Our team spent 10 years teaching machines to read documents.
LLMs improved language understanding, but reliable document extraction still depends on the system around the model.
Here’s what we learned building Azure Content Understanding 🧵
An agent reported 46.1 dB for F-03’s ORL at 1310 nm.
That cell was empty. It borrowed the value from the 1550 nm column after extraction flattened the table.
For your document pipeline, preserving structure is part of preserving meaning.
Our tests showed LLMs outperforming specialized models on contracts and agreements, where meaning lives in the words.
On tax forms and mortgage applications, where layout carries meaning, our existing approach still won.
Match the approach to the document.
Grounding ties extracted values to source evidence. Confidence scores flag uncertainty. A developer-defined JSON schema gives your app structured output.
Together, they help you validate results or route them for review before your app or agent acts.
Oct 2, 2026 · 8:22 PM UTC
Applying LLM reasoning selectively can also reduce token usage.
With recent contextualization work, pre-built analyzers in the current preview cut LLM token consumption by up to 99%.
Read Cha Zhang’s full breakdown of the work and lessons.
commandline.microsoft.com/az…
