POST /reset is always the first call you make. It generates a new episode UUID, loads the appropriate messy CRM dataset for the configured task, snapshots the hidden ground truth into isolated per-episode server memory keyed by that UUID, and returns an initial CRMPipelineObservation. All in-memory Pandas DataFrames are freshly copied so that concurrent episodes cannot corrupt each other’s state.
Task Selection
The task loaded is controlled by theTASK_ID environment variable set on the server process. You do not pass this in the request body.
To switch tasks, set
TASK_ID=t2 (or t3) in your Docker environment or
shell before starting the server. A running server uses the value frozen at
the time of the /reset call.Request
POST /reset accepts an empty JSON body. No fields are required.
Response
The response always wraps aCRMPipelineObservation object alongside top-level convenience scalars.
Task 1 (t1) — Single Source
Task 2 (t2) — Two Sources
Task 3 (t3) — Three Sources with Conflict Rules
Response Fields
object
required
The full
CRMPipelineObservation object representing the initial environment state.float
Top-level convenience mirror of
observation.reward. Always 0.0 after reset.boolean
Top-level convenience mirror of
observation.done. Always false after reset.Extracting the Episode UUID
The episode UUID embedded inlast_action_feedback is the key you pass to /grader/{episode_id}. Parse it out immediately:
GET /state at any point during the episode: