[Platform] Add tooling support for ollama to allow agent and toolbox usage

[JoshuaBehrens]

Q	A
Bug fix?	no
New feature?	yes?
Docs?	no
License	MIT

Running a toolbox agent with ollama was not yet supported. This ensures, that tool messages and responses work with a locally hosted ollama instance. It seems not to be a fix as it was not meant to work and it is not really a mentionable feature as one could expect it to work like the others so I am not really sure to work where to add more content for humans. I tested it with the given new example in examples/ollama/toolcall.php.

A current design decision had me struggling a bit. I was not yet able to solve in a good way the fact, that ollama itself cannot claim every model to have tool support. So someone who uses a model needs to know before its usage whether it has model support. This can be queried easily but the capabilities are part of the model, that you need to create before using the platform to query in the name of the model. So this is like hen-egg. If the model itself would not know its capabilities but there is a service, that has model and platform at hand, could evaluate it (and cache it). Is this something we want to establish?

Possible query:

curl http://localhost:11434/api/chat -d '{
  "model": "llama3:8b-instruct-q6_K",
  "tools": [{}],
  "stream": false
}'
{"error":"llama3:8b-instruct-q6_K does not support tools"}

[chr-hertel]

I general this looks pretty clean and not like your first contribution here - thanks kicking this off - great work!

I made a few comments - can see where your thoughts are coming from - this is between those axis in Platform - working with two bridges.

And rolling out tool calling support to more bridges is def a new feature 👍

[JoshuaBehrens]

I general this looks pretty clean and not like your first contribution here - thanks kicking this off - great work!

ty :) likely not my last one. already looking into ollama embeddings

And rolling out tool calling support to more bridges is def a new feature 👍

anything more you need to have written out for a zero-ver? :D

[chr-hertel]

Please rebase on main and adopt the changed test style, switching from #[Test] attribute to test method prefix, see #214.

[JoshuaBehrens]

I rebased, changed the test and renamed the class. Anything more :) ?

[chr-hertel]

This! Looks! Great! :)

I believe your regex model work deserves some simple unit testing, but will give it a local test already

[chr-hertel]

Tested with mistral, deepseek and llama - works 👍

[JoshuaBehrens]

Great I will fix phpstan, rebase and we can get this merged. The new pull request for the model usage can follow after that :)

[chr-hertel]

And that unit test for Symfony\AI\Platform\Bridge\Ollama\Ollama please :)

[JoshuaBehrens]

Sorry for the repetetive pushes. fabpot had some remarks I was not aware of. Especially this is confusing in the config:


At one part the suggestion is to have a FQCN on global functions and in another place they don't.

Apart from that I should not have any open tasks on this pull request

[OskarStark]

At one part the suggestion is to have a FQCN on global functions and in another place they don't.

We only use the one which have an optimized usage afterwards, but yes, sometimes its hard to guess :-)

[chr-hertel]

Thank you @JoshuaBehrens.

[JoshuaBehrens]

Thank you for the good exchange and merge :) next pull requests are in the making