@langchain/google-genai does not properly handle MALFORMED_FUNCTION_CALL

[jianmin-chen]

Checked other resources

• I added a very descriptive title to this issue.
• I searched the LangChain.js documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain.js rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

try {
  response = await modelWithTools.invoke([
    new SystemMessage(actionPrompt(mappedChatHistory, toolsToUse)),
    new HumanMessage(mappedChatHistory)
  ]);
} catch (err) {
  sparrowMonitor.error(`Ran into error while sampling, try ${tries}`, err);
}

Error Message and Stack Trace (if applicable)

I have an error logger wrapper - this is the stack it returns:

[12:11:32.944] ERROR: Ran into error while sampling, try 0 {"reqId":"ITwa6etzSrhGZdlyILjNL","reqPath":"POST /api/v1/_internal/benchmark/actions","reqUrl":"POST /api/v1/_internal/benchmark/actions","levelLabel":"ERROR"}
    err: {
      "type": "TypeError",
      "message": "Cannot read properties of undefined (reading 'length')",
      "stack":
          Ran into error while sampling, try 0: Cannot read properties of undefined (reading 'length')
              at mapGenerateContentResultToChatResult (/Users/jc/Documents/node_modules/.pnpm/@langchain+google-genai@0.1.11_@langchain+core@0.2.26_openai@4.73.0_zod@3.23.8___zod@3.23.8/node_modules/@langchain/google-genai/dist/utils/common.cjs:229:32)
              at ChatGoogleGenerativeAI._generate (/Users/jc/Documents/node_modules/.pnpm/@langchain+google-genai@0.1.11_@langchain+core@0.2.26_openai@4.73.0_zod@3.23.8___zod@3.23.8/node_modules/@langchain/google-genai/dist/chat_models.cjs:661:87)
              at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
              at async Promise.allSettled (index 0)
              at async ChatGoogleGenerativeAI._generateUncached (/Users/jc/Documents/node_modules/.pnpm/@langchain+core@0.2.26_openai@4.73.0_zod@3.23.8_/node_modules/@langchain/core/dist/language_models/chat_models.cjs:176:29)
              at async ChatGoogleGenerativeAI.invoke (/Users/jc/Documents/node_modules/.pnpm/@langchain+core@0.2.26_openai@4.73.0_zod@3.23.8_/node_modules/@langchain/core/dist/language_models/chat_models.cjs:60:24)
              at async takeSample (/Users/jc/Documents/apps/sparrow/src/controllers/api/v1/_internal/benchmarkController.ts:184:24)
              at async actionsBenchmark (/Users/jc/Documents/apps/sparrow/src/controllers/api/v1/_internal/benchmarkController.ts:233:16)
      "name": "Ran into error while sampling, try 0"
    }

Description

I'm inclined to believe that there's a bug in @langchainjs/google-genai that results in it not being able to properly handle MALFORMED_FUNCTION_CALL outputs from the Google Gemini API, specifically these lines in src/utils/common.ts:

if (candidateContent?.parts.length === 1 && candidateContent.parts[0].text) {

candidateContent appears to not always contain keys based on debugging, but optional chaining will return undefined:

{
  candidates: [
    {
      content: {
      },
      finishReason: "MALFORMED_FUNCTION_CALL",
    },
  ],
  usageMetadata: {
    promptTokenCount: 4512,
    totalTokenCount: 4512,
    promptTokensDetails: [
      {
        modality: "TEXT",
        tokenCount: 4512,
      },
    ],
  },
  modelVersion: "gemini-2.0-flash",
  text: () => {
    if (response.candidates && response.candidates.length > 0) {
        if (response.candidates.length > 1) {
            console.warn(`This response had ${response.candidates.length} ` +
                `candidates. Returning text from the first candidate only. ` +
                `Access response.candidates directly to use the other candidates.`);
        }
        if (hadBadFinishReason(response.candidates[0])) {
            throw new GoogleGenerativeAIResponseError(`${formatBlockErrorMessage(response)}`, response);
        }
        return getText(response);
    }
    else if (response.promptFeedback) {
        throw new GoogleGenerativeAIResponseError(`Text not available. ${formatBlockErrorMessage(response)}`, response);
    }
    return "";
  },
  functionCall: () => {
    if (response.candidates && response.candidates.length > 0) {
        if (response.candidates.length > 1) {
            console.warn(`This response had ${response.candidates.length} ` +
                `candidates. Returning function calls from the first candidate only. ` +
                `Access response.candidates directly to use the other candidates.`);
        }
        if (hadBadFinishReason(response.candidates[0])) {
            throw new GoogleGenerativeAIResponseError(`${formatBlockErrorMessage(response)}`, response);
        }
        console.warn(`response.functionCall() is deprecated. ` +
            `Use response.functionCalls() instead.`);
        return getFunctionCalls(response)[0];
    }
    else if (response.promptFeedback) {
        throw new GoogleGenerativeAIResponseError(`Function call not available. ${formatBlockErrorMessage(response)}`, response);
    }
    return undefined;
  },
  functionCalls: () => {
    if (response.candidates && response.candidates.length > 0) {
        if (response.candidates.length > 1) {
            console.warn(`This response had ${response.candidates.length} ` +
                `candidates. Returning function calls from the first candidate only. ` +
                `Access response.candidates directly to use the other candidates.`);
        }
        if (hadBadFinishReason(response.candidates[0])) {
            throw new GoogleGenerativeAIResponseError(`${formatBlockErrorMessage(response)}`, response);
        }
        return getFunctionCalls(response);
    }
    else if (response.promptFeedback) {
        throw new GoogleGenerativeAIResponseError(`Function call not available. ${formatBlockErrorMessage(response)}`, response);
    }
    return undefined;
  },
}

So it'll return Cannot read properties of undefined. It would be nice to have better error output, or have Langchain handle it. The Mistral integration for example returns Failed to call a function which is a lot more useful.

For context I'm sampling and benchmarking the Google Gemini models and this is a bug that I believe appears with all of them.

System Info

Appropriate Langchain package versions in package.json:

"@langchain/anthropic": "^0.1.21",
"@langchain/community": "^0.2.2",
"@langchain/core": "0.2.26",
"@langchain/google-genai": "^0.1.11",
"@langchain/groq": "^0.1.2",
"@langchain/openai": "^0.3.16",
"@langchain/xai": "^0.0.1",
"langchain": "^0.3.8",

[dosubot]

Hello @jianmin-chen!

To resolve the "Cannot read properties of undefined (reading 'length')" error in @langchain/google-genai when handling MALFORMED_FUNCTION_CALL outputs from the Google Gemini API, ensure that the candidates array in the EnhancedGenerateContentResponse is properly checked for existence and length before accessing its elements. The error likely occurs because the code attempts to access the candidates array without verifying its presence or length.

In the function mapGenerateContentResultToChatResult, the code should check for the existence of response.candidates and its length:

if (
  !response.candidates ||
  response.candidates.length === 0 ||
  !response.candidates[0]
) {
  return {
    generations: [],
    llmOutput: {
      filters: response.promptFeedback,
    },
  };
}

Ensure that similar checks are in place wherever the candidates array is accessed. This will prevent attempts to read properties from an undefined object, which is the cause of the error ^[1].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}

[jianmin-chen]

Yeah, a PR to fix this would just be a minor addition to that if statement. Happy to do that if my intuition is correct and I'm not just overlooking something.