LangChain JS: .content in chain's output returns array of objects instead of string when tool call is requested

[rmbajracharya]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

const llmTool = llm.bindTools([saveNameTool);

export const myChain = PromptTemplate.fromTemplate(
`You are a friendly assistant. Your job is to guide collect the user's name.

You’ll refer to the conversation history to check if the user's **full name** has already been shared. If not, politely ask for it as part of a natural conversation.

Once the name is collected, thank the user and indicate that you’ll now assist with their inquiry.
IMPORTANT.
You have access to tool to save the user's name.
Use the available tool to save the user's name.

Conversation History:
{conversationHistory}
<question>
{question}
</question>

Keep your response short and focused.`
).pipe(llmTool);
await myChain.invoke({conversationHistory,question})

Description

I'm using LangChain JS with an LLM chain that combines a PromptTemplate and an LLM bound to a tool via .bindTools().

Everything works fine when the user input doesn't require a tool — the response from the chain is a string in .content,

However, when the input does trigger a tool call (e.g., providing the user's name), the .content field becomes an array of objects:

[
  {
    type: 'text',
    text: "I'll use the tool to save user's detail"
  }
]

This text field seems to be the model's rationale before making the tool call.

Question:
Is this expected behavior when tools are used? I do not need the model's rationale. Right now I am checking the .content field everytime, and when there is a tool call, I call the tool and send the new response rather than sending tool's rationale.

System Info

"@langchain/aws": "^0.1.8",
"@langchain/community": "^0.3.41",
"@langchain/core": "^0.3.45",
"@langchain/langgraph": "^0.2.64",

$ node -v
v20.17.0

[dosubot[bot]]

Yes, this is expected behavior in LangChain JS when tools are used. When a tool call is triggered, the .content field can become an array of objects, often including the model's rationale or other metadata before the tool is actually invoked. This is part of the standard output format for tool-enabled chains in LangChain JS, and is documented in the official docs and reflected in example code:

When using tools in LangChain JS, the output can include tool calls and content, which may be structured as an array of objects. The .content field in LangChain JS can include rationale or metadata related to tool calls, which is part of the expected output format when tools are used.
See docs

There is currently no configuration option to suppress or automatically exclude the rationale from the .content field. The recommended approach is to check the structure of .content and extract or ignore the rationale as needed in your application logic.

If you only want to send the final response after the tool call (and not the rationale), your current approach—checking .content and handling tool calls before sending a user-facing message—is correct and aligns with how others are handling this pattern in LangChain JS.

For more details and community discussion on handling tool calls and output, see this GitHub discussion.

If you have a specific output format you want, you may need to post-process the .content field to extract only the parts you need.

To continue talking to Dosu, mention @dosu-bot.

---

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