A small model selects one candidate

From intent to selection

Retrieval and reranking reduce a potentially large graph to request-local candidates. The router selects one candidate using the current FSM context.

intent + state + history
  → semantic retrieval
  → candidate reranking
  → router SLM selects one node_id
  → transition validation
  → node execution

Selection output

• The selected_node_id must match a candidate supplied in the request.
• Fallback is represented as a candidate and can be selected when appropriate.
• The runtime, not the model, decides whether the selected transition is legal.

{
  "reasoning": "Identity must be verified first.",
  "selected_node_id": "verify_identity"
}

Training direction

The training pipeline fine-tunes Qwen2.5-1.5B with supervised learning and GRPO. Rewards emphasize valid JSON and correct candidate selection.

Synthetic scenarios span customer support, fintech, supply chain, marketing CRM, and DevOps. Their constraints teach planning patterns; they are not a runtime axiom engine.