Technical Notes
Needl.ai publishes its engineering notes and benchmark results here, from PDF extraction to OfficeQA Pro and context infrastructure.
Benchmarking PDF Extraction Tools
Evaluation criteria, scoring functions, and comparative results across 42 extraction libraries and vision–language models, scored against a human-verified golden reference.
Read the noteAskNeedl on OfficeQA Pro
Against the strongest published frontier agent on 89,000 pages of raw PDFs: 16.5 points more accurate, 13× faster and 4.6× cheaper per question.
Read the noteContext Infrastructure
Architecting a high-accuracy, permission-aware retrieval engine for the agentic enterprise — across unstructured documents and structured data, at hundreds of terabytes.
Read the noteFrequently asked questions
What are Needl.ai technical notes?
Engineering write-ups from the Needl.ai team: how the system is built, how it is measured, and the results.
How does AskNeedl score on OfficeQA Pro?
On OfficeQA Pro, a 133-question benchmark over 89,000 pages of U.S. Treasury Bulletins in raw PDF, AskNeedl is 16.5 points more accurate, 13× faster and 4.6× cheaper per question than the strongest published frontier agent: 64.66% of answers exactly right against 48.12%, in 2m 22s against 31 minutes, for $0.98 against $4.55. The method and caveats are in the OfficeQA Pro note.
Which PDF extraction tools did Needl.ai benchmark?
42 extraction libraries and vision-language models, scored against a human-verified golden reference on seven criteria. See Benchmarking PDF Extraction Tools.
What is context infrastructure?
The layer that finds and assembles the information an AI agent works from, with each source’s permissions applied. The context infrastructure note describes how Needl.ai builds it across documents and structured data.
See Needl.ai run on your own documents
The fastest way to get started is a demo on the documents your team works with.