Teach Model

SUPERVISOR Teach Model is how supervisors and administrators correct the AI's output after it has run. Lives at Admin → Knowledge → Teach Model.

What it writes

Corrections become rows in cl_knowledge_base with seeded_by='admin' (when you author them) or seeded_by='auto' (when the self-learning hook mirrors a non-REJECTED supervisor approval). Schema:

• target_code — the UCS code the correction is about
• correction_text — short text the LLM should bias toward next time
• learning_weight — 1.0 default; admin can override
• quarantined — 0 while target_code is in active Master, 1 if orphaned

What it does NOT do

• It does not write to live PMS data
• It does not retrain a model
• It does not change retrieval weights — only the few-shot context injected into the next AI call

The KB is read-only from the LLM's perspective. The LLM never writes to it. Only humans (and the auto-mirror hook for approved corrections) write to it.

Token-budget cap (v2.31.0.19)

correctionsToFewShotByBudget caps the few-shot context at a token budget rather than a row count. So a long correction takes its share; a short one takes less. Avoids the v2.31.0.18 problem where 50 short rows starved out 10 long, useful ones.