内容历史
least clickbait slop post
China has killed the entire vector database industry. They open-sourced TencentDB Agent Memory. It gives any AI agent long-term memory that runs 100% locally. No Pinecone. No cloud APIs. No repeating yourself every session. - 61% fewer tokens - PersonaMem accuracy: 48% → 76%
RT Imaan Sultan 🇵🇰🇸🇦 i’ve tried 150+ coffee shops in SF. these are the only ones worth your morning (grew up in saudi. we don’t drink, so coffee became a core part of my personality) what’s your go-to? i’ll go try it this week :)
The definitive guide to SF coffee shops: Which ones did I miss or underrate? Fully expect to get roasted for this 😉☕️
We hosted an awesome black tie gala last Friday to close out SF tech week huge shoutout to everyone involved with organizing! @ValyuOfficial @MongoDB @daytonaio @arceuslegal @obvious @thelinqapp @lemma_ @Tiny_Fish + ofc @imaanxsultan from our side
Three events, one award, a very respectable 4th place in hot sauce. Our SF Tech Week: → AI Demo Night w/ @ClickHouse + @artie_labs: @imaanxsultan won best demo → @matissemlk took 4th in the Blazel hot sauce competition → Co-hosted the biggest night of SF Tech Week with
RT LlamaIndex 🦙 Three events, one award, a very respectable 4th place in hot sauce. Our SF Tech Week: → AI Demo Night w/ @ClickHouse + @artie_labs: @imaanxsultan won best demo → @matissemlk took 4th in the Blazel hot sauce competition → Co-hosted the biggest night of SF Tech Week with @SpaceX, @Lemma_, @Daytonaio, @Tiny_Fish, and more! Agent Economy Gala: 500+ founders + investors in suits and gowns Thanks to everyone who joined us for a very busy, but fun SF Tech Week :)
Unlimited OCR is a good example of a low-parameter OCR model that can process any document extremely quickly. It's good at reading order and layout, and decent over tables. It is slightly better than PaddleOCR on tables and slightly worse on layout. There are tradeoffs though - it ignores visual elements, formatting, and the most complex tables. It scores ~46.2% and ranks #115 on ParseBench. There's plenty of opportunities to improve this frontier, but Unlimited OCR is still a good milestone! Check out ParseBench: https://www.parsebench.ai/
🚨 CHINA JUST OPEN SOURCED AN OCR THAT EATS 100 PAGE PDFS IN ONE PASS It’s called Unlimited OCR. Only 3B params. Runs locally. Every other OCR tool chops your doc into pages and loses the thread. This one reads the whole thing in one continuous pass. → One shot full context
View quoted post
Mistral OCR is on OpenDocRouter! For its price, it is decent over tables and reading order. It's also blazing fast at p50 1.8s per page We're constantly bringing in new models since OpenDocRouter launched last week: https://www.opendocrouter.ai/
Today we're announcing OpenDocRouter: every model for document parsing under one API. There are ~4,000 OCR models on Hugging Face, and the frontier labs ship a new one nearly every month. Whatever's best for your docs today won't be by Q1. So stop asking which model to use for
View quoted post
RT LlamaIndex 🦙 What happens when a table contains tables? Usually, a mess. With LlamaParse, it comes out clean. Doc of the Week: Micron’s latest earnings deck. Even with two business units and same row names, LlamaParse keeps all 18 values under the right header. Try it out on your own earnings deck: https://www.llamaindex.ai/llamaparse
i'm pulling out my suit for this Huge shoutout to @imaanxsultan and everyone else for organizing!
The Agent Economy Gala just maxed out at 400 guests Here’s a breakdown: - 203 founders ($1B+ raised combined) - 67 founding operators - 58 investors - 23 creators - 295 guests building agent applications or infrastructure Founders @llama_index, @daytonaio, @Tiny_Fish,
RT Imaan Sultan 🇵🇰🇸🇦 The Agent Economy Gala just maxed out at 400 guests Here’s a breakdown: - 203 founders ($1B+ raised combined) - 67 founding operators - 58 investors - 23 creators - 295 guests building agent applications or infrastructure Founders @llama_index, @daytonaio, @Tiny_Fish, @obvious, @thelinqapp Leads at @FireworksAI_HQ, @mercor, @mintlify, @WisprFlow High levels at OpenAI, Google, Meta, NVIDIA, Microsoft, Netflix, Tesla, and SpaceXAI But honestly none of that matters. What matters is there's no fluff. Prioritizing a quality experience, not turnout, because turnout doesn't matter when it's not relevant to the space EVERY person in this room is a main player, actively building out the agent ecosystem If you still want in on the biggest event of TECH WEEK by a16z, reach out to anyone at Lemma, Linq, LlamaIndex, TinyFish, or Obvious See you tomorrow 🥂
fleet week makes me patriotic af
Love WFH during Fleet Week in the nontechnical north. Finally a break from sirens, dogs, and handymen. Just weapons of war flying low over my apartment while I try to shout over them to sell AI models for doc parsing on Zoom.
LightOn OCR-3 is now available on OpenDocRouter, priced at $0.28 / 1m input and $1.40 / 1m output (it is ~$3.19 / 1k pages on ParseBench). We benchmarked it on ParseBench. It is on the pareto frontier for open-weight OCR models, with similar performance as Gemini 3.8 flash low at ~45% lower price. It is decent at tables, workable for charts, and quite good at grounding. Try it out on OpenDocRouter: https://www.opendocrouter.ai/models For the full list of benchmark results, check out ParseBench: https://www.parsebench.ai/
World-leading document recognition. Beyond text extraction. Introducing LightOnOCR-3, our new family of state-of-the-art AI models for document intelligence. Building on LightOnOCR-2, it recognizes text and handwriting, describes images, extracts data from charts and preserves
RT Logan Markewich OpenDocRouter serves LightOnOCR-3-1B now! When benched in the app, we found that our attribution engine produced better bounding boxes than the model itself, leading to higher grounding scores! 🚀 Try it out today: https://opendocrouter.ai
World-leading document recognition. Beyond text extraction. Introducing LightOnOCR-3, our new family of state-of-the-art AI models for document intelligence. Building on LightOnOCR-2, it recognizes text and handwriting, describes images, extracts data from charts and preserves
Markdown has become a universal representation of information between humans and agents. It preserves core primitives needed to express structure (headings, lists), allows agents to read the information in a semantically interpretable manner, and allows for complex elements like tables and charts. You can natively embed HTML as well. But since most unstructured data is not natively within markdown (yet), the issue becomes the translation layer. We've created best models to convert any unstructured document container into markdown. Check out our blog! https://www.llamaindex.ai/blog/markdown-is-all-you-need
Markdown is all you need. (Mostly.) A parser can get every word on the page right and still lose which column a number belongs to. Then your model has to guess. Markdown keeps headings, lists, and tables intact, and it stays readable when you're debugging a bad answer. For
Claude Haiku 5.5 is now available on OpenDocRouter as one of the cheapest models for document parsing. It is roughly $1.2 / 1k pages and does quite well on tables and text reading order compared to other models at this price point. It struggles a bit on charts, semantic formatting, and bounding boxes. Check out OpenDocRouter, which we just launched today! https://www.opendocrouter.ai/models
Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released. On average, it costs around 75% less to run than Claude Haiku 4.5.
View quoted post
RT Logan Markewich The power of OpenDocRouter is already working: Anthropic just dropped Haiku 5.5 and we already have initial benchmark results. I see pretty comparable results to Luna overal. Slightly better at tables, but worse at chart understanding. Seems like a decent choice for cost-concerned parsing workloads! Try it out at http://opendocrouter.ai
RT Murtaza Khomusi It's SF tech week and here's how we celebrated at @llama_index 🦙: 1. @imaanxsultan pitched our product in a room full of founders and won best pitch of the night 2. @matissemlk placed third in a hot wings eating contest until he was sick, also in a room of founders Peak alpha energy.
OCR is the perfect example of a use case that has been dominated by legacy/brittle systems, and can be solved both accurately and cheaply by applying the right amount of agentic intelligence. A properly tuned agentic OCR needs to dynamically apply extra compute to complex elements, review and correct failures, and create the right semantic meaning throughout the page. It's a task that frontier models are both overengineered for in terms of cost and latency, and struggle at over a long tail of complex edge cases. Check out this blog post by @LoganMarkewich below: https://www.llamaindex.ai/blog/ocr-is-dead-long-live-agentic-ocr
OCR is dead 🪦 Long live agentic OCR! Traditional OCR makes one pass and hands back whatever text it got. Tables get flattened, charts disappear, and multi-column layouts come out scrambled. Nothing checks the output. Agentic OCR treats parsing as a loop instead: ✅
RT LlamaIndex 🦙 The biggest night of SF Tech Week is black tie. 120+ agent founders, 50 investors, and live jazz at the Redwood Room on Friday for the Agent Economy Gala. Invite only. We're excited to be partnering with @SpaceX, @Daytonaio, @lemma_, @Tiny_Fish, @thelinqapp, @Obvious & more. Link to request an invite in below or comment what you're building!
RT LlamaIndex 🦙 OCR is dead 🪦 Long live agentic OCR! Traditional OCR makes one pass and hands back whatever text it got. Tables get flattened, charts disappear, and multi-column layouts come out scrambled. Nothing checks the output. Agentic OCR treats parsing as a loop instead: ✅ layout-aware reading order ✅ smart routing of hard elements to the right model ✅ multi-pass verification and self-correction ✅ multimodal parsing of charts, images, and complex tables Logan Markewich, LlamaIndex's Head of Open Source, wrote about what that shift looks like in practice and where it's still struggling. Link to the breakdown in the comments below.
RT Imaan Sultan 🇵🇰🇸🇦 it's sf tech week and this year is the first time i'm hosting, which means i get to see the chaos from the other side 🌟 a few things i've learned from past years: 1/ paying to attend or pitch at events is a scam 2/ have a reason for every event on your cal. there are thousands this week. know why you're going and what you'd want to walk out with. "free snacks" is a valid reason btw 3/ take muni or walk. ubers are about to get wildly expensive 4/ ask for people's twitters. once you've added 3,000 people on linkedin it's genuinely impossible to remember who you actually talked to 5/ come to our founders + investors gala on oct 9. rsvp link is in the comments!! welcome to sf and lmk if you need any help finding cool events to go to :D
RT AI Engineer 🔜 NYC 🗽 Can your agent read the documents your business runs on? @jerryjliu0 on parsing, search & source citations. World’s Fair 2026 recording: https://www.youtube.com/watch?v=RQi7x-navxU Keep learning with us in NYC, Oct 12–14. Tickets: https://ai.engineer/nyc/2026
RT LlamaIndex 🦙 The BEA's new Q2 GDP-by-state was released, but the key data is entirely in an infographic, without a text-layer. LlamaParse's agentic parsing tier understands documents like this by default, and faithfully records data, turning the page into a 52-row table. Source: https://www.bea.gov/sites/default/files/2026-09/gdp2q26-3rd.pdf
RT LlamaIndex 🦙 How many people show up on a Tuesday night to talk about document processing for AI agents? 🤔 In New York, enough to fill the room and start a waitlist. In San Francisco, almost 600. We were in both cities on the same night to talk about one problem: agents don't read documents the way people do. Look at an invoice and you can tell the monthly rate from the total in a second. An agent often gets the words without the layout, so it has to guess. In New York, we partnered with @usekernel on @braintrust's Agent Builders Night. In San Francisco, we partnered with @WorkOS on @daytonaio's AI Builders event. Thanks to them for having us, and to Abrar and Yong from our team for speaking!
RT Daytona Last night, we hosted our monthly @daytonaio AI Builders Demo Night at the @WorkOS office in San Francisco! Great demos, conversations, and energy from the AI builder community. Huge thanks to @llama_index for partnering with us, @WorkOS for hosting us, and everyone who joined and shared what they’re building. See you soon at the next one!
we comprehensively benchmarked gpt-6.1 sol on document OCR tasks. compared to gpt-6 sol from a week ago (!), we see a sizable increase in table parsing capabilities and reading order, which is pleasantly surprising. gpt-6.1 sol has similar table parsing as gpt-6 astra. these are great results. frontier models are still typically an order of magnitude more expensive than cost-effective document parsing solutions, but this also means that there's a lot of alpha in improving capabilities at the "premium" end (>1c per page) of the spectrum too check out ParseBench: https://www.parsebench.ai/ for a comprehensive set of document parsing capabilities, check out llamaparse: https://cloud.llamaindex.ai/
RT vincent sunn chen awesome discussion with @jerryjliu0 and a great group! one theme: data must be developed to be useful users/enterprises have a trove of interesting & valuable raw data (e.g. patient records, historical images/documents), but making that ready for evaluation and training requires a new form of engineering & research. this requires new interfaces & programming models involving subject matter experts to make the resulting data & environments sufficiently difficult/fair/complex!
Yesterday I hosted a fun dinner conversation with @vincentsunnchen from @SnorkelAI on evals and RL environments. The "data and RL env" companies (like Snorkel) have seen massive growth in the past few years. There's been an explosion of interest in evals. At the same time,
Yesterday I hosted a fun dinner conversation with @vincentsunnchen from @SnorkelAI on evals and RL environments. The "data and RL env" companies (like Snorkel) have seen massive growth in the past few years. There's been an explosion of interest in evals. At the same time, models are ripping through benchmarks with each new release. We talked about the evals everyone is defining, what evals are still left unsolved, what’s left up to frontier models vs. intelligence that you own, and more: - A big challenge for building RL environments is “fairness” - when the model fails on a given environment, can you attribute it to the input, harness, or reward model? - Building proper rewards is hard. Some tasks are not easily quantifiable. You also want to discourage reward hacking. At the same time, you don’t want to be too prescriptive with intermediate rewards. - Long horizon evals are still extremely hard, some business processes can take up to weeks or months before the final outcome - Most regulated industries still need human in the loop to guarantee ~100% accuracy, “80%” accuracy is not good enough - As models get more intelligent, there will be a barbell of boutique data vendors (e.g. any SMB) any scaled up data providers. - Models still exhibit “jagged intelligence” where they still fail on a long tail of edge cases. - There might always be opportunities to gather unique data for a given task and posttrain models for lower cost and higher accuracy. This marks #003 in our founder dinner series. What topic should we discuss next? Let us know your thoughts below!
We built state-of-the-art models for reading forms 📋 Form documents have the following properties that trip up VLMs: ✅ They carry much more structure than can be represented in standard markdown. You need consistent types for checkboxes, textboxes, labels, signature fields. ✅ They can be extremely complicated (forms can be scanned, there can handwriting scribbles, some forms are dense with ~100+ fields) but accuracy requirements need to be close to 100% ✅ Any form parser requires accurate grounding and attribution. Not only should you extract the values, but you should also be able to precisely locate where each value came from in the source doc ✅ Any form needs to be not only accurate, but cheap/fast We've done a deep-dive into what it takes to build a form parser in this blog post: https://www.llamaindex.ai/blog/why-vlms-can-t-read-forms If you want to try out our form models, check out LlamaParse: https://cloud.llamaindex.ai/
Parsing forms still trips up the latest frontier VLMs. A form isn't text on a page. It's a set of fields, grouped into sections, each tied to a specific box. That's why forms need purpose-built parsing, not a bigger general model: ✅️ Detect every field, not just the obvious
if this takes off, then i was 4 years too early 😭 https://x.com/jerryjliu0/status/1590192512639332353 only OGs remember gpt tree index
The entire RAG industry is about to get cooked. Researchers developed a new RAG approach that bypasses almost everything traditional RAG depends on. - No vector DB - No data embeddings - No chunking - No similarity search It's called PageIndex. Instead of splitting your
RT LlamaIndex 🦙 Parsing forms still trips up the latest frontier VLMs. A form isn't text on a page. It's a set of fields, grouped into sections, each tied to a specific box. That's why forms need purpose-built parsing, not a bigger general model: ✅️ Detect every field, not just the obvious ones ✅️ Keep the hierarchy of sections and fields ✅️ Tie every value to the exact box it came from ✅️ Understand handwriting and checkmarks Our latest blogpost breaks down where VLMs fail on real W-2s, 1040s, W-9s and scanned W-4s, and a custom cookbook for LlamaParse to handle them at a fraction of the cost 👇 https://www.llamaindex.ai/blog/why-vlms-can-t-read-forms
We comprehensively evaluated 16 recent frontier VLMs - including Opus 5.5 and GPT-6 Sol/Luna - on whether higher effort led to better document parsing performance. Higher effort typically leads to improvements on other benchmarks (coding, knowledge work), but up until recently it wasn't obvious that this natively improved capabilities for reading PDFs. Results: ✅ Out of the frontier models, Opus 5.5 has the best performance relative to its price. It's especially good at parsing tables. ✅ Astra is also quite good, but starts at a more expensive price than Opus. ✅ GPT-6 Luna is more compelling at the cheaper end of doc parsing If you're parsing documents at scale, you'll still want a dedicated OCR solution like LlamaParse (https://cloud.llamaindex.ai/) that has better performance at a cheaper price. But if you're parsing docs "in the agent loop" within an app like Codex/Claude code, and you're too lazy to integrate a dedicated solution, then Opus 5.5 is currently the leader. Full results on ParseBench: https://www.parsebench.ai/
blank cells in tables can mess up agent decision making many document OCR tools struggle with reading multiple blank cells in a document and cause values to be shifted from the original column and rows. this causes the wrong numerical interpretation downstream. This is a nice illustration of what happens! If you have complex documents where table accuracy is extremely important, come check out LlamaParse: http://cloud.llamaindex.ai/signup
A blank cell can change the meaning of a forecast. Four times a year, the Fed's 18 top policymakers each put their forecasts on paper: where growth, jobs, inflation, and interest rates are headed. This is September 2026's edition, home of the "dot plot" that markets treat as the
RT LlamaIndex 🦙 A blank cell can change the meaning of a forecast. Four times a year, the Fed's 18 top policymakers each put their forecasts on paper: where growth, jobs, inflation, and interest rates are headed. This is September 2026's edition, home of the "dot plot" that markets treat as the Fed tipping its hand. This release is the closest thing to the Fed saying what it plans to do. Most analysts will want to throw this documents to an AI agent, but this messy doc is dense: full of complex tables and charts that hold valuable context. All things that frequently trip up raw LLM APIs. The Fed’s September 2026 projections table includes a 2029 column, but its June comparison row leaves that cell empty. This is a common but silent failure point for document parsers. We parsed page 2 with LlamaParse and checked the displayed GDP median excerpt against the original PDF. All nine numbers matched and the data stays aligned in the returned HTML. Try LlamaParse on a table where headers and missing cells matter. https://cloud.llamaindex.ai/signup Source: https://www.federalreserve.gov/monetarypolicy/files/fomcprojtabl20260916.pdf
come join us at SF @Techweek_ !
building in the agent space? this is your room at SF @Techweek_. think modern gatsby: black tie, red carpet, cocktails, live jazz & art deco glamour. 300+ founders, founding operators & investors across agent products, infra & dev tools. 5 cohosts sending exclusive invites:
RT Imaan Sultan 🇵🇰🇸🇦 sorry to every other sf tech week host in advance, but we may have gotten carried away planning our event... we're about to host the biggest and most lavish black-tie gala sf tech has ever seen. live jazz. drinks under the art deco lights of the famous redwood room. the best founders and investors from across the AI agent economy will be dressed like its the roaring 20s. oct. 9. rsvp below 🥂
Opus 5.5 is the best frontier model for parsing tables in PDFs. We ran it through ParseBench, and it scored 93.9%, a 7%+ increase over Opus 5, and beating Fable/Gemini/Astra. It's comparable to LlamaParse Agentic Plus in table quality, though it still struggles on charts, formatting, and layout. It's also 5.8c a page, which is too expensive to be a production OCR solution. Full results on ParseBench: https://www.parsebench.ai/ If you want Opus 5.5-level table quality, consistent parse quality + metadata for all edge cases, at the fraction of the cost, come check out LlamaParse. We have volume discounts for all modes and can help you set up customized routing between Agentic Plus, Agentic, and cost-effective: https://cloud.llamaindex.ai/
