AMF

ERP applications and tools for AMF

Monthly Operations AI Insights

The monthly Operations KPI Report can enrich its deterministic KPI snapshot with bilingual, evidence-backed OpenAI insights.

Configuration:

1. Install requirements.txt in the target bench Python environment, then run bench --site <site> migrate and bench restart. Example from the bench root: ./env/bin/pip install -r apps/amf/requirements.txt.
2. Enter the API key in the Single Operations KPI Report Settings DocType. AMF encrypts it with the site's encryption key into a Long Text field, avoiding Frappe v12's short __Auth.password column. The key is never displayed again after saving. OPENAI_API_KEY remains an optional deployment fallback.
3. Enable AI Insights, select the model and confidence threshold, and keep Require Human Approval enabled for controlled distribution. Each report has an editable Generate AI Insights checkbox, enabled by default.
4. Generate a monthly or semester report. Validated AI output is stored on the report for review but is excluded from files and email until approved.

For reliable monthly and comparative reports, keep AI reasoning on low, use a timeout close to 600 seconds, and limit the maximum insights to about 10. Higher reasoning with the maximum 15 insights can exceed the OpenAI read timeout on larger KPI snapshots.

Reports can run in Single Period or Comparative mode. Comparative mode keeps the primary month/semester calculations intact, adds a selected comparison month/semester, stores primary-versus-comparison KPI deltas in the snapshot, and asks the AI to prioritize evidence-backed evolution insights.

The KPI calculations remain authoritative. The AI receives a bounded snapshot, external parties and transaction identifiers are aliased by default, API storage is disabled with store=False, Issue root-cause free text is excluded unless explicitly enabled, and every retained insight must cite a real leaf-level source path. Evidence values are replaced server-side with the canonical values from the snapshot before storage.

The read-only investigation methods in amf.amf.utils.operations_ai_tools expose bounded OTIF, shortfall, machining scrap, shipping issue and procurement exception details. They do not provide SQL execution, document mutation, email or other outbound actions.

License

AGPL