Skip to content
Arfniia Router

Search is only available in production builds. Try building and previewing the site to test it out locally.

Episodic Routing

Why Episodes Matter

Some applications evaluate success only after a series of turns: tool-executing agents, troubleshooting flows, tutoring sessions, or coding assistants that gather requirements before delivering a solution. Episodic routing lets Arfniia accumulate the intermediate observations and assign the final reward to the whole trajectory.

Headers to Set

Add the following headers when calling /v1/chat/completions:

  • X-Arfniia-Episode-Id: stable identifier for the session (UUID recommended).
  • X-Arfniia-Episode-Start: truthy value on the first turn.
  • X-Arfniia-Episode-End: truthy value on the last turn.

Intermediate turns only need the episode ID. The router drops these headers before invoking Bedrock.

Example Timeline

TurnHeadersNotes
1X-Arfniia-Episode-Id=123
X-Arfniia-Episode-Start=true		Capture initial prompt and action
2X-Arfniia-Episode-Id=123Feedback from turn 1 is pending
3X-Arfniia-Episode-Id=123
X-Arfniia-Episode-End=true		Final answer generated

After the user or system submits feedback (either sparse KPI or per-response), the learner stitches the cached observations together and updates the policy.

Providing Feedback

  • Send a single sparse KPI once the episode outcome is known:
    PUT /v1/feedbacks/<router>/sparse/<value>.
  • Optionally submit per-turn feedback keyed by response IDs if you want faster corrections.

Operational Tips

  • Keep episode IDs unique per router; stale caches are cleared automatically when Episode-Start is received.
  • If the process aborts early, send Episode-End with the latest turn so the replay buffer is flushed.
  • Combine episodic signals with custom features (e.g. user sentiment or dialog depth) to give the learner richer state.