community: Implement bind_tools for ChatTongyi

[cheese-git]

Description

Implement bind_tools in ChatTongyi. Usage example:

from langchain_core.tools import tool
from langchain_community.chat_models.tongyi import ChatTongyi

@tool
def multiply(first_int: int, second_int: int) -> int:
    """Multiply two integers together."""
    return first_int * second_int

llm = ChatTongyi(model="qwen-turbo")

llm_with_tools = llm.bind_tools([multiply])

msg = llm_with_tools.invoke("What's 5 times forty two")

print(msg)

Streaming is also supported.

Dependencies

No Dependency is required for this change.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	May 7, 2024 5:22pm

[ccurme]

are these tool calls fully formed in a streaming context?

if not you can specify tool_call_chunks on AIMessageChunk instead, with args a (partial json) string. See example here:

langchain/libs/partners/openai/langchain_openai/chat_models/base.py

Line 264 in c010ec8

tool_call_chunks=tool_call_chunks,

[cheese-git]

Thanks a lot, I have made some updates about this.

[ccurme]

this is confusing, because in convert_dict_to_message we are parsing tool calls (e.g., in call to parse_tool_call) but here we are passing them in although they are already parsed.

at this point, is the following true?

if message_chunk.additional_kwargs["tool_calls"]:
    item = message_chunk.additional_kwargs["tool_calls"][0]

assert isinstance(item["args"], dict)  # not a string

[cheese-git]

Sorry for delayed response, having a really tough time with my job 😫
This was a mistake, I've rewritten the method referring langchain_openai's base module.

[cheese-git]

I also did some refactoring.

The main challenge of supporting tools for Tongyi, is that Tongyi does not support using tools with incremental_output=True yet, which means every new chunk contains the content of previous chunks. However, some internal methods of Langchain assume the opposite.

That's why I end up writing a subtract_client_response method, to chop off duplicated content when streaming without incremental_output=True.

[ccurme]

@cheese-git would this represent a breaking change for streaming generally?

[cheese-git]

No, it would not.

Before this PR, incremental_output is always True when streaming:

langchain/libs/community/langchain_community/chat_models/tongyi.py

Lines 416 to 417 in 64c4722

	if params.get("stream"):
	params["incremental_output"] = True

To support tools when streaming, I updated the condition to this:

# According to the Tongyi official docs,
# `incremental_output` with `tools` is not supported yet
if params.get("stream") and not params.get("tools"):
    params["incremental_output"] = True

So "streaming without incremental_output being True" is a new circumstance.

[Arunshmily]

I Use ChatTongyi to bulid a Agent.
Which outputparser I could use?
ToolsAgentOutputParser() could not available to ChatTongyi.
LangChain-core==0.2.1

[jlcoo]

from langchain_core.tools import tool
from langchain_community.chat_models.tongyi import ChatTongyi

@tool
def multiply(first_int: int, second_int: int) -> int:
    """Multiply two integers together."""
    return first_int * second_int

llm = ChatTongyi(model="qwen-turbo")

llm_with_tools = llm.bind_tools([multiply])

msg = llm_with_tools.invoke("What's 5 times forty two")

print(msg)

this code can't call multiply in langchain 0.2.6?