Skip to content

ONDA Overview

ONDA is the core paradigm behind Plugy. It treats the chatbot not as a stateless request-response system but as a coherent observer that reads context, reasons, retrieves knowledge, responds, and then learns from the outcome.

Traditional chatbots are “thermometers” — they measure and forget. ONDA builds “thermostats” — systems that measure, act, observe the result, and modify themselves.

Every incoming message passes through four tacts:

TactNameModeWhat Happens
1SCANExpansionLoad memory (recent + summary + profile), extract session facts
2THINKCompressionFast model analyzes emotion, intent, tone, topic
3RETRIEVEExpansionSemantic search retrieves relevant knowledge base entries
4RESPONDCompressionPrimary model generates the response using enriched context

After the response is sent, asynchronous tasks update memory and compute B-scores.

B-score measures the quality of each bot response as a weighted geometric mean of four components:

ComponentSymbolMeasures
FocusFRelevance of the answer to the user’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 single component is zero, the entire B-score collapses to zero.

Before generating a response, the RESPOND tact applies four verification checks:

  • [A_F] — Is my response focused on the client’s actual need?
  • [A_E] — Does my tone match the recommended tone from THINK analysis?
  • [A_sigma] — Does my response contradict any known facts from client context?
  • [A_Lambda] — Am I using relevant past resolution patterns from the knowledge base?
LevelScopeMechanism
ClientSingle conversationRecent message history, session facts
ProjectOne bot instanceConversation summaries, persona optimization, knowledge rules
GlobalCross-projectClient profiles, cross-project knowledge sharing

The self-learning system forms a feedback loop: the bot generates responses, the ONDA system observes those responses, measures B-scores, proposes improvements, and feeds improved configurations back into the bot. The observer observes itself observing.

See Self-Learning for how ONDA improves your bot over time.