Jerry Liu(@jerryjliu0)

One of the main issues with AI document parsing is that because no solution is 100% accuracy, it's h...

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 18:22:45 GMT

One of the main issues with AI document parsing is that because no solution is 100% accuracy, it's hard to tell if a given page is parsed incorrectly: for instance a misaligned table, missing/hallucinated values, or messy scans 

over the past few months we've invested a lot in models that provide more calibrated confidence scores - it will give you a confidence level that the page is parsed correctly. it is correlated with the parsing mode and how complex the source page is. 

this allows you to bake in human-in-the-loop mechanisms where you can review and correct the outputs, or trigger automated fallback logic, for highly sensitive paperwork-heavy processes. 

come check out our high-effort scores! developers.llamaindex.ai/llamaparse/par…

Come sign up to LlamaParse here: cloud.llamaindex.ai

LlamaIndex 🦙@llama_index

confidence scores in LlamaParse just got an upgrade 🦸‍♀️ 

our new high-effort mode provides granular page-level scores with text explanations to help you intimately understand parsing quality of your documents. we even refer back to the original document for an extra check when generating the score. 

use high-effort only when you need it, at 5 additional credits per page.

try it on your docs:developers.llamaindex.ai/llamaparse/par…h

🔗 View Quoted Tweet

💬8🔄3❤️20👀2015📊10

⚡ Powered by xgo.ing

agreed, one of the biggest bottlenecks right now is my own ability to internalize the information th...

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 01:32:56 GMT

agreed, one of the biggest bottlenecks right now is my own ability to internalize the information that these models convey

the latest models (astra, fable 5.1) are either way too verbose, or try too hard to pack information density to be concise

it comes up with ai-coded jargon that's in to deciphering a math proof

Wes Winder@weswinder

claude and codex still use way too many words

it's exhausting just reading all the word vomit

there has to be a better way

🔗 View Quoted Tweet

💬15🔄0❤️27👀4832📊16

⚡ Powered by xgo.ing

who needs whitepapers when you have this LOL full credits to @hexapode if u want the full copy d...

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 01:13:39 GMT

who needs whitepapers when you have this LOL 

full credits to @hexapode 

if u want the full copy dm me or him

💬0🔄0❤️3👀860📊1

⚡ Powered by xgo.ing

if you'd like to understand why pdf parsing is important, we made a manga on it 💕 so kawaii 💕

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 01:13:38 GMT

if you'd like to understand why pdf parsing is important, we made a manga on it

💕 so kawaii 💕

💬7🔄1❤️15👀2525📊10

⚡ Powered by xgo.ing

Astra is one of the best models we've tested at parsing complex documents with tables, at 93.2% on P...

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 00:52:05 GMT

Astra is one of the best models we've tested at parsing complex documents with tables, at 93.2% on ParseBench.

So it seems well suited for ad-hoc analysis over table-heavy documents e.g. in finance.

To be fair, it's also at 10c per page so not suitable for document-heavy OCR workloads. Fable 5.1 still performs a better at charts.

Check out the full set of results on ParseBench: parsebench.ai

Jerry Liu@jerryjliu0

We benchmarked GPT-6 Astra on hard document parsing and extraction tasks.

It is the best frontier model for one-shotting document extraction over short and medium-sized documents:
📊  97.2% over short documents, achieving a new SOTA on our benchmark
📊 90.6% over medium documents

But a few caveats:
* it's pricy! The average price per page is 11c which is 10x  our cost-effective document extraction solution
* one-shot performance over long documents is inherently tricky, it only achieves 31.7%
* the native OCR capabilities of Astra isn't that much better than other frontier models (e.g. Fable 5.1, gpt-5.6-sol). It does very well on table understanding, but isn't great on charts, layout, and semantic formatting.

Full results for document extraction on ExtractBenchextractbench.aic3

🔗 View Quoted Tweet

💬12🔄10❤️44👀5766📊17

⚡ Powered by xgo.ing

Lightning-fast extraction over any document. ICYMI, we launched Extract Turbo last week, that lets...

Jerry Liu (@jerryjliu0) — Fri, 11 Sep 2026 00:45:57 GMT

Lightning-fast extraction over any document. 

ICYMI, we launched Extract Turbo last week, that lets you process documents from 1 page to dozens of pages with lightning-fast speed. Come check it out! 

Check out the video sequence below on equipment orders. Get all the line items with bounding boxes/citations in seconds.

(disclaimer: astra sped up the video a bit so it's not quite 1x, but it should still complete in seconds!) 

cloud.llamaindex.ai

Your browser does not support the video tag.

🔗 View on Twitter

LlamaIndex 🦙@llama_index

Document extraction just got a new gear. ⚡️

Introducing Turbo mode for Extract, our fastest way yet to pull structured data from documents.

It runs roughly 4× faster than our Cost Effective Tier at comparable accuracy, with a median latency of 3.7 seconds a page. Turbo processes pages in parallel, so latency stays nearly flat as document size grows. 

Available in beta today.

Read the full breakdown → llamaindex.ai/blog/introduci…

🔗 View Quoted Tweet

💬0🔄0❤️15👀3447📊3

⚡ Powered by xgo.ing

We've trained new models for checkbox parsing ☑️ This lets you convert checkboxes of any size and s...

Jerry Liu (@jerryjliu0) — Thu, 10 Sep 2026 15:36:04 GMT

We've trained new models for checkbox parsing ☑️

This lets you convert checkboxes of any size and shape, filled or unfilled, into structured JSON that's suitable for programmatic or agentic use.

Use it on insurance applications, tax forms, compliance questionnaires, KYC forms, and more. 

Check it out in our forms mode in LlamaParse! cloud.llamaindex.ai

Your browser does not support the video tag.

🔗 View on Twitter

💬11🔄1❤️24👀2746📊16

⚡ Powered by xgo.ing

Our agentic extraction harness can extract 54+ fields from an insurance policy with 100% accuracy ✅ ...

Jerry Liu (@jerryjliu0) — Wed, 09 Sep 2026 23:20:17 GMT

Our agentic extraction harness can extract 54+ fields from an insurance policy with 100% accuracy
✅ coverage limits 
✅ nested sublimits
✅ explicit noncoverage

This includes a full result review UI where you can audit every output, see which values have lower confidence scores, and see the source bounding box for every single field. 

Check out the video below, using the "Agentic Plus" mode in our Extract API. It's a great option for extracting a massive amount of output tokens from any-sized document, whether it's 10 pages or 1k pages.

If you're interested in giving it a spin, come check out LlamaParse: cloud.llamaindex.ai

Your browser does not support the video tag.

🔗 View on Twitter

💬8🔄7❤️42👀3957📊19

⚡ Powered by xgo.ing

Astra is awesome, but if you're using it specifically to parse and extract a massive volume of docum...

Jerry Liu (@jerryjliu0) — Wed, 09 Sep 2026 17:24:16 GMT

Astra is awesome, but if you're using it specifically to parse and extract a massive volume of documents, you will break the bank (10c+ per page on parsing and extraction!) 

We've added a native LlamaParse connector to ChatGPT that gives the model access to high-quality document OCR and extraction at the pareto frontier of accuracy and cost.

It contains native access to our parse and extract endpoints. Upload your most complex documents, and you can do the following: 
* digitalize it into extremely accurate context for any QA task
* perform large-scale extraction from a set of documents with accurate grounding and citations

Offload the higher-level reasoning to Astra, leave specialized doc parsing and extraction to us. 

ChatGPT plugin: chatgpt.com/plugins

Come sign up here: cloud.llamaindex.ai

Your browser does not support the video tag.

🔗 View on Twitter

LlamaIndex 🦙@llama_index

Last week we brought LlamaParse to Claude. Now it’s available in ChatGPT, too 🦙 

ChatGPT can already read PDFs. But document work gets harder when the information you need is buried in messy scans, dense tables, spreadsheets, and charts. Strike the right balance between burning tokens and accuracy, with a specialized API for document understanding.

Turn complex documents into clean, structured context and let Astra run its magic. Use our connector to:

✅ Parse documents into clean Markdown, JSON, or HTML. 
✅ Extract specific fields into a schema you define.
✅ Search across document collections to find the information you need. 
✅ Classify files and split documents into logical sections.

Live in the ChatGPT plugin directory now! →chatgpt.com/plugins/plugin…5

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬13🔄14❤️89👀10565📊32

⚡ Powered by xgo.ing

it is 2am ET and they're in the beginning of the 4th set, holy

Jerry Liu (@jerryjliu0) — Wed, 09 Sep 2026 05:59:03 GMT

it is 2am ET and they're in the beginning of the 4th set, holy

US Open Tennis@usopen

This. Point!

And it's only the first set!

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬4🔄0❤️11👀5305📊5

⚡ Powered by xgo.ing

we're giving away arsicault, $3k in llamaparse credits, and a limited edition swag box 🎁 if you're...

Jerry Liu (@jerryjliu0) — Wed, 09 Sep 2026 03:16:06 GMT

we're giving away arsicault, $3k in llamaparse credits, and a limited edition swag box 🎁

if you're building something that depends on processing high volumes of unstructured documents, let us know!

Imaan Sultan 🇵🇰🇸🇦@imaanxsultan

happy to announce i've become an arsicault chocolate almond croissant dealer!

(just for today and tomorrow.) 😁

doing a massive drop-off of pastries, $3000 (!!) in llamaparse credits for each team, and a ton of cool merch + goodies i designed for @llama_index . 

🌟 if you'd like me to stop by your office to deliver some sweet treats today or tmrw, PLS shoot me a message or comment :D 

would love to come say hi to your team and learn about what you're building 💌

🔗 View Quoted Tweet

💬4🔄0❤️10👀4094📊5

⚡ Powered by xgo.ing

i'm fucking dead

Jerry Liu (@jerryjliu0) — Tue, 08 Sep 2026 19:07:10 GMT

i'm fucking dead

cocktail peanut@cocktailpeanut

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬16🔄32❤️1081👀365796📊148

⚡ Powered by xgo.ing

Nathan's just prepping for The Rehearsal season 3

Jerry Liu (@jerryjliu0) — Mon, 07 Sep 2026 14:19:25 GMT

Nathan's just prepping for The Rehearsal season 3

A24@A24

YOU CAN SEE EVERYTHING, a documentary by Nathan Fielder and Lance Oppenheim. Featuring Elizabeth Holmes. Only in theaters this October.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬5🔄0❤️13👀7974📊5

⚡ Powered by xgo.ing

correction: the switching costs weren't enough to stop my desire to try out gpt-6 astra i ported ov...

Jerry Liu (@jerryjliu0) — Sun, 06 Sep 2026 22:10:25 GMT

correction: the switching costs weren't enough to stop my desire to try out gpt-6 astra

i ported over all my skills and conversational context to codex this weekend. planning to use it as my daily driver for now 

if the new model is sufficiently good enough, the switching costs for early adopters might continue to be 0 😅

Jerry Liu@jerryjliu0

i think there's a real opportunity for labs to bake in even higher switching costs between claude code/codex/grok bot etc. 

whenever a new model release comes out, i already feel an inertia to switch between apps because of my existing skills, routines, system instructions, project setup, and more. 

i do maintain an externally generated wiki / artifacts for each project, so technically i could point any app at the same wiki and get similar results, but this loses the nuances of conversation history. 

if memory is turned on, then these apps index and remember your context for subsequent sessions, alleviating the need to retype massive amounts of context for new sessions

the biggest pain point for any of these tools is figuring out how to efficiently feed it the right context every time. maybe the answer really is around building a good, self-improving context graph

🔗 View Quoted Tweet

💬25🔄1❤️41👀17346📊25

⚡ Powered by xgo.ing

everyone's building some video game with astra, and i got fomo, so i made gta 6 at home

Jerry Liu (@jerryjliu0) — Sun, 06 Sep 2026 20:48:29 GMT

everyone's building some video game with astra, and i got fomo, so i made gta 6 at home

Your browser does not support the video tag.

🔗 View on Twitter

💬10🔄2❤️56👀7751📊18

⚡ Powered by xgo.ing

Astra might finally help me automate what I've spent ~hundreds of hours doing since the start of thi...

Jerry Liu (@jerryjliu0) — Sun, 06 Sep 2026 16:00:25 GMT

Astra might finally help me automate what I've spent ~hundreds of hours doing since the start of this company: 

Making live product demos through Loom or Screen Studio 

Check out the video walkthrough on our new Agentic Plus mode for document extraction - where we're able to populate 60 fields worth of financials from a sample earnings release. 

This entire video was fully generated by Astra. With its default set of tools it can already login to my LlamaParse account and piece together a set of slides with static screenshots. I then gave it full access to computer use and Chrome, and let it try out Loom and Screen Studio on my desktop and record the current chrome session. 

In the meantime, if you do have use cases for gnarly extraction from extremely complex documents, come check out our LlamaParse product demo below 🙂cloud.llamaindex.aiA

Your browser does not support the video tag.

🔗 View on Twitter

💬28🔄13❤️187👀18560📊91

⚡ Powered by xgo.ing

We benchmarked GPT-6 Astra on hard document parsing and extraction tasks. It is the best frontier m...

Jerry Liu (@jerryjliu0) — Sat, 05 Sep 2026 15:25:59 GMT

We benchmarked GPT-6 Astra on hard document parsing and extraction tasks.

It is the best frontier model for one-shotting document extraction over short and medium-sized documents:
📊  97.2% over short documents, achieving a new SOTA on our benchmark
📊 90.6% over medium documents

But a few caveats:
* it's pricy! The average price per page is 11c which is 10x  our cost-effective document extraction solution
* one-shot performance over long documents is inherently tricky, it only achieves 31.7%
* the native OCR capabilities of Astra isn't that much better than other frontier models (e.g. Fable 5.1, gpt-5.6-sol). It does very well on table understanding, but isn't great on charts, layout, and semantic formatting.

Full results for document extraction on ExtractBenchextractbench.aic3

💬30🔄17❤️176👀22684📊66

⚡ Powered by xgo.ing

Full results for document parsing on ParseBench:

Jerry Liu (@jerryjliu0) — Sat, 05 Sep 2026 15:25:59 GMT

Full results for document parsing on ParseBench:

💬1🔄0❤️5👀1888📊2

⚡ Powered by xgo.ing

For reference, here were the scores from the original parsebench paper (note the jump in cost-effect...

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 22:28:44 GMT

For reference, here were the scores from the original parsebench paper (note the jump in cost-effective performance)

💬1🔄0❤️3👀1357📊2

⚡ Powered by xgo.ing

We've massively improved our document parsing capabilities across the board in the past ~3 months. ...

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 22:28:43 GMT

We've massively improved our document parsing capabilities across the board in the past ~3 months. 

Our LlamaParse cost-effective mode is one of the cheapest OCR solutions out there at <0.4c per page, and it has improved a lot on parsing since when we first put out the ParseBench paper a few months ago. 
📈 10%+ improvement on complex tables
📈 10%+ improvement on charts
📈 20%+ improvement on grounding 

We plot this along with other modes against a comprehensive set of models on ParseBench, our comprehensive benchmark for real-world docs.

This includes the latest VLMs like Gemini 3.8 Flash / Fable 5.1, along with new modes from specialized OCR providers.

Some providers don't have native chart parsing (or are relegated to other modes), so for the sake of completeness we also include the benchmark that excludes the chart dimension entirely. 

Our team is shipping like crazy and only accelerating 🚀

ParseBenparsebench.aip0OX
LlamaParcloud.llamaindex.aiTFz8

💬13🔄6❤️52👀5946📊25

⚡ Powered by xgo.ing

Unlike other players in the OCR space, we don't block our competitors from accessing our platform. ...

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 20:41:02 GMT

Unlike other players in the OCR space, we don't block our competitors from accessing our platform.

We're also fully committed towards open, reproducible benchmarks for document parsing and extraction - on ParseBench and ExtractBench respectively. It's important to give the industry a sense of how the pareto frontier of accuracy and cost evolves over time.

ParseBench and ExtractBench are fully open, maintained benchmarks that we're constantly adding every day (e.g. 3.8 flash, fable 5.1, r-1) 

Sometimes we make mistakes on benchmarking, and we take into account your feedback on how to fix and improve the results! 

You can access both benchmarks from the links below, which also give you access to Github/Huggingface and more. 

ParseBench: parsebench.ai
ExtractBench: extractbench.ai

In the meantime, if you're looking to ship doc OCR workloads to production, we'd be happy to help you benchmark LlamaParse against other solutions you're considering over your data and compare the results.

Sign up here:  cloud.llamaindex.ai

💬11🔄11❤️88👀10079📊33

⚡ Powered by xgo.ing

you should always time the release of your frontier model to be the day after another frontier model...

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 02:35:45 GMT

you should always time the release of your frontier model to be the day after another frontier model to mog them

💬13🔄0❤️46👀4865📊19

⚡ Powered by xgo.ing

me waiting for GPT-6 Astra access on my ChatGPT Pro subscription

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 01:39:06 GMT

me waiting for GPT-6 Astra access on my ChatGPT Pro subscription

OpenAI@OpenAI

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬9🔄3❤️90👀5833📊15

⚡ Powered by xgo.ing

Check out ParseBench: https://t.co/cLdEsog6a6

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 00:10:38 GMT

Check out ParseBench: parsebench.ai

💬1🔄0❤️4👀1239📊2

⚡ Powered by xgo.ing

Gemini 3.8 Flash might be maxxed on other benchmarks, but not quite on document parsing. We benchma...

Jerry Liu (@jerryjliu0) — Fri, 04 Sep 2026 00:10:08 GMT

Gemini 3.8 Flash might be maxxed on other benchmarks, but not quite on document parsing.

We benchmarked it on ParseBench. It does slightly better on tables + content faithfulness compared to Gemini 3.7 Flash, but a little worse on charts and semantic formatting. The overall scores are similar to those of 3.7 and 3.6 Flash. 

We definitely need to evolve the benchmark towards higher difficulty documents. In the meantime though, parsing real-world docs remains a real challenge for frontier models!

Bindu Reddy@bindureddy

🚨 Gemini 3.8 Flash Is Bench Maxxed And Is A Regression on 3.7 Flash 🫢

-  3.8 is worse on our benchmark than 3.7
- This version appears to be overfit to public benchmarks

It does worse on our hidden questions and regresses in data analysis.

Edit - That said, Gemini Flash 3.7 is an AWESOME model. Google should stop with more Flash releases and drop Gemini 4.0

🔗 View Quoted Tweet

💬12🔄3❤️46👀9245📊17

⚡ Powered by xgo.ing

Check out

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 21:19:11 GMT

Check out

💬1🔄0❤️1👀729📊1

⚡ Powered by xgo.ing

Check out the blog here: https://t.co/vQ63QiZS0T If you're interested in trying it out, come sign u...

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 21:19:10 GMT

Check out the blog here: llamaindex.ai/blog/introduci…

If you're interested in trying it out, come sign up for a LlamaParse account: cloud.llamaindex.ai

💬1🔄0❤️3👀850📊2

⚡ Powered by xgo.ing

Today, we're excited to launch Extract Turbo ⚡️ - the fastest VLM-powered document extraction soluti...

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 21:19:08 GMT

Today, we're excited to launch Extract Turbo ⚡️ - the fastest VLM-powered document extraction solution. 

It is 3-5x faster than all other comparable OCR solutions, including our other LlamaParse tiers, at equivalent or higher accuracy.

LlamaIndex 🦙@llama_index

Document extraction just got a new gear. ⚡️

Introducing Turbo mode for Extract, our fastest way yet to pull structured data from documents.

It runs roughly 4× faster than our Cost Effective Tier at comparable accuracy, with a median latency of 3.7 seconds a page. Turbo processes pages in parallel, so latency stays nearly flat as document size grows. 

Available in beta today.

Read the full breakdown → llamaindex.ai/blog/introduci…

🔗 View Quoted Tweet

💬5🔄20❤️98👀12234📊28

⚡ Powered by xgo.ing

I trust these vibe checks more than I trust most of the benchmark results GPT-6 astra seems more po...

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 20:50:54 GMT

I trust these vibe checks more than I trust most of the benchmark results

GPT-6 astra seems more powerful for e2e automation and better at writing (a massively underrated gap for these models). But it also seems harder to steer

Dan Shipper@danshipper

BREAKING:

OpenAI just dropped GPT-6 ASTRA!!! 🚀✨

We’ve been testing it extensively at @every across coding, writing, and knowledge work. My take: it’s a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end.

Here’s your vibe check:

- The best writing model I’ve tried. It’s fast, produces very little slop, and is easy to steer. It’s a good companion for actually working through the writing I do every day. (Not to mention, it one-shotted the first draft of its own vibe check today!)

- The computer use is wild. It can go for hours at a time using complicated apps to get work done. It did the first cut of our Fable 5.1 vibe check video...kind of mindblowing

- Impressive 3D games and visualizations. It can make beautiful 3D worlds from a single prompt. I one-shotted a historically accurate rendition of the Battle of Waterloo

- It can overcomplicate things. (Especially at higher effort levels.) Ask for a simple interface and you get extra labels, buttons, and features everywhere. It has a habit of turning everything into a landing page.

It just doesnt quite match Fable's ability to intuitively understand your prompt and do something delightful (without overcomplicating.)

Net Result:
If you already live in ChatGPT for Work or Codex and can afford it, it’s an easy upgrade from 5.6-Sol.

The biggest proof of Astra's effectiveness at helping you do work is our vibe check. We found out it was launching at 3 AM this morning, and had a 4,000 word vibe check + video done by 2 PM. Not possible without this model.

I’m reaching for Astra all day, but Fable 5.1 still gets my biggest tasks. On ambitious builds, Fable is better at understanding what I want and taking it further than I would have thought to ask.

State of Play:

Astra is launching to Enterprise customers today, and the rest of ChatGPT users over the coming days. Now, both OpenAI and Anthropic have a higher class of models that cost more to use. That changes who gets to use frontier AI and how. 

It's also a new vector of competition between them: Fable and Astra are priced at the same level. We'll see what that means for adoption in the coming days and weeks.

read our full vibe check @every today: every.to/vibe-check/gpt…L

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬11🔄6❤️54👀15182📊17

⚡ Powered by xgo.ing

We got GPT-6 before GTA-6

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 20:43:48 GMT

We got GPT-6 before GTA-6

OpenAI@OpenAI

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬7🔄2❤️46👀5713📊10

⚡ Powered by xgo.ing

Vendor-neutral startups are uniquely positioned to solve any given task better than the frontier lab...

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 02:55:10 GMT

Vendor-neutral startups are uniquely positioned to solve any given task better than the frontier labs.

They have access to the full set of open-weight and frontier models. They can still use the frontier models where needed, but can optimize the harness e2e for any given task. They can even posttrain the model itself because they have the domain expertise to create specialized evals.

Harry Stebbings@HarryStebbings

What is required to get the most out of models?

“To take the most advantage out of models, you need to do something a little different from just routing.

You need your agent or system to dynamically understand the task it is working on and understand how to allocate intelligence in a much more stateful way.

You need to know what just happened and what is going to happen in the future. You have to be in there, in the task.” @EnoReyes

How do you think about this @jerryjliu0@tomas_hk@sarahwooders@AstasiaMyers

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬13🔄2❤️23👀4633📊14

⚡ Powered by xgo.ing

parsing pdfs isn't that different from building self-driving cars

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 01:48:17 GMT

parsing pdfs isn't that different from building self-driving cars

💬41🔄9❤️211👀15665📊71

⚡ Powered by xgo.ing

We've benchmarked all the latest frontier models on hard document extraction tasks (with Fable 5.1 a...

Jerry Liu (@jerryjliu0) — Thu, 03 Sep 2026 01:40:35 GMT

We've benchmarked all the latest frontier models on hard document extraction tasks (with Fable 5.1 and Gemini 3.8 Flash coming soon!).

ExtractBench is now live on @kaggle . 5.6 Sol leads the pack, followed by other OpenAI models, then Gemini 3 Flash and Opus 5. 

Come check it out! kaggle.com/benchmarks/lla…

If you're looking for a dedicated extraction tool with out of the box grounding, citations at the price/performance frontier, come check out LlamaParse: cloud.llamaindex.ai

LlamaIndex 🦙@llama_index

ExtractBench is now live on @Kaggle.

It tests schema-guided document extraction on the documents most likely to break downstream agents and workflows, including long record lists, noisy scans, handwriting, and complex tables.

The benchmark covers 370 enterprise documents across 8 business domains and 67 document types. See how LlamaParse, Codex, Claude Code, and more stack up.

Read the full story →llamaindex.ai/blog/llamainde…

🔗 View Quoted Tweet

💬9🔄4❤️34👀5643📊13

⚡ Powered by xgo.ing

Most document OCR solutions have parse (doc->markdown) and extract (doc + schema -> structured outpu...

Jerry Liu (@jerryjliu0) — Wed, 02 Sep 2026 22:48:33 GMT

Most document OCR solutions have parse (doc->markdown) and extract (doc + schema -> structured output) endpoints.

Sometimes you want to extract information from a semi-structured document like a form, without spending the work defining on defining the exact schema.

This is where LlamaParse form mode comes in. Simply upload any form, partially filled in or not, and we’ll not only extract out the markdown, but give you key-value pairs on the precise form fields and values.

Docs: developers.llamaindex.ai/llamaparse/par…

Signup to LlamaParse here! cloud.llamaindex.ai

💬7🔄5❤️55👀4589📊22

⚡ Powered by xgo.ing

welcome to the team @imaanxsultan !! 🦙

Jerry Liu (@jerryjliu0) — Wed, 02 Sep 2026 16:34:22 GMT

welcome to the team @imaanxsultan !! 🦙

Imaan Sultan 🇵🇰🇸🇦@imaanxsultan

SO excited to share my next chapter: i’m joining @llama_index to lead all things growth marketing, brand, and community! 🦙🩷

when i came along them, it clicked incredibly fast: a team building the solution that made life SO much easier for developers, startups, enterprises, and everyone in between. 

in case you’re wondering, “wtf does llamaindex do?” we're solving one of the biggest problems in infra to date. 

the world runs on documents, yet 90% of enterprise data is unstructured, disconnected, and inaccessible to your favorite models. that's what we exist to solve, and our tech continuously tops on benchmarks (and we've processed billions of pages to date).

stay tuned for a ton of insane events, growth hacking, community building, and much more. if you're not using us yet, you will be very soon 👀

p.s. @jerryjliu0 is a natural at selfies

🔗 View Quoted Tweet

💬2🔄0❤️24👀7629📊4

⚡ Powered by xgo.ing

Fable 5.1 didn't come with document OCR benchmarks, so we benchmarked it on ParseBench: a comprehens...

Jerry Liu (@jerryjliu0) — Wed, 02 Sep 2026 16:21:07 GMT

Fable 5.1 didn't come with document OCR benchmarks, so we benchmarked it on ParseBench: a comprehensive parsing benchmark that's comprised of 2k+ real-world documents across finance, legal, insurance, and more.

As a positive note, it is one of the first frontier models in recent memory where document OCR performance actually improved between model generations! 
✅ It beats Fable 5 across the board on all of our metrics, from tables/content faithfulness/formatting/charts/visual grounding
✅ It is one of the best models out there for parsing tables

This is great, because while frontier models have rapidly advanced on reasoning benchmarks, they have not advanced on visual understanding. Prior to this the latest models (Opus 5, 5.6-Sol, 3.7 Flash) have stagnated on visual understanding tasks. 

Obviously you shouldn't use Fable 5.1 as a document OCR tool on its own; besides missing needed metadata, it also costs 15c a page - which is 10-15x more expensive than our default agentic mode on LlamaParse (which performs similarly). 

If you're interested in tracking the 90+ solutions that we've benchmarked on doc OCR, check out ParseBench: parsebench.ai

If you're looking for a dedicated, SOTA OCR tool to power your doc processing workloads, check out LlamaParse: cloud.llamaindex.ai

💬24🔄10❤️75👀9768📊42

⚡ Powered by xgo.ing

Yesterday I hosted a founder dinner with @GuangyuRobert from @Fundamental (the company behind @trysh...

Jerry Liu (@jerryjliu0) — Wed, 02 Sep 2026 01:28:49 GMT

Yesterday I hosted a founder dinner with @GuangyuRobert from @Fundamental (the company behind @tryshortcutai) to talk about a key topic:

How to build a moat as an AI startup in 2026

We didn’t arrive at a crystallized answer, but we touched on a lot of interesting conversation threads:

✅ While some companies were building products that were a bit more adjacent to the frontier labs (Shortcut, maybe us with document OCR), other companies were deeply focused on adjacent domains, from evals to infrastructure to vertical-specific services. So then the question evolved into not only how to compete against frontier labs, but how to build a moat against any startup

✅ It’s increasingly easier for any startup to duplicate your business, with everyone using AI and AI coding tools. There’s a question of whether everyone can split the pie, in which case businesses might make good money but not capture the entire market opportunity, which makes it less appealing for venture. Or can you truly build an outlier business.

✅ There’s a line of thinking that the main moats are not related to the inherent nature of the business strategy or product, but mostly around money, talent, execution. Get in first, hire the best people, raise enough $$, and it’s structurally hard for others to compete.

✅ On the other hand, there can be some legitimate moats around:
- Domain expertise in the given space
- In general, the context layer that’s locked up within an organization and not inherent to a model
- Deep-tech investment that’s hard to vibe-code
- Vendor-neutrality for application-layer companies where they can use any mixture of open-weight, posttrained, and frontier models to optimize for a given task.

It was a fun conversation. The most important bit is to keep your head straight, find the fulfillment in the journey itself, eat healthy and exercise 🙂

Thanks to everyone who participated!

This was founder dinner 002. We hope to do this every month and are hoping to host founder dinner 003 soon! If you have topics in mind, let us know.

💬14🔄1❤️42👀4600📊21

⚡ Powered by xgo.ing

There’s generally a massive difference in quality between specialized OCR providers, “simple” open-w...

Jerry Liu (@jerryjliu0) — Tue, 01 Sep 2026 01:30:36 GMT

There’s generally a massive difference in quality between specialized OCR providers, “simple” open-weight OCR models, and free/OSS solutions.

1️⃣ Specialized OCR models (including LlamaParse) solve for the long-tail of complex documents, and make sure that everything is digitalized properly with lower hallucinations. They typically use posttrained VLMs to cover a wide range of real-world docs. They have tuned bounding boxes and annotations for each section, letting agents trace citations back to the source. They also usually come with additional endpoints like extraction and splitting.

2️⃣ Open-weight VLMs (e.g. Paddle, MinerU, UnlimitedOCR) are reasonable over relatively simple documents like text and tables and can do basic visual reasoning. They can seem somewhat cheap to host but can be unreliable in quality.

3️⃣ Free OSS libs (including liteparse) are meant to be universally accessible, fast text extractors. They’re not meant to do any sort of visual reasoning, so won’t perform any linearization, or reasoning over complex tables, or OCR over non-native docs. AI agents like Claude will by default use these tools to do a light pass over documents. But I would caution using these for retrieval, because they will drop entire sections that are not digitalized.

At this point we’ve benchmarked over 92 tools on ParseBench. Come check it out!

parsebench.ai

💬30🔄21❤️127👀13468📊79

⚡ Powered by xgo.ing

We’re excited to launch the official LlamaParse connector on Claude! Yes Claude can already read PD...

Jerry Liu (@jerryjliu0) — Mon, 31 Aug 2026 17:38:27 GMT

We’re excited to launch the official LlamaParse connector on Claude!

Yes Claude can already read PDFs, and it probably works fine for relatively simple queries over relatively simple docs. But if you want to:
✅ Parse complex tables, graphs, filled-in dense forms, redlines from docs
✅ Include visual citations/bounding boxes for auditability
✅ Extract a *lot* of docs (e.g. 1k+) at scale without breaking the bank

Then you’re going to want a dedicated OCR tool that can handle any document at the Pareto frontier of accuracy and cost.

Install the connector here: claude.ai/directory/llam…

If you haven’t already, come sign up for a LlamaParse account! cloud.llamaindex.ai

Your browser does not support the video tag.

🔗 View on Twitter

LlamaIndex 🦙@llama_index

LlamaParse is officially a verified connector on Claude! 🦙

Claude is perfectly fine for the occasional text-only PDF. But most of the context your agent needs lives in messy PDFs, spreadsheets, scanned forms, dense tables, and charts. 

Feeding those docs straight to the model burns through tokens and can lead to hallucinations when layouts get scrambled, chart values are missed, or numbers are pulled from the wrong column. 

LlamaParse converts those documents into clean, structured context Claude can work with. With the connector you can

✅️ Parse documents into clean Markdown, JSON, or HTML
✅️ Extract specific fields directly into your schema
✅️ Search across document collections with filesystem-style tools
✅️ Classify documents or split them into predefined sections

Live in the Claude Connectors Directory now !
claude.ai/directory/llam…4

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬9🔄0❤️42👀6943📊15

⚡ Powered by xgo.ing

Thanks @juliafedorin for a great conversation! Some tidbits from the podcast on the AI space: - We...

Jerry Liu (@jerryjliu0) — Mon, 31 Aug 2026 17:19:47 GMT

Thanks @juliafedorin for a great conversation! 

Some tidbits from the podcast on the AI space:
- We've pivoted from building a RAG framework towards building the document infrastructure for agents
- Documents are 90% of the world's unstructured context, and solving the context layer is the most important bottleneck towards unlocking general intelligence
- There's a lot of ways to hack building basic RAG, but building production retrieval systems is still hard

Julia Fedorin@juliafedorin

My conversation with @jerryjliu0.

He's building @llama_index: document infrastructure for AI agents, the layer underneath everything models need to work.

Did you know he's the guy whose tool half the industry used for RAG? Then he called RAG a hack, two years before the field agreed.

Here's the problem he's pointing at: 90% of the world's information is locked inside PDFs, PowerPoints, and Word documents. The models are already smarter than humans. But until an agent can read and reason over all that, it just sits still.

RAG was the hack everyone reached for to bridge that gap: chunk, embed, retrieve, done. The easiest thing that kind of worked, so people could ship demos. Anything complex, and it falls apart.

We got into:
•Why RAG has no first principles, and what's replacing it
•Why context, not intelligence, is the real bottleneck now
•Why even MCP doesn't fully solve context
•How you get agents to follow a company's culture, not just its instructions
•What he'd build if he started LlamaIndex today

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
 (00:00) The Context Bottleneck
 (02:00) One Commit, One Tweet: The Origin
 (05:00) Feeding GPT-3 Your Private Data, Pre-RAG
 (08:00) From Quora to Uber's Self-Driving Lab
 (11:00) Meeting His Co-Founder at Uber ATG
 (14:00) Document Infrastructure for AI Agents
 (18:00) "RAG Is a Hack" — What He Meant
 (22:00) Why Production Search Is Still Unsolved
 (24:00) How a Research Background Changes How You Build
 (33:00) The Open Problem: Context Across Sources
 (38:00) Specs, Cultural Context, and Underspecified Tasks
 (40:00) What He'd Build Differently Today

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬4🔄1❤️27👀4882📊8

⚡ Powered by xgo.ing

we're hosting a founder dinner on this exact topic tonight! is the moat in: - product - data - b...

Jerry Liu (@jerryjliu0) — Mon, 31 Aug 2026 16:23:19 GMT

we're hosting a founder dinner on this exact topic tonight! 

is the moat in: 
- product 
- data
- brand
- workflows
- network effects
- working harder than everybody else
- anything else?

Event here: luma.com/llamai-8hry

even if you can't make the dinner, we'd love to get your general thoughts. we'll also be hosting more events like this on a regular cadence!

Auren Hoffman@auren

there are no more moats anymore.  moats are dead.
was talking to a friend of mine who's a CEO of a rapidly growing startup (~$1B revenue). we're small investors. He told me that he has only taken one one-week vacation in the last eight years of his startup. 
I was like, "Hey this is crazy. You're gonna burn out. This is pretty serious."

he said, "Auren, I don't know what to do. We don't have no moat." 

am like, "What are you talking about? The business is going great!" 

he said, "We have incredible competitors. We sell AI products and so we're selling to the smartest people in the world. If we don't have the best product that week, they won't buy us. If we don't have the best product for two weeks, our customers will churn. We don't have a moat."

I started thinking about it: who does have a moat? 
There are just no moats anymore. Nobody has a moat. Everyone will switch their software providers in a heartbeat for better product or for cheaper product or for faster product. Anyone will switch.

We looked at our 200 companies at Flex Capital and how they're buying software and AI services today. Almost all of their vendors are on month-to-month contracts or on usage-based contracts. That means they can switch immediately. The reason they have everyone on month-to-month is that they have no idea if they're going to be using those tools two months from now. They may not even be using those tools two weeks from now. Everything is changing so fast all the time.

The only moat you have in software is that you have the best product. There are just no other moats. 

It used to be back in the day that you have a mediocre product and you'd be using it for years. we just dealt with all these of mediocre software systems, mediocre cloud providers, mediocre data providers, mediocre APIs. we dealt with it because the switching costs are really hard. 

Nowadays it's really easy to change software. It's really easy to change APIs. it's really easy to change AI infrastructure. It's super easy. You can change it in a heartbeat and every other vendor will help you switch. Of course, your AI tools will help you switch as well. All these things that used to give very high customer lock-in just don't exist anymore and so they're just no more moats. 

Moats are dead.  you just need to work insane hours all the time because the only thing you have is your product.  goodbye moat, goodbye predictable revenues, goodbye 2010s era.

🔗 View Quoted Tweet

💬7🔄0❤️25👀8219📊9

⚡ Powered by xgo.ing

I had to write Java during my junior year internship in 2016. And my manager asked me what IDE I was...

Jerry Liu (@jerryjliu0) — Mon, 31 Aug 2026 04:36:06 GMT

I had to write Java during my junior year internship in 2016. And my manager asked me what IDE I was using. I told him, I'm using vim with no plugins.

My whole life, I wanted to see if I was a badass or not. Am I a fake badass that coded in IntelliJ / NetBeans / VSCode, or am I really a badass on the inside?

dnap@dnapway

Sequoia partner Doug Leone reveals he got a root canal with no Novocaine to see if he was really a badass

"I had to get a root canal. And I had to go to a meeting right after that. I told the dentist, drill without any Novocaine.”

“My whole life, I wanted to see if I was a badass or not. Am I a fake badass that has this envelope nobody could tell, or am I really a badass on the inside?"

"He drilled and I jumped. And I told the dentist, I won't move again. And I didn't move again. I used the trick of thinking what it must be like to go to war and lose an arm or two legs. That's pain with consequences. Mine was pain with no consequences."

“It’s just nerve endings from there to my brain to my spine, f*ck it I’m not moving.”

"I had to go home and change my shirt because I was a puddle of sweat. But I didn't move. And so I love fear. It does wonders for me."

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬12🔄4❤️148👀24204📊26

⚡ Powered by xgo.ing

at the end of the day, we all have to stand by our company and the values it represents 🙂 * fronti...

Jerry Liu (@jerryjliu0) — Sat, 29 Aug 2026 04:06:18 GMT

at the end of the day, we all have to stand by our company and the values it represents 🙂

* frontier labs will want to own their own application/harness/model e2e. 
or at least, openai will to start with, and then anthropic will do the opposite of whatever openai does

* everybody else will push for democratized access to a wide mixture of proprietary and open-weight models, so that we can optimize performance and margin for any given task and we're not beholden to the whims of the provider 

(this of course includes us with LlamaParse and any other startup building verticalized agents)

Tibo@thsottiaux

We unfortunately have decided that we cannot continue providing access to our models through Cursor and are ending our partnership. It boils down to trust and we’ve asked that this takes effect on November 12 to give you some time to plan.

Many have used the GPT models through Cursor and here are options we know should work in the future:

- We will continue to allow using your own OpenAI API key and similarly will continue to provide access through our IDE extensions for Cursor.
- We will keep working with the broadest range of tools and harnesses, some of which are OSS, but also many many closed-source ones.

We are as committed as ever to continue supporting developers and the flourishing ecosystem of tools, harnesses and products. We will also continue to invest in our own open-source initiatives and believe in broad optionality for developers.

You can read more about our decision in the blog:
openai.com/index/our-deci…

🔗 View Quoted Tweet

💬14🔄0❤️46👀16282📊16

⚡ Powered by xgo.ing

openai's hatred for elon must've outweighed any aura loss this would cause them

Jerry Liu (@jerryjliu0) — Sat, 29 Aug 2026 04:02:52 GMT

openai's hatred for elon must've outweighed any aura loss this would cause them

OpenAI@OpenAI

We’re ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor’s direct access to our models would end on November 12.

We know that the people most affected by this decision are the developers who rely on OpenAI models in Cursor. We care about their experience in this transition and we’re ready to go above and beyond to support them.

openai.com/index/our-deci…

🔗 View Quoted Tweet

💬13🔄2❤️84👀10280📊18

⚡ Powered by xgo.ing

Agentic spreadsheet extraction 📊 is hard to execute well from a price/performance standpoint becaus...

Jerry Liu (@jerryjliu0) — Fri, 28 Aug 2026 17:16:38 GMT

Agentic spreadsheet extraction 📊 is hard to execute well from a price/performance standpoint because your agent needs to ingest a massive quantity of input tokens from a highly complex interface (e.g. the Excel file format), and emit a massive number of output tokens that accurate represents the source document.

We've tuned an agentic harness within LlamaParse that can extract from even the most complex spreadsheets and schemas, and emit a large volume of structured output with extremely high accuracy.

It can be used for any document, and now it can be used for spreadsheets too! Simply toggle "spreadsheet extraction" in the advanced processing options (see screenshots below) 

Come check it out:cloud.llamaindex.aiA

Jerry Liu@jerryjliu0

We've introduced native, agentic spreadsheet extraction 📊 into LlamaParse. 

Spreadsheets are a wildly different format from PDFs (or any other document format). They can span arbitrarily many rows/columns, lack any guarantees on tabular structure, and can have information linked across multiple sheets. The native ways to deal with spreadsheets is usually through a code interpeter as opposed to OCR.

We've introduced a tuned agentic engine (model+harness) that is equipped for large-scale schema guided extraction from spreadsheets. You can extract out dense sheets like balance sheets into clean structured fields. 

Check out the docs for how to integrate into the API:developers.llamaindex.ai/llamaparse/ext…W

On the UI, select "agentic plus", and see "spreadsheet options" in the advanced options section. 

Signup for LlamaParse here!cloud.llamaindex.aiZ

🔗 View Quoted Tweet

💬15🔄3❤️47👀5719📊22

⚡ Powered by xgo.ing

static embeddings are extremely fast/cheap compared to dense embeddings, and still quite powerful. ...

Jerry Liu (@jerryjliu0) — Fri, 28 Aug 2026 15:20:51 GMT

static embeddings are extremely fast/cheap compared to dense embeddings, and still quite powerful.

one issue is that they lose accuracy over longer text, since the tokens get pooled and averaged

late interaction models score with MaxSim instead of average pooling, and avoid the pooling accuracy degradation. Can we apply these concepts to static models? 

This is a fun writeup by @LoganMarkewich - come check it out! llamaindex.ai/blog/exploring…

LlamaIndex 🦙@llama_index

Static embeddings offer unmatched throughput, but they also suffer from accuracy loss compared to traditional embedding models. Can we make them better for retrieval?

We tried several things:
✅️  raw maxsim scoring on the per-token embeddings
✅️  training a small adapter model
✅️  changing the distillation training target and teacher

While none of these saw the results we wanted, it's an excellent dive into static embeddings and how they do (and don't) work!

Blog: llamaindex.ai/blog/exploring…

🔗 View Quoted Tweet

💬10🔄8❤️53👀6681📊20

⚡ Powered by xgo.ing

The hardest road paid off

Jerry Liu (@jerryjliu0) — Fri, 28 Aug 2026 10:23:46 GMT

The hardest road paid off

Kevin Durant@KDTrey5

Excited for the new release of #HuggingFace durant.ly/huggingface - proud investor! durant.ly/hfrelease

🔗 View Quoted Tweet

💬1🔄2❤️17👀8862📊3

⚡ Powered by xgo.ing

Grok bot is surprisingly fun. Coming from Claude Code/Cowork/Codex I wasn't expecting to like it so ...

Jerry Liu (@jerryjliu0) — Fri, 28 Aug 2026 04:55:15 GMT

Grok bot is surprisingly fun. Coming from Claude Code/Cowork/Codex I wasn't expecting to like it so much, but I severely underestimated the niceness of having a persistent cloud computer

💬39🔄14❤️273👀20606📊77

⚡ Powered by xgo.ing

PDF parsing is fun because there's an infinite variety of enteprise documents 📑. For each document ...

Jerry Liu (@jerryjliu0) — Fri, 28 Aug 2026 04:53:55 GMT

PDF parsing is fun because there's an infinite variety of enteprise documents 📑. For each document category, there's a long tail of work to build more precise bounding boxes, confidence scores, and domain-specific annotations so that you provide any downstream agent rich metadata without it having to reinvent this from scratch. 

Take forms for example. Besides simply outputting it into markdown, we put in the work to detect every annotation, field, checkbox, and section. That way you 
 immediately get structured information as to whether a form is filled without a separate LLM extraction step. You also get source citations for free! 

Doing this well is hard. 
✅ There's a rabbit-hole of optimizations you can do, including extracting annotations for every document type. 
✅ It's hard to properly render visual formats like charts, handwriting into digitalized information. The more you skip this step, the more work you're creating for any downstream agent.
✅ Precise bounding boxes are a necessity for precise citations on any type of document. 

You can aggressively tune the model + harness so that the accuracy/cost on any document subtype is much more competitive than the frontier models.

Whether you're parsing forms (see the enriched forms option in "processing options") or any other type of doc, check out LlamaParse !
cloud.llamaindex.ai8

💬20🔄7❤️137👀12183📊62

⚡ Powered by xgo.ing

We've introduced native, agentic spreadsheet extraction 📊 into LlamaParse. Spreadsheets are a wil...

Jerry Liu (@jerryjliu0) — Thu, 27 Aug 2026 17:29:48 GMT

We've introduced native, agentic spreadsheet extraction 📊 into LlamaParse. 

Spreadsheets are a wildly different format from PDFs (or any other document format). They can span arbitrarily many rows/columns, lack any guarantees on tabular structure, and can have information linked across multiple sheets. The native ways to deal with spreadsheets is usually through a code interpeter as opposed to OCR.

We've introduced a tuned agentic engine (model+harness) that is equipped for large-scale schema guided extraction from spreadsheets. You can extract out dense sheets like balance sheets into clean structured fields. 

Check out the docs for how to integrate into the API:developers.llamaindex.ai/llamaparse/ext…W

On the UI, select "agentic plus", and see "spreadsheet options" in the advanced options section. 

Signup for LlamaParse here!cloud.llamaindex.aiZ

💬10🔄8❤️50👀13586📊29