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 Claude. 🛡️ 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 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://lnkd.in/e4cMNN2Z Find the code on GitHub: https://lnkd.in/eaMBdfdv

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 line values. 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! 

LlamaSheets is our new way to handle complex, messy spreadsheets that come as many sheets disguised as one, multiple regions that provide different sets of information, and much more. Check out this example of a (generated, fake) company budget sheet. It actually has 4 sub-sheets, each containing multiple regions. LlamaSheets identifies each sub-region, creates summaries about what information they provide, and returns all of this informaiton as a parquet file! Try it out and let us know what you think while it's in public beta! Get started here: https://lnkd.in/e_Y9refB

"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://lnkd.in/eTGZJB-9

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://lnkd.in/e6DkkFdc Docs: https://lnkd.in/eU5rZFVS

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 while maximizing throughput 📊 Get detailed progress tracking and summary statistics for batch operations This is perfect for processing large document collections, research papers, or any scenario where you need to parse dozens or hundreds of PDFs quickly and reliably. Full tutorial with working code examples: https://lnkd.in/eFSDB7R8

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 high-quality research from top conferences (NeurIPS, CVPR, ACL, etc.) to 0-1 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://lnkd.in/g_gxQTvv