Перейти к содержимому

Self-Learning

Это содержимое пока не доступно на вашем языке.

ONDA (Observer-Native Dialog Architecture) is Plugy’s self-learning system. It continuously improves your bot’s performance by analyzing conversations, proposing optimizations, and safely applying changes.

The ONDA system runs automatically on a schedule (default: daily). Each iteration follows this pipeline:

  1. Benchmark — Run test dialogs against your current bot configuration
  2. Analyze — Identify the weakest quality dimension (Focus, Empathy, Consistency, or Experience)
  3. Propose — Use AI to generate an improvement targeting the weak dimension
  4. Re-benchmark — Test the proposed change against the same dialogs
  5. Safety check — A guardian system verifies the change is safe and improves quality
  6. Apply or revert — Only apply changes that pass all safety checks

The self-learning system operates across multiple levels, each optimizing a different aspect of your bot:

LevelWhat It OptimizesDescription
L1PersonaYour bot’s personality, tone, and communication style
L2Knowledge RulesIndividual knowledge base entries and responses
L3RetrievalHow your bot searches and selects relevant knowledge
L4Model RoutingWhich AI model handles which type of question
L5Meta-OptimizerThe optimization parameters themselves
L9Cross-ProjectKnowledge sharing between your projects

When one level converges (no more improvements possible), the system advances to the next level.

The ONDA Guardian system ensures changes never degrade your bot:

  • Quality floor — Responses must maintain a minimum quality score
  • Component check — No individual quality dimension can drop below a threshold
  • Rollback — Every change can be instantly reverted
  • Auto-halt — After 3 consecutive rejected changes, the system pauses and waits for review

Track your bot’s improvement in the dashboard:

  • View iteration history with before/after quality scores
  • See which changes were approved or rejected and why
  • Monitor the current ONDA level and optimization target
  • Review the B-score trend over time

The B-score is Plugy’s quality metric, computed as a weighted geometric mean of four components:

ComponentSymbolWhat It Measures
FocusFRelevance of the answer to the customer’s question
EmpathyEAlignment with instructions, emotional attunement
Consistency1-sigmaNon-contradiction with prior answers and knowledge base
ExperienceLambdaUse of conversation history and feedback

Weak-link rule: If any component is zero, the entire B-score collapses to zero.