government, and a bunch of others,"
  },
  {
    "startMs": 816000,
    "text": "there's just a lot of complexity in a"
  },
  {
    "startMs": 817800,
    "text": "lot of these document types. And if"
  },
  {
    "startMs": 819200,
    "text": "you're actually trying to unlock context"
  },
  {
    "startMs": 821040,
    "text": "at scale,"
  },
  {
    "startMs": 822320,
    "text": "most of the models are not up for the"
  },
  {
    "startMs": 823760,
    "text": "task. And so, we've created this thing"
  },
  {
    "startMs": 825920,
    "text": "called ParseBench, which is a"
  },
  {
    "startMs": 827160,
    "text": "comprehensive enterprise document"
  },
  {
    "startMs": 828840,
    "text": "benchmark for agents. We we think about"
  },
  {
    "startMs": 831640,
    "text": "it as the most comprehensive enterprise"
  },
  {
    "startMs": 833280,
    "text": "document benchmark. You know, it"
  },
  {
    "startMs": 834640,
    "text": "contains 2,000 human-verified pages. It"
  },
  {
    "startMs": 837560,
    "text": "measures tables, charts, content"
  },
  {
    "startMs": 839600,
    "text": "faithfulness, semantic formatting."
  },
  {
    "startMs": 841720,
    "text": "And it's optimized for"
  },
  {
    "startMs": 843520,
    "text": "just how AI agents are actually able to"
  },
  {
    "startMs": 845640,
    "text": "understand these documents instead of"
  },
  {
    "startMs": 847480,
    "text": "like syntactic correctness."
  },
  {
    "startMs": 849160,
    "text": "We've If you look at parsebench.ai,"
  },
  {
    "startMs": 851640,
    "text": "which is, you know, kind of the it's a"
  },
  {
    "startMs": 853560,
    "text": "fully public page on the internet. It's"
  },
  {
    "startMs": 855800,
    "text": "also available on Hugging Face and"
  },
  {
    "startMs": 857520,
    "text": "Kaggle."
  },
  {
    "startMs": 858600,
    "text": "We benchmark probably like 50 different"
  },
  {
    "startMs": 860440,
    "text": "frontier models, open weight models,"
  },
  {
    "startMs": 862080,
    "text": "specialized OCR solutions. And you"
  },
  {
    "startMs": 864360,
    "text": "really want the Pareto curve to kind of"
  },
  {
    "startMs": 866200,
    "text": "be towards the left and up in terms of"
  },
  {
    "startMs": 869040,
    "text": "accuracy, like extremely high accuracy,"
  },
  {
    "startMs": 871640,
    "text": "but also extremely low cost. And there's"
  },
  {
    "startMs": 874680,
    "text": "just so many different types of"
  },
  {
    "startMs": 875960,
    "text": "documents, where some are a little bit"
  },
  {
    "startMs": 877520,
    "text": "simpler, maybe some are a little bit"
  },
  {
    "startMs": 879480,
    "text": "more complex. And ideally, you want to"
  },
  {
    "startMs": 881720,
    "text": "cover all the points on the curve to"
  },
  {
    "startMs": 883320,
    "text": "deal with the dynamic distribution of"
  },
  {
    "startMs": 885080,
    "text": "various types of documents out there."
  },
  {
    "startMs": 886960,
    "text": "It's pretty clear, even if you look at"
  },
  {
    "startMs": 888440,
    "text": "this graph, that it's"
  },
  {
    "startMs": 890320,
    "text": "like you can increase the complexity of"
  },
  {
    "startMs": 892200,
    "text": "the benchmark, and that document"
  },
  {
    "startMs": 893560,
    "text": "understanding is definitely not a 100%"
  },
  {
    "startMs": 895839,
    "text": "solved. But, you know, you fundamentally"
  },
  {
    "startMs": 898560,
    "text": "need to kind of advance a lot of the"
  },
  {
    "startMs": 900000,
    "text": "core capabilities to make sure they're"
  },
  {
    "startMs": 901720,
    "text": "able to process, unlock the vast trove"
  },
  {
    "startMs": 903680,
    "text": "of enterprise context out there."
  },
  {
    "startMs": 908720,
    "text": "There's kind of like a few different"
  },
  {
    "startMs": 910240,
    "text": "points on this accuracy cost latency"
  },
  {
    "startMs": 912400,
    "text": "Pareto curve. There's what I call like"
  },
  {
    "startMs": 914440,
    "text": "the high accuracy regime, where like,"
  },
  {
    "startMs": 916240,
    "text": "you know, some institutions basically"
  },
  {
    "startMs": 917640,
    "text": "need like 99 to almost 100% accuracy,"
  },
  {
    "startMs": 920440,
    "text": "because basically, the downside of an"
  },
  {
    "startMs": 922040,
    "text": "incorrect extraction is you completely"
  },
  {
    "startMs": 923640,
    "text": "mess up your financial model, you"
  },
  {
    "startMs": 925440,
    "text": "completely, you know, you basically get"
  },
  {
    "startMs": 927320,
    "text": "flagged for fraud or a bunch of other"
  },
  {
    "startMs": 929120,
    "text": "really really bad things."
  },
  {
    "startMs": 930760,
    "text": "In regulated industries like insurance"
  },
  {
    "startMs": 932880,
    "text": "and financial services, we see this a"
  },
  {
    "startMs": 934520,
    "text": "decent amount. This typically means"
  },
  {
    "startMs": 936320,
    "text": "you're willing to pay a little bit more"
  },
  {
    "startMs": 937640,
    "text": "money per page for like deeper agent"
  },
  {
    "startMs": 939360,
    "text": "tech reasoning to at least make sure"
  },
  {
    "startMs": 941040,
    "text": "that you get back the the information in"
  },
  {
    "startMs": 943120,
    "text": "the right format. There's also like the"
  },
  {
    "startMs": 945200,
    "text": "low cost regime. Let's say you're just"
  },
  {
    "startMs": 947120,
    "text": "trying to index, you know, the million"
  },
  {
    "startMs": 948640,
    "text": "plus documents per day within your"
  },
  {
    "startMs": 951040,
    "text": "that's you know, being continually"
  },
  {
    "startMs": 952440,
    "text": "updated within your SharePoint just for"
  },
  {
    "startMs": 954480,
    "text": "like rag knowledge base search. You"
  },
  {
    "startMs": 956160,
    "text": "know, in these cases, you obviously want"
  },
  {
    "startMs": 957800,
    "text": "it to be not like terribly inaccurate,"
  },
  {
    "startMs": 959880,
    "text": "but even if it's a little bit messed up,"
  },
  {
    "startMs": 961480,
    "text": "it's okay, too. Because in the end if"
  },
  {
    "startMs": 963400,
    "text": "you have a sufficiently good agent, it"
  },
  {
    "startMs": 964800,
    "text": "can always dive deeper into the document"
  },
  {
    "startMs": 966840,
    "text": "and surface the right information with"
  },
  {
    "startMs": 968480,
    "text": "the right citations and grounding."
  },
  {
    "startMs": 970440,
    "text": "So for these, you know, being able to"
  },
  {
    "startMs": 971800,
    "text": "create some sort of scalable offline"
  },
  {
    "startMs": 973880,
    "text": "indexing pipeline that has the best like"
  },
  {
    "startMs": 975960,
    "text": "cost constraints"
  },
  {
    "startMs": 977360,
    "text": "is something that is optimal."
  },
  {
    "startMs": 979800,
    "text": "One thing about VLMs based approaches"
  },
  {
    "startMs": 981440,
    "text": "though is that, you know, they're"
  },
  {
    "startMs": 983080,
    "text": "typically not very fast. And I think a"
  },
  {
    "startMs": 986120,
    "text": "lot of times if you have like real-time"
  },
  {
    "startMs": 988200,
    "text": "file uploads, let's say you're using"
  },
  {
    "startMs": 989920,
    "text": "Cloud Co-work, you upload a thousand"
  },
  {
    "startMs": 991600,
    "text": "documents and you need to process it"
  },
  {
    "startMs": 993080,
    "text": "within a minute a minute, like having a"
  },
  {
    "startMs": 995640,
    "text": "bunch of VLMs process that at scale is"
  },
  {
    "startMs": 998120,
    "text": "really tough for basically every single"
  },
  {
    "startMs": 999480,
    "text": "OCR service out there. And and to be"
  },
  {
    "startMs": 1001640,
    "text": "fair, that includes ours, too. I think"
  },
  {
    "startMs": 1003960,
    "text": "in general, there's also some sort of"
  },
  {
    "startMs": 1006480,
    "text": "need for an extremely low latency"
  },
  {
    "startMs": 1008760,
    "text": "solution so that you can actually"
  },
  {
    "startMs": 1010440,
    "text": "process stuff in real-time even if you"
  },
  {
    "startMs": 1012600,
    "text": "have like deeper VLM enabled processing"
  },
  {
    "startMs": 1015160,
    "text": "for kind of like deeper visual"
  },
  {
    "startMs": 1016280,
    "text": "inspection and analysis."
  },
  {
    "startMs": 1018440,
    "text": "And so that's what I call kind of being"
  },
  {
    "startMs": 1020360,
    "text": "in the agent loop. So besides Lama"
  },
  {
    "startMs": 1022440,
    "text": "Parse, which is kind of our commercial"
  },
  {
    "startMs": 1023760,
    "text": "service around like document processing"
  },
  {
    "startMs": 1025560,
    "text": "and extraction, we also created this"
  },
  {
    "startMs": 1027520,
    "text": "tool called Light Parse."
  },
  {
    "startMs": 1029400,
    "text": "It is surprisingly really really good. I"
  },
  {
    "startMs": 1031600,
    "text": "don't know if you've been following some"
  },
  {
    "startMs": 1032720,
    "text": "of the Twitter threads, but it is"
  },
  {
    "startMs": 1035079,
    "text": "Rust-based. It is the fastest"
  },
  {
    "startMs": 1036880,
    "text": "open-source parser out there. It is"
  },
  {
    "startMs": 1038720,
    "text": "completely free, um, and there is"
  },
  {
    "startMs": 1040400,
    "text": "basically no strings attached. I think"
  },
  {
    "startMs": 1041680,
    "text": "it's like MIT or Apache license. Um, and"
  },
  {
    "startMs": 1044480,
    "text": "it basically is the most accurate like"
  },
  {
    "startMs": 1046199,
    "text": "markdown parser out there that doesn't"
  },
  {
    "startMs": 1048199,
    "text": "use a VLM or any sort of kind of like"
  },
  {
    "startMs": 1049920,
    "text": "deeper model. Um, and so this is kind of"
  },
  {
    "startMs": 1053040,
    "text": "nice because you can use it as a default"
  },
  {
    "startMs": 1055960,
    "text": "in the assistive agent loop. Um, let's"
  },
  {
    "startMs": 1058040,
    "text": "say you're uploading a a bunch of"
  },
  {
    "startMs": 1059520,
    "text": "documents to Quad Code, Quad Code Work,"
  },
  {
    "startMs": 1061400,
    "text": "Codex, and you want it to process like a"
  },
  {
    "startMs": 1063720,
    "text": "thousand PDFs extremely quickly. You can"
  },
  {
    "startMs": 1066040,
    "text": "always do that, um, and then, you know,"
  },
  {
    "startMs": 1068120,
    "text": "equip a VLM-based parser like Llama"
  },
  {
    "startMs": 1070160,
    "text": "Parser or other frontier models as a"
  },
  {
    "startMs": 1071920,
    "text": "tool. So, what these agents will do is"
  },
  {
    "startMs": 1074800,
    "text": "they'll do like a fast pass over all the"
  },
  {
    "startMs": 1076600,
    "text": "documents first, uh, uh, kind of like"
  },
  {
    "startMs": 1078920,
    "text": "just scan through all the context"
  },
  {
    "startMs": 1080240,
    "text": "extremely efficiently, um, and then if"
  },
  {
    "startMs": 1082440,
    "text": "actually needs to dive into a page with"
  },
  {
    "startMs": 1084400,
    "text": "like tables, with like charts, and"
  },
  {
    "startMs": 1086000,
    "text": "actually needs to more deeply understand"
  },
  {
    "startMs": 1087560,
    "text": "the values, um, it will use a VLM-based"
  },
  {
    "startMs": 1089760,
    "text": "tool, slower processing, to actually"
  },
  {
    "startMs": 1091440,
    "text": "make sure it reads the information"
  },
  {
    "startMs": 1092640,
    "text": "correctly. This is available as a"
  },
  {
    "startMs": 1094520,
    "text": "one-click installable skill, um,"
  },
  {
    "startMs": 1096320,
    "text": "complements kind of any other deeper"
  },
  {
    "startMs": 1098000,
    "text": "VLM-based OCR tool you want to use, um,"
  },
  {
    "startMs": 1100680,
    "text": "and we kind of designed to make it as"
  },
  {
    "startMs": 1102520,
    "text": "fast as possible and also easy to plug"
  },
  {
    "startMs": 1104400,
    "text": "in to your favorite AI agent."
  },
  {
    "startMs": 1108680,
    "text": "So, I kind of speed ran through a bunch"
  },
  {
    "startMs": 1110920,
    "text": "of this stuff, but basically, you know,"
  },
  {
    "startMs": 1112840,
    "text": "uh, we spent a bunch of time on the, uh,"
  },
  {
    "startMs": 1115040,
    "text": "parsing layer. There's also other, um,"
  },
  {
    "startMs": 1117520,
    "text": "general components around like the"
  },
  {
    "startMs": 1118920,
    "text": "semantic and storage layer in terms of"
  },
  {
    "startMs": 1120840,
    "text": "document extraction and search, and of"
  },
  {
    "startMs": 1122800,
    "text": "course like document workflows. Um, and"
  },
  {
    "startMs": 1125120,
    "text": "due to time, I'll probably kind of just,"
  },
  {
    "startMs": 1126800,
    "text": "uh, skip some of the, uh, unexplored"
  },
  {
    "startMs": 1128760,
    "text": "areas like agent native document"
  },
  {
    "startMs": 1130280,
    "text": "formats, hill climbing as a service, and"
  },
  {
    "startMs": 1132240,
    "text": "others, but I'll share the full set of"
  },
  {
    "startMs": 1133720,
    "text": "slides online. In terms of the semantic"
  },
  {
    "startMs": 1136160,
    "text": "and storage layer, you know, besides"
  },
  {
    "startMs": 1137840,
    "text": "document parsing, um, a lot of use cases"
  },
  {
    "startMs": 1140480,
    "text": "also require actually getting back, uh,"
  },
  {
    "startMs": 1142600,
    "text": "structured information at scale from"
  },
  {
    "startMs": 1144280,
    "text": "documents. Whether you're processing,"
  },
  {
    "startMs": 1146200,
    "text": "you know, a million invoices or expense"
  },
  {
    "startMs": 1147880,
    "text": "reports or receipts or claims, you need"
  },
  {
    "startMs": 1150200,
    "text": "to make sure that, you know, you want to"
  },
  {
    "startMs": 1151680,
    "text": "actually get back structured outputs"
  },
  {
    "startMs": 1153760,
    "text": "that you can put into a downstream"
  },
  {
    "startMs": 1155160,
    "text": "database or system."
  },
  {
    "startMs": 1156760,
    "text": "And so, a lot of these use cases"
  },
  {
    "startMs": 1158560,
    "text": "basically revolve around the form of,"
  },
  {
    "startMs": 1160640,
    "text": "you know, how do you automate a lot of"
  },
  {
    "startMs": 1162560,
    "text": "workloads that humans typically do in"
  },
  {
    "startMs": 1164480,
    "text": "scanning a lot of paperwork and doing"
  },
  {
    "startMs": 1166240,
    "text": "data entry. Whether it is kind of,"
  },
  {
    "startMs": 1168760,
    "text": "again, invoices, claims, contracts,"
  },
  {
    "startMs": 1170680,
    "text": "receipts, or others."
  },
  {
    "startMs": 1172120,
    "text": "We kind of created these capabilities"
  },
  {
    "startMs": 1173880,
    "text": "within Llama Parse as well."
  },
  {
    "startMs": 1175720,
    "text": "A lot of our capabilities are actually"
  },
  {
    "startMs": 1177880,
    "text": "tuned towards like low cost while"
  },
  {
    "startMs": 1179840,
    "text": "extremely high accuracy. And you get"
  },
  {
    "startMs": 1182000,
    "text": "back granular citations all the way back"
  },
  {
    "startMs": 1184120,
    "text": "to the source document for every"
  },
  {
    "startMs": 1185480,
    "text": "extracted output. And of course, you can"
  },
  {
    "startMs": 1187320,
    "text": "run this in a pipeline at scale with"
  },
  {
    "startMs": 1189280,
    "text": "confidence scores"
  },
  {
    "startMs": 1190680,
    "text": "and also, you know, being able to"
  },
  {
    "startMs": 1192880,
    "text": "actually flag whether or not we're"
  },
  {
    "startMs": 1194400,
    "text": "certain about a certain value before"
  },
  {
    "startMs": 1196520,
    "text": "deciding to put it into some sort of"
  },
  {
    "startMs": 1198120,
    "text": "system."
  },
  {
    "startMs": 1200160,
    "text": "There's also document search, which is a"
  },
  {
    "startMs": 1202160,
    "text": "basically expanded tool set as I"
  },
  {
    "startMs": 1203720,
    "text": "mentioned around retrieval, BM25, grep,"
  },
  {
    "startMs": 1206560,
    "text": "vector search, reading, and scrolling."
  },
  {
    "startMs": 1208880,
    "text": "And so, all these capabilities are"
  },
  {
    "startMs": 1210560,
    "text": "available within some of our commercial"
  },
  {
    "startMs": 1212120,
    "text": "platforms as well as open source"
  },
  {
    "startMs": 1213520,
    "text": "offerings. But I also just wanted to"
  },
  {
    "startMs": 1215280,
    "text": "paint a picture of the general concepts"
  },
  {
    "startMs": 1216760,
    "text": "out there today."
  },
  {
    "startMs": 1218200,
    "text": "So, I'll skip this section about kind of"
  },
  {
    "startMs": 1219640,
    "text": "what's next and then maybe just go all"
  },
  {
    "startMs": 1222000,
    "text": "the way to the end. I know I'm a little"
  },
  {
    "startMs": 1223920,
    "text": "bit over time. So, really appreciate you"
  },
  {
    "startMs": 1225680,
    "text": "all spending time today and then let me"
  },
  {
    "startMs": 1227760,
    "text": "just how do I get to this part really"
  },
  {
    "startMs": 1230000,
    "text": "quick? Oh, right."
  },
  {
    "startMs": 1233400,
    "text": "I'm going to skip this piece. I'll put"
  },
  {
    "startMs": 1235160,
    "text": "put this online."
  },
  {
    "startMs": 1236720,
    "text": "Our booth is at LG 47. If you guys are"
  },
  {
    "startMs": 1239360,
    "text": "interested in stopping by and we're"
  },
  {
    "startMs": 1241240,
    "text": "hosting a giant pickleball tournament"
  },
  {
    "startMs": 1242760,
    "text": "today. So, thank you for your time."
  },
  {
    "startMs": 1245824,
    "text": ">> [applause]"
  }
]