Samuel Colvin

@sydney-runkle I'd add that ```py schema = core_schema.typed_dict_schema( { 'a': core_schema.typed_dict_field(core_schema.int_schema()), 'b': core_schema.typed_dict_field(core_schema.int_schema()), }, config=core_schema.CoreConfig(extra_fields_behavior='forbid'), ) ``` Doesn't solve the problem — with both 2.9 and 2.10 `additionalProperties` is missing.

Nice, I'm sure I can work it out from there. Thank you. Feel free to close this.

See https://github.com/pydantic/pydantic-ai/issues/116, more and more people want documentation in pure markdown to suit LLMs.

I've reviewed #3122, I think there is a much simpler and more scalable approach we should take.

Thanks @adriangb. I think sampling and memory are the most immediately problematic concerns which we need conceptual answers for before continuing work on #3122. My proposal: * can we defer...

I do think that an initial version of this should only support pre-signed URLs as described above. > some of this data may be highly sensitive (more so than other...

Currently we extract descriptions of individual parameters from the docstring by inferring the format of the docstring, then using griffe to pass the docstring. It would be easy to think...

> Can the `retriever` pattern be reused for vector search? `tool` calls already work well for vector search, see https://ai.pydantic.dev/examples/rag/. The main thing we can add is a model agnostic...

Worth saying that Jinja is inherently not type-safe, and you can often get the same result with type safe code, e.g. ```py ... return f""" blah blah blah {'\n'.join(thing.foo for...

Honestly I think #93 should cover 99% of cases.