fix: improve tool type, bump pydantic and outlines

[drbh]

This PR resolves a couple

• adjusts the tool response to align with openai's tools response type
• bumps pydantic to 2.6.4 in all apps (resolves dependency issue when running tests)
• bump outlines version and fix import for new name

[drbh]

this PR resolves
#1624
#1653

[Narsil]

Some nits.

[joumenharzli]

the "name" here should be "get_current_weather"

[joumenharzli]

also here the "name" here should be "get_current_weather"

[joumenharzli]

Thanks @drbh for your PR, I've noticed that you changed the type of "tool_calls" from an object to a list, which is great.
Additionally, there's another issue where the returned function name is consistently "tools".
is this issue will be addressed in this PR?
the bug comes from here

text-generation-inference/router/src/server.rs

Line 953 in 7eb3d75

name: "tools".to_string(),

Would it be feasible to to have the tool_grammar in the format of {name:"",arguments:{}} instead of {function:{}} ?

text-generation-inference/router/src/server.rs

Line 787 in 7eb3d75

let tool_grammar = if let Some((req_tools, tool_choice)) = req.tools.zip(req.tool_choice) {

this would instruct the LLM to respond with the choosen function name and it's argument, and it can be parsed later here

text-generation-inference/router/src/server.rs

Line 948 in 7eb3d75

let tool_calls = vec![ChatCompletionMessageToolCall {

I never worked with Rust, otherwise I would subbmited a PR!

[drbh]

Hi @joumenharzli thanks for highlighting that issue! Unfortunately the current expected value is "name": "tools" because of the issue you noted above (the LLM is not specifying the function it chose in it's response).

Updating the tool's formatting is interesting, and in general the currently implementation is a bit rigid. However a change like that is outside of the scope of this PR and should be addressed in a new PR. I'll open a new issue to continue this conversation: #1657 and I'll open a new PR when we have a clear path forward

[Narsil]

LGTM