davsclaus commented on code in PR #26027: URL: https://github.com/apache/camel/pull/26027#discussion_r3911035424
########## components/camel-ai/camel-langchain4j-agent/src/main/docs/langchain4j-agent-component.adoc: ########## @@ -159,6 +159,48 @@ Agents are configured using the `AgentConfiguration` class which provides a flue * Retrieval Augmentor (for RAG functionality) * Input and Output Guardrails * Concurrent tool execution (`withExecuteToolsConcurrently`) for parallel Camel route tools and MCP tools within one LLM round trip +* Tool-calling control: round-trip limits, hallucinated tool handling, and error compensation +* AiServices builder customizer for advanced LangChain4j options + +==== Tool-calling options + +`AgentConfiguration` exposes the most common LangChain4j `AiServices` tool-calling settings as first-class fluent methods. These are applied automatically when Camel creates an agent from the configuration bean (inline agent mode or a registered `AgentWithMemory` / `AgentWithoutMemory` bean): + +[cols="1,2"] +|=== +| Method | Purpose + +| `withMaxToolCallingRoundTrips(int)` | Limits how many tool-calling round trips the LLM may perform per request (`0` = unset) +| `withHallucinatedToolNameStrategy(...)` | Handles requests for tool names that do not exist (for example a hallucinated `task_complete` tool) +| `withToolExecutionErrorHandler(...)` | Converts tool execution exceptions into tool result messages for the LLM +| `withToolArgumentsErrorHandler(...)` | Handles invalid or unparsable tool arguments +| `withCompensateOnToolErrors(Boolean)` | Sends tool errors back to the LLM so it can recover instead of failing the exchange +| `withExecuteToolsConcurrently()` / `withExecuteToolsConcurrently(Executor)` | Runs multiple tool calls from one LLM turn in parallel +|=== + +._Java-only: recover from a hallucinated tool name_ Review Comment: Minor accuracy nit: this example and the `withMaxToolCallingRoundTrips`/`withCompensateOnToolErrors` rows in the table above are labeled "Java-only", but CAMEL-23929 (already on `main`) exposes these two options directly as endpoint URI params (`maxToolCallingRoundTrips`, `compensateOnToolErrors`) — documented further down in the "Concurrent tool execution" section of this same file. Worth a cross-reference or a tweak to the "Java-only" label so readers don't conclude these two are Java-only, unlike `withHallucinatedToolNameStrategy`/`withToolExecutionErrorHandler`/`withToolArgumentsErrorHandler`/`withAiServicesCustomizer`, which genuinely are (functional-interface based). Not blocking. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
