Skip to main content

Supervised Pipeline

A supervised pipeline adds a Supervisor node to which several AI Agent nodes set to "supervised" mode are connected as subagents. This is the right structure when several specialized agents may produce overlapping or contradictory annotations, and you need one consistent final result rather than a plain concatenation.

What happens at execution time​

  1. Each agent, as a subagent, analyzes the document independently and in parallel — this step is deterministic, no extra AI call is involved.
  2. The supervisor groups the produced annotations by their position in the document (overlapping passages) — again, no AI call.
  3. A group containing a single annotation (no conflict) is kept as-is, with no further AI call.
  4. A group containing several conflicting or overlapping annotations triggers one single arbitration AI call per group, which keeps, merges or drops the annotations involved — never inventing new ones.

The total number of arbitration calls is capped (20 by default, configurable): beyond that, the remaining groups are kept unchanged rather than multiplying AI calls.

Supervisor node settings​

  • Arbitration AI model — used only to settle conflicts, independent of the models used by the agents it supervises; it is recommended to use one at least as capable as them.
  • Severity:
    • Conservative (default) — only drops an annotation in case of a demonstrable false positive or an exact duplicate; keeps it when in doubt.
    • Strict — also arbitrates partial overlaps between agents (keeps the more precise one, or merges them); still keeps it when genuinely in doubt.
  • Maximum number of arbitrations — 20 by default, 0 = unlimited.
  • Additional instructions — optional free text to guide arbitration.

Each retained annotation keeps track of the agent that originally produced it, which preserves the manager validation loop described in AI Agents.