— Tag

#vlm

1 post tagged vlm.

#crs (4) #data-centers (3) #fieldintel (3) #gotchas (3) #landledger (3) #row (3) #transmission (3) #field-workflow (2) #market (2) #qc (2) #ai (1) #archive (1) #cost-and-roi (1) #datum (1) #ferc (1) #ingest (1) #institutional-memory (1) #local-grid (1) #mobile (1) #pnezd (1) #product-tour (1) #project-management (1) #prompt-engineering (1) #re-visit (1) #state-plane (1) #theo (1) #verification (1) #vlm (1)

— vlm

Why We Replaced 'Extract' With 'Transcribe' (And Got 7x More Data)

FieldIntel ingests handwritten field books by asking a local vision-language model to read each page. The original prompt asked the model to extract and organize what it saw -- and produced confident hallucinations on roughly half the regions we tested. Replacing 'extract and organize' with 'transcribe verbatim' lifted total entity extraction by 6.7x on the same model, same hardware, same input. Here's the test, the failure mode, and the one-line config change that shipped in v1.2.2.

June 2026 · 6 min read

#vlm

Why We Replaced 'Extract' With 'Transcribe' (And Got 7x More Data)

Get new posts in your inbox