Loading video...
Video Failed to Load
We tuned an AI agent that can do large-scale document extraction from long docs (50+ pages, some with 10k-100k fields) with 94%+ accuracy 📈 It uses a harness + model set that is tuned specifically for reasoning over extracting out complex information from complex docs. Each extracted field comes... show more
18,359 views • 26 days ago •via X (Twitter)
25 Comments
You have the power to help, and we are eagerly awaiting it. Don't just pass by without doing something. [Campaign Link:
This is cool, I see it as an expert agent that can be consulted by others via A2A or another protocol. I also need to remind myself to update LiteParse version in Agent Zero. Saw many improvements across the board : )
ok that's genuinely cool. Bounding boxes on every field is the best part.
Provenance on every field is the right instinct. The EvaEsi Index does the same for claims, keeping the source attached and flagging you when a cited page changes underneath.
94% on extraction is the demo number. In production it dies when the same field has two legal values in one packet: invoice total vs line total, both labeled Amount. The model picks one and never flags the collision.
94%+ extraction across 100k fields is the kind of result that deserves a reproducible corpus, dedup rules, bounding-box checks, and an error taxonomy. Long documents punish vague benchmarks.
the confidence score + bounding box is what really stands out. accuracy tells you it works; provenance tells you where to look when it doesn't.
94% on 50k+ fields is a solid leap, but enterprise compliance needs deterministic guarantees for the remaining 6%. Pairing the agent with schema-level cross-field validation rules and deterministic constraint solvers could bridge that gap without adding manual review.
94%+ on 100k fields is insane 🔥 Big win for LlamaParse
The confidence score + bounding box combination is the part that actually matters for production use. General coding agents are optimized for “looks plausible.” Specialized harnesses that force evidence and location for every field are optimized for “can be audited.” That gap is still under-appreciated.
agentic plus is what makes the demo land. Showing the before and after in the same frame would make the change even easier to judge.
Yes. The harness matters. In production, the model still needs a human to sign off when it improvises off-spec. I see that here in the States every day.
confidence scores and bounding boxes make automated extraction much easier to audit.
the reason most people miss this is that they don't realize the majority of human knowledge still lives in documents, not structured data. our job is to bridge that gap
the 10k to 100k field range is where every extraction pipeline ive used starts double counting rows, hows yours handling dedupe
94% accuracy on 100k fields is impressive 📈
Long-document extraction reaching 94% accuracy makes structured evaluation the real breakthrough.
94% on docs that long is honestly wild, we struggle with 10-page PDFs sometimes. What’s the biggest bottleneck you hit when tuning it?
On long docs the extraction was never the hard part, building ground truth to trust it was. Our worst error class on 50+ page contracts: confidently pulling the right field from the wrong table across a page break. Layout-aware chunking helped more than a bigger model.
The bounding box is what makes the confidence score usable. A score on its own only tells you to doubt a field. A score plus a location means the reviewer lands on the exact spot instead of rereading 50 pages, which is what makes a threshold policy affordable in the first place. 94% matters far less than whether the system knows which 6% it got wrong.
Specialized > generalized for extraction
i'll be reading more on this
narrow harness beats a general one
How does wall clock look on a 100k field doc versus the coding agent harnesses? Accuracy lift is easy to sell, waiting an hour per doc less so.
confidence score plus bounding box is the right instinct. curious what happens below some threshold. does a low-confidence field get flagged for a human to check, or does it ship with the number attached and the burden's still on whoever reads the output.
