[Feature] Bump VizroAI to pydantic V2

[maxschulz-COL]

Description

[Update]

• We agreed on finishing pydantic v2 migration by still using langchain with_structured_output.
• Keep the door open for eventually switching to PydanticAI when the timing is right. (Also excited to see the support for structured output is on PydanticAI's roadmap: PydanticAI Roadmap pydantic/pydantic-ai#913)
• The good news is that now apart from OpenAI models, claude and gemini (possibly other vendors too but I didn't try) can work with the dashboard/plot tasks fairly well too.

---

This is a draft of a first working version, will hand off to you @lingyielia now, because you are probably better positioned to do the fine-tuning. I also left one todo for you @nadijagraca , maybe you already have some time this week?

Overall considerations

• PydanticAI
  • using tool_calling only - hence it doesn’t use structured_output , that’s why it doesn’t fail
• Langchain
  • gives the choice (for OpenAI only) to use structured_output or tool_calling
  • structured_output doesn’t work with pattern
    • https://platform.openai.com/docs/guides/structured-outputs/some-type-specific-keywords-are-not-yet-supported#supported-schemas
    • pydanticAI also discussing this
      • Structured outputs as an alternative to Tool Calling pydantic/pydantic-ai#582 (comment)
      • Weird tool call arguments, resulting in UnexpectedModelBehaviour / validation error pydantic/pydantic-ai#81 (comment)
  • hence we need to use function calling anyway
• the question is what framework is better?
  • running gpt-40 and max_retry=1 on both frameworks with a reasonably complex dashboard request to produce the plan suggests that in PydanticAI it gets it while in Langchain it doesn’t - even on 3 retries langchain doesn’t get it
  • maybe due to: “If the result type is a union with multiple members (after remove str from the members), each member is registered as a separate tool with the model in order to reduce the complexity of the tool schemas and maximise the chances a model will respond correctly.”
    • I do not think this is the case for us as we do not directly provide a union
• trying on vm.Dashboard
  • PydanticAI
    • seems to be able to do it with latest anthropic model (claude-3-5-sonnet-latest) but not latest OpenAI model (gpt-40) apart from the impossible things
  • Langchain
    • behaves like pydanticAI

--> Ergo overall I am still confident in us changing to pydanticAI - but I left the core code in in case you disagree @lingyielia .

ToDos

• Finish all TODOs in the code marked as TODO[LZ] or TODO[NG] respectively
• Delete remainders of langchain if we decide against it
• Change examples, documentation and code for new way of configuring models
• Fix the remaining unit tests

Screenshot

Notice

• I acknowledge and agree that, by checking this box and clicking "Submit Pull Request":

  • I submit this contribution under the Apache 2.0 license and represent that I am entitled to do so on behalf of myself, my employer, or relevant third parties, as applicable.
  • I certify that (a) this contribution is my original creation and / or (b) to the extent it is not my original creation, I am authorized to submit this contribution on behalf of the original creator(s) or their licensees.
  • I certify that the use of this contribution as authorized by the Apache 2.0 license does not violate the intellectual property rights of anyone else.
  • I have not referenced individuals, products or companies in any commits, directly or indirectly.
  • I have not added data or restricted code in any commits, directly or indirectly.

[github-actions]

View the example dashboards of the current commit live on PyCafe ☕ 🚀

Updated on: 2025-02-24 17:44:26 UTC
Commit: 686477b

Compare the examples using the commit's wheel file vs the latest released version:

vizro-core/examples/scratch_dev

View with commit's wheel vs View with latest release

vizro-core/examples/dev/

View with commit's wheel vs View with latest release

vizro-core/examples/visual-vocabulary/

View with commit's wheel vs View with latest release

vizro-ai/examples/dashboard_ui/

View with commit's wheel vs View with latest release

[lingyielia]

langchain_openai with_structured_output supports "function_calling" and "json_schema".

• before langchain 0.3.0, method defaults to "function_calling", so that's why we didn't get validation error on "pattern".
• after langchain 0.3.0, method defaults to "json_schema". (see https://python.langchain.com/api_reference/openai/chat_models/langchain_openai.chat_models.base.ChatOpenAI.html#langchain_openai.chat_models.base.ChatOpenAI.with_structured_output)

This means if we now use

llm_model.with_structured_output(
                response_model,
                method="function_calling",
            )

it's equivalent to what we did before pydantic v2 migration.

If this conclusion is true, I'd prefer to first ship pydantic v2 migration code along with .with_structured_output(method="function_calling") in this PR. And we open another PR to do the pydanticAI migration (this task is time consuming, with a lot of docs change and unit tests change if I'm correct)

Let know wdyt @maxschulz-COL :)

[maxschulz-COL]

⭐ Looks good to me! ⭐

Thanks for partially reverting stuff, and great inquisition as to the default behaviour before 0.3.0!

[lingyielia]

looks good on my side now. @maxschulz-COL may I have another review from you?

[maxschulz-COL]

Nice!

[nadijagraca]

Looks good. Great job. 🚀

I noticed just small visual issue in the chart UI (not caused by this PR) that I will fix in a separate small PR this week.