Decision questions - Bring AI Home

Defined in qvac-service/reasoning.ts; wired up in qvac-service/index.ts, evaluation-cli.ts and offline-cli.ts. Since lesson 06.

Model

Item	Value
SDK constant	`LAYA_421M_Q8_0`
Health ID	`laya-421m-q8`
File	`laya-Q8_0.gguf` (450 MB)
`loadModel`	`modelType: "llamacpp-decision"`, `modelConfig: { device: "gpu" }`
Call	`decide({ modelId, states, questions })`, one state per OCR line

Official page: Decisions.

Questions

Every question is type: "noul": the answer is the probability that the statement holds.

Question ID	Statement	Asked at
`name`	This text contains a person’s name.	Private, Confidential
`email`	This text contains an email address.	Private, Confidential
`phone`	This text contains a telephone number.	Private, Confidential
`address`	This text contains a street address.	Private, Confidential
`account`	This text contains a bank, card or account number.	Private, Confidential
`identity`	This text contains a passport, national ID or employee ID number.	Private, Confidential
`amount`	This text contains a specific amount of money.	Confidential
`date`	This text contains a sensitive date such as a birth date.	Confidential
`reference`	This text contains a tracking, case or project reference code.	Confidential
`business`	This text contains a private company detail.	Confidential
`direction`	This text is something the user asked to hide: (the typed direction)	When a direction is typed

From answers to suggestions

For each OCR line:

Take the question with the highest probability.
If it is below DECISION_THRESHOLD (0.5), make no suggestion.
Otherwise propose that question’s label (other for direction) with the probability as confidence.
Apply the evidence rules in Labels: the rule’s label wins, and a line the rules don’t support is dropped, except name and other.

The decide() call fails the analysis (503 analyzer_unavailable) if it returns a different number of results than lines.

Copy check

Since lesson 11. After Done, the browser sends each flattened page to POST /v1/analyze-page with the same level and direction. Any suggestion returned is listed under Still readable on the copy. It doesn’t block Save copy. If the service is busy, preparing the next document waits up to 60 s for it.

Measured

Apple Silicon, @qvac/sdk 0.22.0, the four evaluation pages, two runs with identical results:

Item	Value
Decide one page (12 lines)	0.6–1.0 s
Model load, cached	about 0.8 s
F1 (clear / rotated / low contrast / obscured)	0.92 / 0.92 / 0.92 / 0.83
Same answers on repeat runs	Yes, with no sampling settings

​Model

​Questions

​From answers to suggestions

​Copy check

​Measured

Model

Questions

From answers to suggestions

Copy check

Measured