anthropics / anthropic-tools Goto Github PK

Please add plugin, custom action like functionality so we can build external tooling on Claude platform. Our users are asking for it to access their long term memory and using it inside Claude.AI .

Regards,
Ozgur
Co-Founder of Keymate.AI

Has Claude 3 been fine-tuned on this XML-based function calling format, or is it simply a matter of in-context learning from the prompt?

Trying to see if it's worth building an XML -> JSON (Zod) layer, or if I can just in-context use JSON-Schema.

@nmarwell-anthropic hey do we have a ts version of this. I would love to create one!!

Running the file throws the error

line 7, in <module> from .tools.base_tool import BaseTool ImportError: attempted relative import with no known parent package.

Steps to reproduce:

1. Clone the repository
2. Run calculator_example.py inside tool_use_package

Hi Anthropics!

Great work with the function calling support! We want to evaluate the function calling performance of Claude-3 and are wondering:

1. In the current parsing schema from JSON, there isn't anywhere where we can specify whether the parameters are optional or not. Would this feature be supported in near future? Or we can actually specify in XML format now?
2. When it comes to dictionary/Map and List/Array parameters. The current schema does not support specifying element types in dictionary or list. In another word, How are we able to formulate function parameters that require nested documentations.

Thank you!

Fanjia

Amazon Bedrock Claude3 Sonnet model support request

It seems there's a mismatch between the description of function calling in Anthropic docs and this library's implementation:

• As of the time of filing the bug, the docs say to provide the function call results in a new role=user message: https://web.archive.org/web/20240324060505/https://docs.anthropic.com/claude/docs/functions-external-tools#example-tool-use-prompt
• But this library instead keeps appending the results to the preexisting role=assistant message.

Since this is an LLM I'm sure it will approximately understand both ways of providing the results, but it would be great if you could clarify which one's officially correct.

Thanks!

For most chatbot applications, it's important to stream the response, especially in cases where a tool isn't actually being used.

Not just that, but it's also important to know whether a tool is being used, or whether the bot is just being slow to respond.

Either way, the user will be waiting (in complex cases this might take several minutes) without knowing what's going on.

The bot might be generating a function call, or generating a message. As far as I can tell, when we call use_tools we don't know which of those 2 is happening until it's done, so we can't show the user what is happening.

And more importantly, if claude is generating a normal message, we can't stream the response to the user so they can read it as it is being created instead of waiting and then having to read it all at once. This costs the user lots of time overall.

If i'm wrong please let me know, otherwise I would say this is quite an important feature request, and I'll use the XML in the meantime.