System

Configuration and component status.

LLM backend

Connected — gemini

Gemini vision extraction using gemini-3.1-flash-lite. Page images are sent directly; the model is instructed to return null rather than guess. $2.7626 of $4.00 spent across 1007 requests. 24/400 requests today (UTC), spaced to 300/minute.

Free-tier keys are not suitable for real customer documents. Google's terms for the unpaid Gemini API allow submitted content to be used to improve Google products, and state that human reviewers may read it. Google explicitly advises against sending confidential or personal information to unpaid services.

A funding packet contains social security numbers, dates of birth, addresses, income and bank details. Sending those on a free key means disclosing them to a third party for training — almost certainly a breach of CRIF's obligations as a credit bureau.

Enable billing on the key before processing real packets. Paid Gemini API usage is covered by Google's data processing terms and is not used to train models. The code is identical either way; only the key's billing status changes. Redacted test packets on a free key are fine.

Classifier

Trained	No
Document types	0
Training samples	0
Leave-one-out accuracy	Not measurable — needs ≥2 samples for some type
Last trained	—

Directories & thresholds

Inbox	/root/Mo Crif/crif-document-intelligence/data/inbox
Outbox	/root/Mo Crif/crif-document-intelligence/data/outbox
Failed	/root/Mo Crif/crif-document-intelligence/data/failed
Watcher	Enabled — every 5s
OCR	Enabled — tesseract (ita+eng)
Confidence threshold	0.55
Ambiguity margin	0.25