Outdated dependency versions are blocking adoption

[CaliLuke]

The current dependency pins in pgai are quite outdated, which creates conflicts when trying to integrate pgai into projects that use newer versions of common AI/ML libraries.

For example:

• litellm is pinned to >=1.65.0,<1.73.0 (current: 1.81.5)
• openai is pinned to >=1.44,<2.0 (current: 2.16.0)
• ollama is pinned to >=0.4.5,<0.6.0 (current: 0.6.1)
• tiktoken is pinned to >=0.7,<1.0 (current: 0.12.0)
• anthropic is pinned to 0.49.0 (current: 0.77.0)

This makes it difficult to use pgai alongside other libraries that depend on newer versions of these packages.

We've tested with relaxed version constraints (removing upper bounds) and observed no major issues - the unit tests pass and imports work correctly with the latest versions. Are there plans to bump these dependencies? What's blocking the upgrade?

[Askir]

Since you mention anthropic I assume you mean the pgai postgres extension and not the corresponding python library.
However I am confused why this would conflict with any other dependencies. We include pgai in multiple docker images and are pinning the deps to have a consistent build process. We could upgrade them but we do not want to have non deterministic builds.

Are you running other pl/python3u extensions on your postgres instance which have other requirements and therefore have conflicts? If yes could you link what else you are installing? Maybe we can figure out where the conflicts are and resolve them.