agents-core/ai.koog.agents.core.agent.session/AIAgentLLMSession

AIAgentLLMSession

sealed class AIAgentLLMSession : AutoCloseable(source)

Represents a session for an AI agent that interacts with an LLM (Language Learning Model). The session manages prompt execution, structured outputs, and tools integration.

This is a sealed class that provides common behavior and lifecycle management for derived types. It ensures that operations are only performed while the session is active and allows proper cleanup upon closure.

Inheritors

AIAgentLLMReadSession

AIAgentLLMWriteSession

Properties

model

open val model: LLModel

Represents the active language model used within the session.

prompt

open val prompt: Prompt

Represents the current prompt associated with the LLM session. The prompt captures the input messages, model configuration, and parameters used for interactions with the underlying language model.

tools

open val tools: List<ToolDescriptor>

Provides a list of tools based on the current active state.

Functions

override fun close()

requestLLM

open suspend fun requestLLM(): Message.Response

Sends a request to the underlying LLM and returns the first response. This method ensures the session is active before executing the request.

requestLLMForceOneTool

open suspend fun requestLLMForceOneTool(tool: ToolDescriptor): Message.Response

open suspend fun requestLLMForceOneTool(tool: Tool<*, *>): Message.Response

requestLLMMultiple

open suspend fun requestLLMMultiple(): List<Message.Response>

Sends a request to the language model, potentially utilizing multiple tools, and returns a list of responses from the model.

requestLLMOnlyCallingTools

open suspend fun requestLLMOnlyCallingTools(): Message.Response

requestLLMStructured

open suspend fun <T> requestLLMStructured(structure: StructuredData<T>, retries: Int = 1, fixingModel: LLModel = OpenAIModels.Chat.GPT4o): Result<StructuredResponse<T>>

Coerce LLM to provide a structured output.

requestLLMStructuredOneShot

open suspend fun <T> requestLLMStructuredOneShot(structure: StructuredData<T>): StructuredResponse<T>

Expect LLM to reply in a structured format and try to parse it. For more robust version with model coercion and correction see requestLLMStructured

requestLLMWithoutTools

open suspend fun requestLLMWithoutTools(): Message.Response