Slot-filling precision decay during multi-bot orchestration in Cognigy.AI

General Discussion
1

Hi all,

What’s the correct approach for maintaining slot-filling precision when handing off context between specialized bots in a multi-bot orchestration setup? The natural language understanding confidence scores are plummeting during the transition. This behavior mirrors the degradation patterns discussed in the “Attention is All You Need” era of transformer-based intent classification, where context window shifts disrupt entity recognition.

Environment details:

  • Platform: Cognigy.AI
  • Integration: Bot Connector API
  • Orchestration: Multi-bot handover via specialized domain bots
  • Issue: Slots are returning as null or incorrectly mapped after the second bot takes over

Option A - Manually pass the context object as a JSON string via the context property during the handover.
Pros: Ensures data persistence.
Cons: Increases payload size and risks overwriting the target bot’s local state.

Option B - Use a centralized state manager or external database to store slot values, querying them via API calls in each bot.
Pros: Decouples state from the conversational AI session.
Cons: Introduces significant latency and complicates the natural language understanding pipeline.

The logs show the handover is successful, but the slot-filling precision for the target bot is essentially zero.

{
 "intent": "UpdateAccount",
 "confidence": 0.42,
 "slots": {
 "accountNumber": null,
 "customerName": null
 }
}
1 Like
2 
{
 "context": {
 "slot_filling": "enabled",
 "preserve_context": true
 }
}

Are the slots being passed in the context object or via a custom payload? Cognigy’s docs on Context Management say that “context is shared across all bots in the orchestration”, but I’ve seen weird latency spikes in the US-East region that mess with the handoff timing. Spun up a quick Lambda to patch this once by forcing a state sync.