Create Intent Map

Train LoRA adapters for converting natural language queries into structured QuerySpec JSON. Uses GRPO (Group Relative Policy Optimization) with execution feedback from ArangoDB.

Prompt Iteration Rule (NON-NEGOTIABLE)

System prompts for intent mapping SFT data MUST be validated through /prompt-lab before training. NEVER hand-craft system prompts in Python strings.

• Before training: /prompt-lab eval the QuerySpec system prompt against ground truth
• Comparing prompt variants: /prompt-lab compare across models
• Only after prompt-lab validation → proceed to LoRA training

Training Approaches