Principle:HKUDS AI Trader LLM Invocation
| Knowledge Sources | |
|---|---|
| Domains | LLM_Agents, Inference |
| Last Updated | 2026-02-09 14:00 GMT |
Overview
An inference pattern that sends conversation messages to a LangChain agent and receives responses including tool call results.
Description
LLM Invocation is the core inference step where conversation messages are sent to the LangChain agent runtime via the Runnable interface's ainvoke() method. The agent processes the messages, optionally invokes MCP tools, and returns a response containing the assistant's reply along with any tool call results. The invocation respects a recursion limit to prevent infinite tool-calling loops.
This is the fundamental building block of the agent's reasoning loop: each invocation may produce a direct response or trigger tool calls that are then fed back as additional context.
Usage
Use this principle within the agent's trading session loop whenever a new message (user query or tool result continuation) needs to be processed by the LLM. Each invocation may result in the agent calling tools, generating analysis, or outputting the finish signal.
Theoretical Basis
# Pseudocode for LLM invocation via LangChain
response = await agent.ainvoke(
{"messages": conversation_history},
{"recursion_limit": 100}
)
# response["messages"] contains:
# - AIMessage (assistant's response)
# - ToolMessage (results from any tool calls)
Key properties:
- Async: Uses ainvoke for non-blocking execution
- Recursion limit: Prevents infinite tool-calling chains
- Stateless per call: Each invocation is independent; state is managed by the caller
- Tool integration: Agent automatically invokes MCP tools when needed