[ 
https://issues.apache.org/jira/browse/CAMEL-23943?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Federico Mariani updated CAMEL-23943:
-------------------------------------
    Attachment: LangChain4jToolHallucinatedToolNameTest.java

> camel-langchain4j-tools: unbounded tool-calling loop, crash on hallucinated 
> tool names, tool errors swallowed
> -------------------------------------------------------------------------------------------------------------
>
>                 Key: CAMEL-23943
>                 URL: https://issues.apache.org/jira/browse/CAMEL-23943
>             Project: Camel
>          Issue Type: Bug
>          Components: camel-langchain4j-tools
>    Affects Versions: 4.21.0
>            Reporter: Federico Mariani
>            Priority: Major
>         Attachments: LangChain4jToolExecutionErrorPropagationTest.java, 
> LangChain4jToolHallucinatedToolNameTest.java, 
> LangChain4jToolUnboundedLoopTest.java
>
>
> The tool-calling loop in {{LangChain4jToolsProducer.toolsChat}} has three 
> related defects (all with attached failing reproducer tests):
> # *Unbounded loop*: the loop is {{do { ... } while (true)}} with no iteration 
> cap ({{LangChain4jToolsProducer.java:134-165}}). A model that keeps answering 
> with tool calls makes the producer loop forever. Reproducer 
> {{LangChain4jToolUnboundedLoopTest}} (mock model that always returns a tool 
> call) performed *3025 LLM round trips and 3025 tool executions in 15 seconds* 
> before the test timed out. LangChain4j itself caps this via 
> {{maxToolCallingRoundTrips}} / {{maxSequentialToolsInvocations}}; the 
> component reimplements the loop without an equivalent guard.
> # *Hallucinated tool name crashes the exchange*: the tool is resolved with 
> {{.filter(...).findFirst().get()}} 
> ({{LangChain4jToolsProducer.java:216-217}}); when the LLM hallucinates a tool 
> name the exchange dies with {{java.util.NoSuchElementException: No value 
> present}} instead of sending a corrective {{ToolExecutionResultMessage}} back 
> to the model. Reproducer: {{LangChain4jToolHallucinatedToolNameTest}}.
> # *Tool exceptions swallowed and garbage fed to the LLM*: {{invokeTools}} 
> catches route exceptions into {{toolExchange.setException(e)}} (with a 
> literal {{// How to handle this exception?}} comment, lines 267-270), then 
> sends {{toolExchange.getIn().getBody(String.class)}} (not an error message) 
> back to the LLM as the tool result, and {{ExchangeHelper.copyResults}} leaks 
> the swallowed exception onto the main exchange. Reproducer: 
> {{LangChain4jToolExecutionErrorPropagationTest}}.
> *Suggested fix*: add a {{maxIterations}} endpoint option (behavior change, 
> needs an upgrade-guide entry), return a corrective tool-result message for 
> unknown tool names, and feed tool failures to the LLM as an explicit error 
> string without failing the main exchange.
> Related: CAMEL-23928 (the agent component gained these exact knobs from 
> AiServices; this component reimplements the loop manually).
> _This issue was drafted by Claude Code on behalf of Federico Mariani._



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to