Skip to content
KortexyaKortexyaReasoning Layer SDK

Interface: ConversationTrainingConfigDto

@kortexya/reasoninglayer

@kortexya/reasoninglayer / ConversationTrainingConfigDto

Interface: ConversationTrainingConfigDto

Section titled “Interface: ConversationTrainingConfigDto”

Defined in: src/types/rl-training.ts:45

Configuration for training on conversation data.

Remarks

Section titled “Remarks”

Used when RlTrainRequest.environmentName starts with "conversation". Controls how conversation transcripts are sourced (real or synthetic) and how each turn is rewarded.

With useRealConversations, a run trains on two sources together: the cognitive agent’s recorded conversations and the tenant’s chat conversations. conversationIds, agentId and numTranscripts select from each source. The run falls back to synthetic transcripts only when neither source has a conversation; RlTrainResponse.transcripts reports where they came from.

Serialized to the snake_case wire shape (num_transcripts, reward_step_cost, …). The engine refuses an unknown key, and refuses null for a field it defaults: leave a field out to take its default.

Example

Section titled “Example”
const config: ConversationTrainingConfigDto = {
  numTranscripts: 25,
  maxTurns: 12,
  useRealConversations: true,
  rewardUserFeedback: 0.8,
};

Properties

Section titled “Properties”

agentId?

Section titled “agentId?”

optional agentId?: string | null

Defined in: src/types/rl-training.ts:50

Keep only the conversations of this agent (UUID), in each source. Applies with useRealConversations.

conversationIds?

Section titled “conversationIds?”

optional conversationIds?: string[]

Defined in: src/types/rl-training.ts:55

Keep only these conversations (UUIDs), in each source. When empty and useRealConversations is true, the most recent conversations are used.

maxTurns?

Section titled “maxTurns?”

optional maxTurns?: number

Defined in: src/types/rl-training.ts:57

Maximum turns per episode (default: 20).

numTranscripts?

Section titled “numTranscripts?”

optional numTranscripts?: number

Defined in: src/types/rl-training.ts:63

With useRealConversations, the most conversations taken from EACH source, most recent first. Without it, or when neither source has a conversation, the number of synthetic transcripts generated (default: 10).

rewardGoalCompletion?

Section titled “rewardGoalCompletion?”

optional rewardGoalCompletion?: number

Defined in: src/types/rl-training.ts:65

Reward for goal completion (default: 1.0).

rewardStepCost?

Section titled “rewardStepCost?”

optional rewardStepCost?: number

Defined in: src/types/rl-training.ts:67

Step cost penalty (default: -0.01).

rewardSuccessfulProof?

Section titled “rewardSuccessfulProof?”

optional rewardSuccessfulProof?: number

Defined in: src/types/rl-training.ts:69

Reward for a successful proof (default: 0.3).

rewardUserFeedback?

Section titled “rewardUserFeedback?”

optional rewardUserFeedback?: number

Defined in: src/types/rl-training.ts:76

Reward scale for a user-rated chat answer (default: 0.5). An answer rated helpful earns rewardUserFeedback × satisfaction; one rated unhelpful takes the failed-goal penalty instead. Ratings come from conversation.submitAnswerFeedback.

topics?

Section titled “topics?”

optional topics?: string[]

Defined in: src/types/rl-training.ts:78

Conversation topics for synthetic transcript generation.

useRealConversations?

Section titled “useRealConversations?”

optional useRealConversations?: boolean

Defined in: src/types/rl-training.ts:85

When true, trains on real conversations instead of synthetic transcripts: the cognitive agent’s recorded conversations and the tenant’s chat conversations together, the most recent numTranscripts of each. Synthetic transcripts are only the fallback when neither source has one.