New model (and it’s named after a baby llama) 🦙 | LlamaIndex Newsletter

🚢 JUST SHIPPED

• Boo Bench 👻: We built a haunted house for document parsers. Boo Bench is 179 pages built to scare parsers into loops: endless dot leaders, near-identical rows, and scans in scripts models rarely see. One model looked at a page with 246 x’s on it and wrote 23,004. See who got spooked here.
  Read the paper: Boo Bench | Dataset: Hugging Face
• Document Parsing Protocol: Two parsers can get every word on a page right and still disagree about what a table means. DPP 0.1 is our draft answer key for parsing, with 140 rules for tables, forms, charts, and diagrams, each shown beside its expected output. Try it here, and if a rule breaks on one of your docs, let us know on Github.
• ExtractBench Pro: We built 100 documents from the ways real customers broke our extraction, including legends that change the meaning of other pages, house jargon, and values that only exist as a line in a diagram. Scores drop 18 to 47 points compared with the original ExtractBench. The dataset is open, so you can see how your extractor holds up!

👋🏽 IRL

RECENTLY:

• Doc Chai Launch Party: We launched our very own drink, Doc Chai, with a party at Corgi Cafe in SF. More than 900 people signed up to try it. Yes, a doc parsing company made a chai :)

• AI Leaders @ Formula 1 US Grand Prix: We spent F1 race weekend in Austin with AI leaders and customers. Days at the track, yummy Southern food at a private loge in the main grandstands, and a lot of great conversations!
• Unindexed Ep.4: Our behind-the-scenes series is back with a new episode, this time at the Agent Economy Gala. We co-hosted with SpaceX, Lemma, Daytona, TinyFish, and more, and 500+ founders and investors showed up in black tie.

COMING UP:

When a value gets attached to the wrong claim or location in a submission, the underwriter prices the wrong risk. On November 4th, we’ll build an intake workflow with LlamaParse that takes ACORD apps, SOVs, loss runs and bordereaux, whether they’re PDFs, Word docs or multi-tab Excel, and maps them into one schema across every carrier and producer layout. Anything uncertain goes to a reviewer, with a citation back to the exact spot in the source. [RSVP HERE]

Nov 5- Marginalia: An Insurtech Leaders Private Dinner (SF)

🚨 SIGNAL

• Anthropic filed to go public.
  Anthropic made its S-1 public on Oct 26 and plans to list on Nasdaq under the ticket ANT. No price or share count has been set just yet. We parsed and synthesized every single page in the S-1 and built this reader so that you can get the highlights.
• Apple is putting agents inside Pages, Keynote, and Numbers.
  Nvidia-backed Apple's new assistant can draft documents, build presentations, and analyze spreadsheets. macOS 27 also adds Model Context Protocol support, so third-party agents can connect to those apps directly.
• Big banks are locking AI agents out.
  Meta's Muse agent hit 10 million US downloads in its first 40 days, and people started handing it their bank logins. Morgan Stanley, Bank of America, and Wells Fargo are now rewriting their terms to stop agents from acting on their own in customer accounts. JPMorgan has decided to not let any external agents in their platform at all.

• Building a new agent? Get in touch with our team.

• Need support? Join our Discord.

• Check out our Github for the latest code releases and documentation.

Thanks for reading this week's newsletter :)

We want your thoughts! Feel free to reply directly or email [email protected].