Content History
A lot of PDFs have charts, and parsing charts is hard. All frontier models out of the box are notoriously bad at interpreting chart values without explicit guidance. There’s a lot of ways you can throw compute/tokens at the problem (gpt-5.2-pro is pretty good), and there’s still a bunch of edge cases we need to fix. We’ve created some awesome visual understanding capabilities within LlamaParse at a cheap price. Our agentic mode supports good approximate chart parsing at 1c per page (lower with cost optimization/discounts), less than our comparable solutions at a few cents a page. In the picture below we overlaid a source line chart with our parsed chart on top. For simple line charts it basically matches perfectly. Whether it’s a financial report or marketing presentation, check out our agentic mode + agentic parsing toggles to help OCR your docs! Sign up: https://www.llamaindex.ai/signup
Excel parsing is underrated Transforming it into a 2D structured format makes it a lot easier for agents to work with. Come check out our webinar on the 29th! We’re going to do a deep-dive on using LlamaSheets to structure complex Excel tables into tabular formats and use it in downstream workflows. https://landing.llamaindex.ai/messy-spreadsheets-to-ai-ready-data?utm_source=socials&utm_medium=li_social
Tired of wrestling with messy Excel files that break traditional parsers? 📊 Join our workshop on January 29th at 11 AM PT to see how LlamaSheets transforms chaotic spreadsheets into clean, AI-ready Parquet files while preserving all the context that matters. 📈 Handle complex
RT LlamaIndex 🦙 Context engineering isn't just about throwing more data at your LLM - it's about giving it the right context at the right time. In this talk for @OReillyMedia, our Developer Relations Engineer @tuanacelik walks through how memory blocks help you build agents that maintain structured context for complex tasks. She demonstrates artifact memory blocks using a restaurant order tracking bot - showing how to distill entire conversations down to just the essential structured information (pizza type, toppings, address) rather than processing the full chat history. Key concepts covered: · Different types of memory blocks (static, fact extraction, vector, artifact) · Context ratio management - balancing chat history vs system prompts vs memory · Using agent workflows to construct and optimize context step-by-step The example shows how artifact memory blocks can transform a meandering conversation into a clean, structured order - exactly what you need for production agents handling real-world tasks. Watch the full talk: https://www.youtube.com/watch?v=POO3ckpJ8NQ Original tweet: https://x.com/llama_index/status/2011484220494098473
RT Tuana Ok, we've scheduled our first technical webinar of the year. Join us (me and @LoganMarkewich) on Jan 29th for a deep-dive on LlamaSheets (by @llama_index). Which is one of our latest tools in LlamaCloud designed to get sheets (which can actually be multiple sheets) with complex layouts into a format that both LLMs and traditional data tools can work with. If you've ever tried to parse a spreadsheet with merged cells, hidden columns, or data scattered across multiple tabs, you know the pain. We'll show you how LlamaSheets handles all of it: https://landing.llamaindex.ai/messy-spreadsheets-to-ai-ready-data Original tweet: https://x.com/tuanacelik/status/2011435695500562807
Do filesystem tools 🗃️ kill vector search❓ File search is all the rage right now. Give claude code access to the CLI and it turns into a generalized computer use agent. This is a great blog post from @itsclelia that benchmarks file search vs. vector search. Up until ~100 files, the speed is comparable. From 1k files onwards, vector search is way faster to get an initial result. File search does enable the agent to do dynamic reasoning to answer more complex questions (e.g. linking across different sections in the same doc). The real answer is probably blending both, ala what we did with semtools! https://github.com/run-llama/semtools Check out our blog: https://www.llamaindex.ai/blog/did-filesystem-tools-kill-vector-search?utm_source=socials&utm_medium=li_social We used LlamaParse to parse all files into LLM-readable formats. fs-explorer repo: https://github.com/run-llama/fs-explorer LlamaCloud: https://www.llamaindex.ai/signup
Can filesystem tools really replace vector search? We put agentic file exploration to the test against traditional RAG. Our experiment with fs-explorer agent vs. hybrid RAG revealed some surprising insights about when each approach shines: 🏃 RAG is faster - averaging 3.81
View quoted postUnderstanding embedded visual elements is a huge pain point for document OCR. With our latest release, we offer one of the cheapest ways for you to get awesome visual parsing over charts, diagrams, etc. 💫 Screenshotting the page into a frontier VLM leads to a ton of failure modes; especially when you’re resolution constrained by the API provider. A lot of embedded elements like charts have a ton of visual detail that require specialized understanding. We upgraded our agentic mode within LlamaParse and it’s fantastic at parsing visual subelements within a document, and stringing it together in to coherent markdown. It also comes with all the requisite metadata like bounding boxes, confidence scores, embedded images, and more. Check out the parsed page below. The parser is able to accurately infer color mappings in the bar chart 📊 and infer intermediate, nonmarked values in the line chart 📈. Sign up below! https://www.llamaindex.ai/signup
The hack to dealing with any long context is just offloading it to files and letting the agent + file system tool loop 🤖🔁🗂️do its magic. Includes the agent's own chat/context history + MCP responses. Don't do the naive ReAct agent + tool loop anymore
Learn about how we use the filesystem to improve context efficiency for tools, MCP servers, skills, terminals, chat history, and more. https://cursor.com/blog/dynamic-context-discovery
RT LlamaIndex 🦙 Transform messy Excel files into AI-ready data with LlamaSheets—our solution for parsing complex spreadsheets while preserving semantic context and hierarchical structure. Join our live workshop: https://landing.llamaindex.ai/messy-spreadsheets-to-ai-ready-data?utm_source=socials&utm_medium=li_social January 29th, 11 AM PT 📊 Handle merged cells, multi-level headers, and visual formatting that traditional parsing tools miss 🤖 Build spreadsheet-specific agents for financial analysis, budget parsing, and automated reporting ⚡ Convert spreadsheet chaos into structured Parquet files with just a few lines of code 🔗 Integrate seamlessly with LlamaAgents for complete workflow automation We'll demonstrate real examples including building financial analysis agents and consolidating multi-region data from large sheets. Perfect for anyone dealing with budget files, quarterly revenue tables, or any spreadsheet data that needs to feed AI systems. Original tweet: https://x.com/llama_index/status/2008587677680234659
As 2025 comes to a close 🎁, I want to highlight the evolution of @llama_index as a company 🚫 We are no longer “a RAG framework” ✅ We’ve built best-in-class technology around document OCR + workflows to provide high-quality context to *any* emerging agent application out there Context and agent engineering are ever evolving (e.g. RAG 1.0 is dead, MCP might be dead too, nowadays it’s all about coding agents + skills + file search/coding execution). But has remained constant is the need for highly-accurate/cost-effective modules to parse the most complex document types. No matter how good the models become, there is no way for them to read file binaries on their own. You need an accurate translation layer to make sense of the vast amounts of unstructured file containers. Our mission is to automate knowledge work over your documents. And that starts by making sense of your document-based data. We’ve seen a massive spike in users, have processed half a billion pages, and have crossed 25M+ monthly downloads on our LlamaCloud client package. Check out everything we’ve released this year: ⭐️ LlamaParse v2: massive upgrades to our parsing quality ⭐️ LlamaSheets: Excel API ⭐️ LlamaClassify/LlamaSplit: complementary blocks to classify/postprocess your docs ⭐️ LlamaAgents: initial e2e templates to help your orchestrate your document workflows Whether you’re an AI startup or F500 enterprise, if you need document processing tech, come follow us / check us out! Here’s to a wonderful 2026 📈🍾 https://www.llamaindex.ai/blog/llamaindex-newsletter-2025-12-30
As we wrap up 2025, we're incredibly proud of what this team shipped. We set out to solve document AI reliability and help you develop your own task specific document agents. Take a look back on the year 2025 with us, from the launch of LlamaAgents to brand new MCP support 👇
RT LlamaIndex 🦙 As we wrap up 2025, we're incredibly proud of what this team shipped. We set out to solve document AI reliability and help you develop your own task specific document agents. Take a look back on the year 2025 with us, from the launch of LlamaAgents to brand new MCP support 👇 Original tweet: https://x.com/llama_index/status/2006050920074006609
Supercharge Claude Code with better Excel understanding 📊 Coding agents are general enough to do any type of knowledge work, including reading/creating docs. There are some pre-built skills for Claude Code to read Excel sheets, but they kind of suck 🚫 - it requires the agent to iterate cell-by-cell through raw openpyxl code. Our LlamaSheets API is specifically designed to parse complex multi-table, hierarchical Excel sheets into well-structured 2d representations. @tuanacelik wrote a brand new guide showing you how to use it with your favorite coding agent including Claude Code! Check it out: https://developers.llamaindex.ai/python/cloud/llamasheets/examples/coding_agent/
I went offline for a couple of days to be with family and it seems like all we talked about on this platform has been coding agents. A completely new way of thinking about documentation, oss projects, developer tools we provide has actually been figuring out how to structure
View quoted postRT LlamaIndex 🦙 Santa receives thousands of letters every year—Processing these manually takes a lot of time, so this year, we're helping him out 👇 We built an agent to automate extracting wish-list items from letters 🎅 📤 Upload to LlamaCloud ✂️ LlamaSplit categorizes pages into logical segments (letters vs. index pages) using AI-powered document understanding 📋 LlamaExtract extracts structured data from each child's letter—name, age, wishlist items, address, and whether they've been good or bad (using Pydantic schemas for type-safe extraction) 🔄 LlamaAgent Workflows orchestrates the process with a fan-in pattern: split the document into segments, then extract data from each letter segment in parallel The result? Transform a messy multi-page PDF into clean, structured JSON for every child's wishlist—automatically handling document segmentation, parallel extraction, and data validation. Try it yourself in this Colab notebook: https://colab.research.google.com/drive/1fZRoOTm46BO2j5iSTHFZy1QxqC71RrmJ?usp=sharing Original tweet: https://x.com/llama_index/status/2003548247365693887
The DOJ messed up some redactions on the latest Epstein files 🗄️🔏 - they didn’t flatten the PDF layers and you can highlight/copy the underlying text. If you want to extract this text at scale, you *can’t* just feed everything to a VLM (gpt-5.2, sonnet-4.5, gemini 3). VLMs only look at the top-level visual layer of the page, and will output the redacted blocks. You need to also reconstruct the text from the PDF binary itself, which is more in line with “traditional” techniques. LlamaParse uses a combination of both VLMs along with reading the underlying binary. * If you try out our agentic mode by default, it will output the redacted blocks in the markdown `md` field, but extract out the full text in the `text` field * With a simple prompt change you can also extract out the full text in `md`. Prompt: "Do not output redactions if the underlying extracted text already exists - output the full extracted text instead" Whether you want to comb through any set of released government documents or any other file, come check out LlamaParse! Source reddit thread: https://www.reddit.com/r/Epstein/s/ax3wevkF1F File: https://www.justice.gov/multimedia/Court%20Records/Matter%20of%20the%20Estate%20of%20Jeffrey%20E.%20Epstein,%20Deceased,%20No.%20ST-21-RV-00005%20(V.I.%20Super.%20Ct.%202021)/2022.03.17-1%20Exhibit%201.pdf To use LlamaParse, sign up to LlamaCloud: https://cloud.llamaindex.ai/
RT LlamaIndex 🦙 We love seeing all of the out of office emails around the holidays! So we thought we'd keep things light this week. We'll share some fun little demos, great way to get started with LlamaIndex too🎄 Today: check out this example where we build an agent that helps navigate a corrupted database! The challenge: Santa's toy database is corrupted. Parent requests are flooding in, but toy names are gone—just placeholders. The fix: An automated workflow using LlamaClassify (routing), LlamaExtract (data extraction), LlamaSheets (database search), and LlamaIndex Workflows (orchestration) to process thousands of requests without manual work. Try the Colab: https://colab.research.google.com/drive/1GSsrv-3kDG3jfxEx_T2e08YcIkmXS5S2?usp=sharing Original tweet: https://x.com/llama_index/status/2003165699238473736
With our latest release of LlamaParse v2, we’ve created one of the cheapest ways to add visual understanding capabilities to document parsing ⚡️📑 Our new cost effective mode (≤ 0.3c a page 💵) can effectively read and parse charts and diagrams into coherent tables. It matches color codes/labels, and semantically reconstructs the text. This is a great way to ingest millions of docs for your RAG or agentic workflows! It is cheaper and better than trying to manually prompt ChatGPT LlamaCloud: https://cloud.llamaindex.ai/
LlamaParse v2 is here: simpler configuration, better performance, and up to 50% cost reduction for document parsing. We'v redesigned our approach based on user feedback to make production-ready document ingestion accessible without becoming a parsing expert. 🚀 Four simple
Introducing LlamaParse v2 🔥 We’ve released a big new update here to push the frontiers of document processing and create an easier/more stable experience. ✅ We’ve standardized across 4 fixed tiers: fast, cost-effective, agentic, agentic plus ✅ We created massive performance improvements on our cheaper modes (cost-effective/agentic). They have fewer hallucinations and native visual understanding capabilities. ✅ We made our best mode (agentic plus) cheaper ✅ We’ve created versioned parsing - you can anchor to an existing version to guarantee parsing quality, or always stay on `latest` Blog: https://www.llamaindex.ai/blog/introducing-llamaparse-v2-simpler-better-cheaper Sign up to LlamaCloud: https://cloud.llamaindex.ai/
LlamaParse v2 is here: simpler configuration, better performance, and up to 50% cost reduction for document parsing. We'v redesigned our approach based on user feedback to make production-ready document ingestion accessible without becoming a parsing expert. 🚀 Four simple
RT LlamaIndex 🦙 LlamaParse v2 is here: simpler configuration, better performance, and up to 50% cost reduction for document parsing. We'v redesigned our approach based on user feedback to make production-ready document ingestion accessible without becoming a parsing expert. 🚀 Four simple tiers replace complex configuration modes - just pick Fast, Cost Effective, Agentic, or Agentic Plus based on your needs 💰 Significant cost improvements across all tiers, with our Agentic Plus tier offering 50% price reduction ⚡ Enhanced accuracy and reduced hallucinations, especially for complex documents with multimodal content Instead of tuning dozens of parameters and choosing between parsing modes, you now focus on outcomes while we handle the optimal model routing automatically. This means we can ship performance upgrades more frequently without breaking your workflows. Read the full announcement and learn about the new tier system: https://www.llamaindex.ai/blog/introducing-llamaparse-v2-simpler-better-cheaper?utm_source=socials&utm_medium=li_social Original tweet: https://x.com/llama_index/status/2001748317072756880
I love the idea of an AI chief of staff We’re excited to partner with @intelligenceco. They’re building Cofounder, an autonomous agent that can integrate with every system within a startup (Notion, Slack, Gmail, Github), ingest global company context, and take actions on behalf of the user. As a heavy Claude/ChatGPT user, I haven’t scratched the surface on time efficiency. There are infinite tasks that still take too long to prompt, setup tools, and automate completely. They’re using LlamaParse for helping to OCR their unstructured docs. Come check out our joint blog! https://www.llamaindex.ai/customers/how-the-general-intelligence-company-turns-business-documents-into-agent-ready-context-with
See how @intelligenceco built Cofounder, an AI chief of staff that turns business documents into agent-ready context at scale. 📄 LlamaParse handles continuous ingestion from @gmail, @SlackHQ, @linear, @notionhq, and @github every 30 minutes - processing PDFs, images, and
RT LlamaIndex 🦙 See how @intelligenceco built Cofounder, an AI chief of staff that turns business documents into agent-ready context at scale. 📄 LlamaParse handles continuous ingestion from @gmail, @SlackHQ, @linear, @notionhq, and @github every 30 minutes - processing PDFs, images, and attachments with agentic OCR 🤖 Two-stage retrieval system combines vector similarity with agent reasoning to filter by time, source, and ownership across multiple business tools 💰 Achieved lower costs and latency compared to managed RAG solutions while avoiding weeks of custom parser development ⚡ Freed engineering time to focus on core differentiator - building agents that can act - instead of document infrastructure "It probably would've taken us a month or more to build a worse document parser ourselves. LlamaParse let us focus on the agents instead of reinventing infrastructure." Read the full case study: https://www.llamaindex.ai/customers/how-the-general-intelligence-company-turns-business-documents-into-agent-ready-context-with?utm_source=socials&utm_medium=li_social Original tweet: https://x.com/llama_index/status/2001385174261760093
Building “RAG 2.0” is just making Claude Code running over your filesystem 🤖🗂️ To make this work well, you need to solve three things 1️⃣ Virtualize your filesystem to prevent the agent from messing stuff up. AgentFS by @tursodatabase is a nice example of how you can give the agent access to a copy of all your files without messing up your raw data. 2️⃣ Parse unstructured documents like PDFs, pptx, Word into an LLM-ready format. Agentic OCR solutions like LlamaParse can help here 3️⃣ Creating an agentic loop with human-in-the-loop. If you want to control the agent implementation instead of using Claude Code out of the box, you can use @llama_index workflows to help orchestrate these long-running agent tasks. Shoutout @itsclelia, check it out! Blog: https://www.llamaindex.ai/blog/making-coding-agents-safe-using-llamaindex Repo: https://github.com/run-llama/agentfs-claude
Secure your coding agents with virtual filesystems and better document understanding. Building safe AI coding agents requires solving two critical challenges: filesystem access control and handling unstructured documents. We've created a solution using AgentFS, LlamaParse, and
RT LlamaIndex 🦙 Secure your coding agents with virtual filesystems and better document understanding. Building safe AI coding agents requires solving two critical challenges: filesystem access control and handling unstructured documents. We've created a solution using AgentFS, LlamaParse, and @claudeai. 🛡️ Virtual filesystem isolation: agents work with copies, not your real files, preventing accidental deletions while maintaining full functionality 📄 Enhanced document processing: LlamaParse converts PDFs, Word docs, and presentations into high-quality text that agents can actually understand ⚡ Workflow orchestration: LlamaIndex Workflows provide stepwise execution with human-in-the-loop controls and resumable sessions 🔧 Custom tool integration: replace built-in filesystem tools with secure MCP server alternatives that enforce safety boundaries This approach uses AgentFS (by @tursodatabase) as a SQLite-based virtual filesystem, our LlamaParse for state-of-the-art document extraction, and Claude for the coding interface - all orchestrated through LlamaIndex Agent Workflows. Read the full technical deep-dive with implementation details: https://www.llamaindex.ai/blog/making-coding-agents-safe-using-llamaindex?utm_source=socials&utm_medium=li_social Find the code on GitHub: https://github.com/run-llama/agentfs-claude Original tweet: https://x.com/llama_index/status/2000612235505467824
GPT-5.2 Thinking is really good at parsing charts 📊 I threw in some charts into the raw ChatGPT UI after OpenAI hyped up GPT-5.2’s visual capabilities 👇 The native visual understanding capability of GPT-5.2 is not amazing - see the plotted graph for GPT-5.2. But both GPT-5.2 Thinking and Pro make up for that by spending a *ton* on reasoning tokens in order to break down the chart image and plot every point. The plotted points by GPT-5.2 Thinking and Pro are spot on (there are maybe small discrepancies but it’s also really hard to tell by the human eye) If you look at the reasoning trace within the ChatGPT UI, you’ll find that GPT-5.2 will spend a lot of reasoning tokens on writing code to break down the image, analyzing each axis, and getting the lines. Check out the results in the image 🖼️ The cool finding here is that models can make up for poor “one-shot” understanding by just adding a ton of thinking tokens on top. ⚠️ Of course if you’re actually trying to parse a bunch of chart data efficiently this isn’t very practical and quite slow/expensive. If you’re looking for good/much cheaper chart understanding check out LlamaCloud!
We made a simple cli command `ask` which lets you ask questions over any arbitrary folder in your filesystem. It's a specialized agent that does super-efficient indexing/search. Can natively help you parse all your pdfs/powerpoints/word docs. You can also plug it into Claude Code to give it an efficient way to query your filesystem for context, without polluting the context window with a ton of raw file text. Check out semtools! https://github.com/run-llama/semtools
"ask" and you shall receive! SemTools now ships with a dedicated "ask" CLI command - performs agentic search over documents - combine with `parse` to create QA workflows over unstructured data - cache your indexes with `workspaces` Learn more: https://github.com/run-llama/semtools/discussions/44
View quoted postWe just launched a specialized agent for document splitting 📑✂️ This is like semantic chunking on steroids, across complex document packets. A lot of documents are stapled together collections of mini “sub-documents”. Each document packet can contain a bunch of subdocs of one or multiple types: - A packet of resumes - Expense reports containing reimbursement form + receipt images - Court filings: complaint/exhibits/orders in a single PDF Our agent lets you do this automatically and route it to downstream workflows: extraction with separate schemas per doc, document parsing with different settings, or higher-order chunking for knowledge base/RAG/agentic workflows. Come check it out 🔥: https://www.llamaindex.ai/blog/split-document-into-clear-targeted-sections-with-llamasplit?utm_source=socials&utm_medium=li_social Docs: https://developers.llamaindex.ai/python/cloud/split/getting_started/ Signup: https://cloud.llamaindex.ai/
Split documents into distinct sections automatically with our new LlamaSplit API 📄✂️ We're excited to introduce LlamaSplit (now in beta), which uses AI to automatically separate bundled documents into clear, targeted sections based on categories you define - no more manual
RT LlamaIndex 🦙 Split documents into distinct sections automatically with our new LlamaSplit API 📄✂️ We're excited to introduce LlamaSplit (now in beta), which uses AI to automatically separate bundled documents into clear, targeted sections based on categories you define - no more manual splitting of document stacks. 📋 Analyze page content and classify pages into your defined categories with natural language descriptions 🎯 Get back precise segments with exact page ranges and confidence scores for each section ⚡ Handle real-world scenarios like resume stacks, mixed financial documents, court filings, and research paper collections 🔗 Combine with LlamaExtract to run targeted extraction on each segment or route to appropriate agent workflows Perfect for processing resume bundles,handling mixed document types, legal teams organizing court filings, categorizing patient charts and more. Watch an example of segmenting (an AI generated) bundle of resumes below 👇 Read the full announcement and get started with LlamaSplit: https://www.llamaindex.ai/blog/split-document-into-clear-targeted-sections-with-llamasplit?utm_source=socials&utm_medium=li_social Docs: https://developers.llamaindex.ai/python/cloud/split/getting_started?utm_source=socials&utm_medium=li_social Original tweet: https://x.com/llama_index/status/1998516266907394185
Scalably Parsing 1M+ PDFs with AI Agents 📈📑 Here’s a simple tutorial we wrote up showing you how to parse a directory of an arbitrary number of PDFs through our service in a reliable, efficient manner. LlamaParse is designed to handle very large workloads; with some simple primitives like asyncio/semaphores along with setting worker parameters, you’ll be able to get the results much more quickly than doing a forloop through everything. Check it out! https://developers.llamaindex.ai/python/cloud/llamaparse/examples/async_parse_folder/?utm_source=socials&utm_medium=li_social
Need to parse multiple PDFs efficiently? Learn how to use LlamaParse with async batch processing. 📁 Process entire folders of PDFs simultaneously instead of one-by-one ⚡ Use asyncio and semaphores to control how many files parse concurrently 🎯 Prevent API rate limit errors
View quoted postEvery NeurIPS/ICML/ICLR "Related Works" section is just deep research? 👇
There are a LOT more papers that need to be desk rejected at ICLR. Somehow, this hallucination wasn't caught. So, I went on a crazy rabbit hole and found 50 more (many are just as funny). We're approaching crisis levels Paper titles included below 👇 https://gptzero.me/news/iclr-2026/
We’re building out an applied research team to push SOTA on document understanding using LLMs/VLMs and other emerging techniques 📈📑 We’re on a mission to understand and orchestrate the most complex document types, from PDFs to Excel. You’re responsible for research, evals, and productization. The work you do will impact thousands to millions of developers across large enterprise to digital-native startups in unlocking context from any unstructured data. Simon and I have a deep appreciation for both high-quality research from top conferences (NeurIPS, CVPR, ACL, etc.) and also 0-1 startup work. We have a lot of ideas and GPUs but need additional resources to help us out! If this sounds fun come join us: https://www.llamaindex.ai/careers/multimodal-ai-engineer-document-understanding
Document understanding is a huge use case for VLMs, but historically there's been no single "good" benchmark to measure progress here (unlike SWE-bench for coding). This past week I did a deep dive into OlmOCR-Bench, a recent document OCR benchmark that is a huge step in the right direction. ✅ It covers 1400+ PDFs containing formulas, tables, tiny text, and more ✅ It uses binary, verifiable unit tests that are super cheap to run. That said there's still some room to go: 🟡 There's a lot of types of data that still needs to be covered - complex tables, chart understanding, form rendering, handwriting, foreign language, and more 🟡 The binary unit tests are still quite coarse + sometimes use brittle exact matching. Check out my blog: https://www.llamaindex.ai/blog/olmocr-bench-review-insights-and-pitfalls-on-an-ocr-benchmark FWIW we do quite well over this and recently upgraded our default modes too: https://cloud.llamaindex.ai/
OCR benchmarks matter, so in this blog @jerryjliu0 analyzes OlmOCR-Bench, one of the most influential document OCR benchmarks. TLDR: it’s an important step in the right direction, but doesn’t quite cover real-world document parsing needs. 📊 OlmOCR-Bench covers 1400+ PDFs with
RT LlamaIndex 🦙 OCR benchmarks matter, so in this blog @jerryjliu0 analyzes OlmOCR-Bench, one of the most influential document OCR benchmarks. TLDR: it’s an important step in the right direction, but doesn’t quite cover real-world document parsing needs. 📊 OlmOCR-Bench covers 1400+ PDFs with binary pass-fail tests, but focuses heavily on academic papers (56%) while missing invoices, forms, and financial statements 🔍 The benchmark's unit tests are too coarse for complex tables and reading order, missing merged cells, chart understanding, and global document structure ⚡ Exact string matching in tests creates brittleness where small formatting differences cause failures, even when the extraction is semantically correct 🏗️ Model bias exists since the benchmark uses Sonnet and Gemini to generate test cases, giving advantages to models trained on similar outputs Our preliminary tests show that LlamaParse shines at deep visual reasoning over figures, diagrams, and complex business documents. Read our Jerry's analysis of OCR benchmarking challenges and what next-generation document parsing evaluation should look like: https://www.llamaindex.ai/blog/olmocr-bench-review-insights-and-pitfalls-on-an-ocr-benchmark?utm_source=socials&utm_medium=li_social Original tweet: https://x.com/llama_index/status/1996637860032925898
RT Tuana Have you used LlamaSheets and/or LlamaAgents? We'd love to hear from you! Over the next few weeks I'm going to try and speak to as many of you as I can. Let me know if you'd like to have a chat by filling in one of these forms 👇 Chat about LlamaSheets: https://forms.gle/WLSQGqM71tKNi6cXA Chat about LlamaAgents: https://forms.gle/JoH6m6r8sBRpZKcB9 Original tweet: https://x.com/tuanacelik/status/1996236247291797512
Announcing LlamaSheets in beta 🔥 Transform your messy spreadsheets into AI-ready data with our newest LlamaCloud API 📊 LlamaSheets (in beta) is a specialized API that automatically structures complex spreadsheets while preserving their semantic meaning and hierarchical
RT LlamaIndex 🦙 Calling all community members: Join us this Thursday for an office hours in our Discord server, all about LlamaAgents and LlamaSheets. This is a chance to ask anything on your mind about two of our latest releases, and learn about what's coming up next. Drop in anytime from 11AM to 12PM, @tuanacelik, @LoganMarkewich and @itsclelia will all be there. Original tweet: https://x.com/llama_index/status/1995906570002350205
We've been cooking 🧑🍳 📊 Spreadsheet parsing (LlamaSheets) 📑 Multi-step document workflows (LlamaAgents) ❓ 3 new releases coming out this week and next week If you have core use cases around document processing, we'd love to talk to you. Come check out our office hours hosted by @tuanacelik!
The team at @llama_index have been cooking! 🧑🍳 🍳 Over the last few weeks, we released: LlamaAgents: This is agent workflows that come with complete, deployable templates (more coming on this this week!) LlamaSheets: Another addition to LlamaCloud that parses, extracts
RT Tuana The team at @llama_index have been cooking! 🧑🍳 🍳 Over the last few weeks, we released: LlamaAgents: This is agent workflows that come with complete, deployable templates (more coming on this this week!) LlamaSheets: Another addition to LlamaCloud that parses, extracts information and deep context (also hidden in metadata) from tables and spreadsheets, as well as identifying sub-groups from complex sheets That's a lot, so this week, we want to take the time to hear from you! On Thursday, me, @LoganMarkewich and @itsclelia will be in our Discords voice channel for an hour dedicated to chatting about these two new tools. Drop by for our office hours, we'd love to hear from you. And of course, bring along any questions too! Original tweet: https://x.com/tuanacelik/status/1995866683723186340
We wrote a tutorial on extracting massive structured tables from documents 📃 Using naive LLM structured output for document extraction fails if the number of output tokens is large - the LLM will end up dropping or hallucinating results. A lot of documents are basically rendering massive tables in PDF form, like this Blue Shield document showing network coverage across 380+ CA hospitals. We created a new mode in LlamaExtract that lets you extract every single row from this document with 100% accuracy. This lets you: * ETL it into a structured database * Do structured queries over it. * And more! Check out our blog post: https://www.llamaindex.ai/blog/extracting-repeating-entities-from-documents?utm_source=socials&utm_medium=li_social
Stop losing 80% of your data when extracting from long documents with repeating entities like catalogs, tables, and lists. Our new Table Row extraction target in LlamaExtract solves the core problem: instead of trying to extract everything at once (where LLMs get overwhelmed),
We launched a new API today to let you parse any Excel sheet in a structured table. Take a look at this example on core production costs 🌽: 1️⃣ The table is located at the center of the sheet with headers, footnotes, and a hierarchical column layout 2️⃣ We get back a structured table with summarization, along with parsed row/column representations This lets you directly run text-to-pandas/SQL over this data if you’re building an AI agent, or do ETL yourself over it. Check out our blog and come take a look! Blog: https://www.llamaindex.ai/blog/announcing-llamasheets-turn-messy-spreadsheets-into-ai-ready-data-beta Try it out: https://cloud.llamaindex.ai/
Announcing LlamaSheets in beta 🔥 Transform your messy spreadsheets into AI-ready data with our newest LlamaCloud API 📊 LlamaSheets (in beta) is a specialized API that automatically structures complex spreadsheets while preserving their semantic meaning and hierarchical
Introducing LlamaSheets 🦙 - a specialized AI agent that can convert complex spreadsheets into normalized, structured data. Excel files are arbitrarily complicated - they contain semi-structured numerical data, complex formatting, visual hierarchies. A lot of Excel files “look like tables” to the human eye but are impossible to parse as actual tables if you tried to read the raw cell values. Parsing Excel files requires a completely different stack from parsing any other document format. We’ve been hard at work over the past few months doing applied research on understanding Excel files. The result is we’ve created a powerful algorithm to help identify, segment, and output structured tables - and this includes preserving multi-level row and column hierarchies. The output is a structured dataframe that you can directly run queries over, OR feed it as a tool to an upstream AI agent. Any AI agent will have a much easier time understanding these structured values vs. trying to write code (e.g. openpyxl) to manipulate the raw Excel values - it raises accuracy and reduces cost. This was a teamwide effort and huge shoutout to everyone who contributed to this. Check out our video, blog, docs, and signup below! Video: https://www.youtube.com/watch?v=eOp6_vbA5Kc Blog: https://www.llamaindex.ai/blog/announcing-llamasheets-turn-messy-spreadsheets-into-ai-ready-data-beta?utm_source=socials&utm_medium=li_social Docs: https://developers.llamaindex.ai/python/cloud/llamasheets/getting_started/ Sign up: https://cloud.llamaindex.ai/ Happy Thanksgiving week 🦃LlamaIndex 🦙: Announcing LlamaSheets in beta 🔥 Transform your messy spreadsheets into AI-ready data with our newest LlamaCloud API 📊 LlamaSheets (in beta) is a specialized API that automatically structures complex spreadsheets while preserving their semantic meaning and hierarchical Link: https://x.com/llama_index/status/1993362324070318286
We’re looking for strong AI engineers to help us build specialized agents for document understanding. It’s an extremely technical role, and a blend of hard ML tech with agent engineering techniques: ✅ You should know how to train VLMs for solving OCR + extraction tasks ✅ You should know first-principles context/workflow engineering for building agents ✅ You should know how to do applied research but aggressively prioritize in a fast-moving startup ✅ Ideally you’ve also shipped e2e product Simon and I both have extensive ML backgrounds, and we strongly value good technical talent here. In-person in SF. Apply here: https://www.llamaindex.ai/careers/multimodal-ai-engineer-document-understanding
It's pretty clear that AI agents can inhale documents and perform knowledge work. It's also pretty clear humans need to be able to observe this process to prevent things from going off the rails. One of the biggest benefits towards orchestrating document workflows through code (vs. a no-code builder like traditional RPA tools) is native integrations with OpenTelemetry and all the wonderful AI observability tools out there. This is exactly what LlamaAgents provides. This is super important because documents are inherently big containers of unstructured data. Workflows transform this mess of tokens into structured outputs over time. Trying to manually inspect everything is a nightmare. These integrations allow you to dump the full LLM traces into any supported observability tool out there, giving both non-technical and technical users extremely granular insights into what's going on, and allowing steerability in the case of unexpected outputs. Check out our awesome blog by @itsclelia here! https://www.llamaindex.ai/blog/observability-in-agentic-document-workflows LlamaAgents: https://developers.llamaindex.ai/python/llamaagents/overview/LlamaIndex 🦙: Agentic Document Workflows are crucial for AI-driven knowledge work and automation, but they are often treated as black boxes, which leads to silent failures and unexpected behaviors. With our Agent Workflows you don't have to worry about not knowing what is happening behind the Link: https://x.com/llama_index/status/1991183958164553959
I did some weekend reading into the recently released OlmOCR2 model by @jakepoznanski et al. 📄🔎 A cool insight here is that you can scalably do RL on your document parsing model (specifically RLVR) in an automated fashion - without needing humans for feedback or creating the reward function for each document. The authors use Sonnet to generate an HTML scaffold to then generate tailored unit tests for each document. The unit tests themselves are deterministic, but the generation is through an LLM. The parsed outputs are scored against the unit tests, which creates feedback signals for the model. This is nice because: - Manually generating parsed ground-truth for documents is a pain in the ass - Manually generating tests for documents is also a pain in the ass - So scaling up data and signals for document understanding has been painful Blog + paper: https://allenai.org/blog/olmocr-2 From our side, we’re constantly benchmarking all the latest OCR models in our LlamaParse pipeline. If you’re interested in parsing complex docs come check it out here! https://cloud.llamaindex.ai/
This past week the entire @llama_index team got together in Mexico City 🇲🇽 We reaffirmed our focus: we’re all in on building AI agents to solve document processing and workflows 📑🤖 We are not building simple LLM framework abstractions. We are building incredibly deep OCR technology across multiple price points to unlock content from any unstructured document. Most importantly, we ate a ton of tacos 🌮🌮🌮🌮 We have a ton of releases slated from now until EOY, stay tuned! Also we’re hiring: https://www.llamaindex.ai/careers Until next time CDMX :)
We’ve created a specialized agent tuned for row-level table extraction 🧩🤖 A lot of document workflows involve converting a complex table in a .pdf/.docx file into an Excel spreadsheet. Oftentimes this work is done manually. We created a mode in LlamaExtract that only deals with table extraction. Simply define the schema you want for each row, and the agent will extract out all rows from every table corresponding to this schema! It can not only populate the row context, but also the global table context. Check it out in LlamaCloud: https://cloud.llamaindex.ai/
RT LlamaIndex 🦙 Chart OCR just got a major upgrade with our new experimental "agentic chart parsing" feature in LlamaParse 📈🧪 Most LLMs struggle with converting charts to precise numerical data, so we've created an experimental a system that follows contours in line charts and extracts values. Automate chart analysis without spending hours manually correcting extracted values. Try it now in LlamaParse: https://cloud.llamaindex.ai/?utm_source=socials&utm_medium=li_social
One of the biggest use cases for agentic document automation is insurance underwriting ✍️ Underwriting depends on processing *massive* volumes of unstructured documents, from medical reports, scanned forms, and way more. It's also historically been a massively manual process. We're super excited to feature this case study with Pathwork AI - Pathwork is hyperfocused on building underwriting agents for life insurance. They're able to use LlamaCloud as a core module in order to process the massive volume of docs, from medical documentation to carrier guidelines. Check it out: https://www.llamaindex.ai/customers/pathwork-automates-information-extraction-from-medical-records-and-underwriting-guidelines-with?utm_source=socials&utm_medium=li_social LlamaCloud: https://cloud.llamaindex.ai/LlamaIndex 🦙: See how @pathwork scaled their life insurance document processing from 5,000 to 40,000 pages per week using LlamaParse. 📄 Process complex medical records, lab results, and decades-old scanned PDFs with 8x improved throughput 🤖 Automatically extract and index carrier Link: https://x.com/llama_index/status/1988290671279829204
RT LlamaIndex 🦙 See how @pathwork scaled their life insurance document processing from 5,000 to 40,000 pages per week using LlamaParse. 📄 Process complex medical records, lab results, and decades-old scanned PDFs with 8x improved throughput 🤖 Automatically extract and index carrier underwriting guidelines to keep risk rules current ⚡ Replace fragile, manual pipelines with robust automation that handles everything from digital forms to 1970s faded scans 🎯 Free up engineering time from maintenance to focus on building new product features @pathwork's Case Underwriter, Knowledge Assistant, and Pre-App Manager products all rely on transforming unstructured insurance documentation into structured data for faster decision-making. By integrating LlamaParse, they eliminated bottlenecks that were directly limiting customer growth and built future-proof infrastructure that automatically improves over time. Read the full case study: https://www.llamaindex.ai/customers/pathwork-automates-information-extraction-from-medical-records-and-underwriting-guidelines-with?utm_source=socials&utm_medium=li_social
We've gotten super, super deep in the wonderful world of document OCR through the history of @llama_index - and we'd love to share it with you! 🌟 1. There's a lot of benefits to "traditional" methods of reading the PDF binary for fast, cheap parsing. 2. You can use LLMs in the loop for general reading order reconstruction. 3. VLMs are obviously useful and we've benchmarked every frontier model there is out there to give high quality results over the most complex pages within our pipeline. State-of-the-art document parsing is super important for building agentic automation over any set of docs, and we've invested in it for the past 2 years. Register here: https://landing.llamaindex.ai/beynd-ocr-how-ai-agents-parse-complex-docs?utm_source=socials&utm_medium=li_socialLlamaIndex 🦙: We probably shouldn't tell you how to build your own document parsing agents, but we will 😮. AI agents are transforming how we handle messy, real-world documents that break traditional OCR systems. Join our live webinar on December 4th at 9 AM PST where the LlamaParse team Link: https://x.com/llama_index/status/1986810928713855235
We're partnering with @browserbase, @braintrust, @modal for an awesome afterparty at re:invent - come join us!LlamaIndex 🦙: There are Vegas parties and there is Late Shift 🎉 Join us for an exclusive re:Invent afterparty that brings together the best minds in AI and tech for a night you won't forget. 🍸 Cocktails and disco balls at Diner Ross Steakhouse in The LINQ 🤖 Connect with the teams behind Link: https://x.com/llama_index/status/1988003781448266141
RT LlamaIndex 🦙 There are Vegas parties and there is Late Shift 🎉 Join us for an exclusive re:Invent afterparty that brings together the best minds in AI and tech for a night you won't forget. 🍸 Cocktails and disco balls at Diner Ross Steakhouse in The LINQ 🤖 Connect with the teams behind @browserbase, @braintrust, @modal_labs, and LlamaIndex 🌙 Late-night tech conversations when the conference sessions end 🎟️ Limited spots with approval-required registration We're teaming up with our friends at @browserbase, @usebraintrust, and @modal_labs to host the most fun you'll have all conference. After your evening sessions, meet us for cocktails, networking, and the kind of tech chatter that makes re:Invent legendary. RSVP now - spots are limited: https://luma.com/lateshift
For the first time in human history, you can: 1️⃣ Take a bucket of docs/PDFs 🪣📑 2️⃣ Make sense of it 3️⃣ Extract insights / search over it with super high accuracy with effectively 0 humans involved. This is a neat joint stack we copublished with @MongoDB, check it out! https://youtube.com/watch?v=5mEPkPtoNyYLlamaIndex 🦙: Last week, we teamed up with @MongoDB to break down one of the most persistent challenges in production AI systems: turning messy, real-world documents into reliable insights. Enterprise documents don't come in neat, uniform packages. Invoices, SEC filings, reports—they all have Link: https://x.com/llama_index/status/1986117341911130163
Build an AI agent to automate your finance team’s entire invoice/expense workflow! 🧾 @TuanaCelik has built a fantastic example that shows you how to construct an agentic workflow that can triage incoming emails + attachments, detect whether it’s an invoice or expense, and process it accordingly. It uses our core agentic classification / extraction capabilities under the hood in LlamaCloud, and is backed by @llama_index workflows. Check it out: https://github.com/run-llama/workflows-py/blob/main/examples/document_agents/finance_triage_agent.ipynbLlamaIndex 🦙: Here's a common scenario: Your finance team gets emails all day with invoices from partners and expense reports from employees. Each one needs different handling. Invoices need acknowledgment and payment scheduling. Expenses need budget validation before approval etc. In this Link: https://x.com/llama_index/status/1986476949687140503
RT LlamaIndex 🦙 We probably shouldn't tell you how to build your own document parsing agents, but we will 😮. AI agents are transforming how we handle messy, real-world documents that break traditional OCR systems. Join our live webinar on December 4th at 9 AM PST where the LlamaParse team reveals industry secrets for parsing complex documents: 📋 Blueprint for building next-generation document parsing workflows using agents instead of OCR alone 🔧 Practical strategies for handling handwriting, rotated scans, nested tables, and visually dense layouts 🤖 Latest LlamaCloud capabilities showing how vision language models automate extraction from previously unparseable PDFs, forms, and images ⚡ When to apply each component in your parsing pipeline and why it matters We'll show you how to move beyond simple text extraction to actually automate understanding of documents with multi-column layouts, embedded charts, skewed scans, and tables within tables. Register now: https://landing.llamaindex.ai/beynd-ocr-how-ai-agents-parse-complex-docs?utm_source=socials&utm_medium=li_social
Our new bounding box approach in LlamaParse gives you clean bounding boxes while preserving clean reading order of the text through agentic reconstruction. The issue with traditional parsing methods is that the quality of the output is directly dependent on the layout detector - if the predicted boxes are wrong / in the wrong sequence, then your output is garbled. Here we use LLMs to reconstruct the entire semantic flow of the text, but still allow bounding box processing in parallel for additional metadata! Now available in LlamaParse: https://cloud.llamaindex.ai/
RT LlamaIndex 🦙 Here's a common scenario: Your finance team gets emails all day with invoices from partners and expense reports from employees. Each one needs different handling. Invoices need acknowledgment and payment scheduling. Expenses need budget validation before approval etc. In this example we build an agent that automatically triages incoming emails with attachments, extracts the right information, and takes appropriate action. Our approach uses three of our tools working together: 1️⃣ LlamaClassify handles the first decision point. It looks at each attachment and determines: is this an invoice that needs to be paid out to a partner, or an expense that needs reimbursement? It also provides reasoning for the decision. 2️⃣ LlamaExtract does the heavy lifting on data extraction. We create two specialized agents with different schemas for invoices vs expenses. 3️⃣ Agent Workflows orchestrates the entire process. It connects classification to extraction to business logic: in this case, checking expenses against a budget threshold and generating appropriate email responses via LLM. Classify incoming documents → extract relevant data → apply business rules → take action. Need to add a new document type? Add a classification rule and an extraction schema. Need different business logic? Modify the workflow steps. The components stay the same. Check out the full example: https://github.com/run-llama/workflows-py/blob/main/examples/document_agents/finance_triage_agent.ipynb
RT LlamaIndex 🦙 Last week, we teamed up with @MongoDB to break down one of the most persistent challenges in production AI systems: turning messy, real-world documents into reliable insights. Enterprise documents don't come in neat, uniform packages. Invoices, SEC filings, reports—they all have irregular layouts, embedded tables, images, and context that traditional text extraction just can't handle. In this session, we walked through a complete document processing workflow that works at scale: LlamaParse acts as an agentic parsing tool that understands document structure—not just text extraction. It handles complex layouts, preserves table formatting, and extracts images with context. It outputs clean markdown that LLMs can work with. The architecture is : S3 → LlamaParse → MongoDB Atlas → LLM. The recording is up now: https://www.youtube.com/watch?v=5mEPkPtoNyY
RT LlamaIndex 🦙 MavenBio transformed complex scientific visuals in biopharma documents into searchable, analyzable intelligence using LlamaParse. Before LlamaParse, MavenBio's AI platform could process text-heavy documents but missed critical insights locked in charts, figures, and conference posters that drive real biopharma decisions. 🔬 Visual content parsing: Conference posters, regulatory filings, and scientific publications with complex diagrams now become fully searchable 📊 10x-20x faster workflows: Users can run comparative trial assessments and opportunity prioritization with unprecedented speed and depth 🎯 Enhanced accuracy: Visual context integration improved the precision of structured analyses across their platform ⚡ Engineering focus: Team reallocated resources from building parsing infrastructure to core product innovation "LlamaParse bridges the gap between static visual data and structured language," says @bernardffaucher, Founding Senior Backend Engineer at MavenBio. The webhook-based asynchronous processing scaled their throughput while maintaining low latency across their always-on ingestion pipeline. Read the full case study: https://www.llamaindex.ai/customers/maven-bio-turns-the-unstructured-world-of-complex-scientific-visuals-into-intelligence-with?utm_source=socials&utm_medium=li_social
Haiku 4.5 is better than GPT-5 at document OCR over tables 📋 Better reasoning doesn’t correlate to visual understanding 💡. I fed the NYC MTA timetable as screenshots into both GPT-5 and Haiku 4.5. - (Left) GPT-5 ignores the spaces between table values - (Right) Haiku almost perfectly reconstructs the table including spaces in between. The extra columns don’t materially impact the correctness of the results. Haiku is shaping up to be a great lightweight contender for document parsing. You can play with it and other models within LlamaCloud! LlamaCloud: https://cloud.llamaindex.ai/
RT LlamaIndex 🦙 Augment your LlamaIndex agent workflows with memory and persistent states: Check out @itsclelia's talk at @qdrant_engine Vector Space Day to learn how to build context-rich AI systems leveraging vector search and workflow engineering. Take a look at the YT video: https://youtu.be/CDyFukgpayY Learn more about LlamaIndex agent workflows: https://developers.llamaindex.ai/python/llamaagents/overview?utm_source=twitter&utm_medium=li_social
