langgraph/how-tos/tool-calling-errors/

[giscus[bot]]

langgraph/how-tos/tool-calling-errors/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/tool-calling-errors/

[zcgnull]

My prebuilt ToolNode did not catch any errors.

[hinthornw]

What version did you use?

[zcgnull]

[worldback]

In the recommendation scenario, for example, if a user asks "Do you have a cup?", this will start a SQL tool call, and then the agent will give a reply based on the tool call results. But if the user asks again "No patterned ones", can the agent understand that the parameters of the SQL tool call are updated based on the previous conversation? In other words, in the next round of conversation, can the LLM see the tool call information of the previous round or the input and output of any intermediate steps?

[bnbabu55]

https://langchain-ai.github.io/langgraph/concepts/low_level/#messagegraph message graph is your answer I think, it is used for chatbot scenarios.

[jojohannsen]

great example, switching to more capable model on failure, very nice

[pen-ho]

me too

[liondr123]

Using with chat models:

llm = ChatOpenAI(
model_name="meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
base_url = "https://api.aimlapi.com/v1",
api_key = "c7ddab3d0e964d11a8d9e8fd085066fa",
temperature=0
)

tools = [get_weather, get_coolest_cities]
tool_node = ToolNode(tools)
llm_with_tools = llm.bind_tools(tools)

llm_with_tools.invoke("what's the weather in sf?").tool_calls

tool_calls return nothing: []

[pen-ho]

from langchain_openai import ChatOpenAI
model_with_tools = ChatOpenAI(
# model="chinese-llama-alpaca-2",
openai_api_key="EMPTY",
openai_api_base="http://192.168.xx.xx:xxxx/v1",
max_tokens=512,
temperature=0
).bind_tools(tools)

llm_with_tools.invoke("what's the weather in sf?").tool_calls

tool_calls return nothing: []

version:
langgraph 0.2.15
langchain-openai 0.1.23

[ywhijn]

how can I use complex tools defined as a class inherited from BaseTool? I have 'environment' as an instance initialized ahead and used in my tool node.

[itsihsan]

How can we use the MapReduce approach here? For instance, when asked, "What's the weather in the coolest cities?" and NYC and SF are identified as the coolest cities, the get_weather tool fetches the weather for each city sequentially. Since these tasks are independent, can we parallelise this process using MapReduce?

[hinthornw]

ToolNode execetues tools in parallel already (assuming the model has invoked both in a single forward pass.

[itsihsan]

Thank you. I just referred to the documentation for the ToolNode. How can we automate passing multiple tool calls to AIMessage's tool_call parameter? Please refer to the text 'You can also do parallel tool calling using ToolNode if you pass multiple tool calls to AIMessage's tool_calls parameter' in 'Manually call ToolNode' section of https://langchain-ai.github.io/langgraph/how-tos/tool-calling/#define-tools

[vbarda]

@itsihsan this is automatically done by the chat model you're using (i.e. ChatOpenAI / ChatAnthropic in LangChain), and is supported by largest providers like OpenAI and Anthropic. When the model has access to multiple tools, it can automatically choose to provide multiple values for the tool_calls field, when necessary. Additionally, with OpenAI models you can also explicitly control whether you have parallel tool calling enabled or not (via ChatOpenAI().bind_tools(tools, parallel_tool_calls=False))

[itsihsan]

Thank you @vbarda. I verified that using traces.

[BNN0]

Can i use two nodes for manage my tools and how? For example use one toolNode for weather tool and other for the list. I tried to do this but when i execute the code I only got a bunch of xml tags and not content of the tool.