Did you try pyright without strict?

@alejandrodnm without strict it looks a lot better:

4 errors, 0 warnings, 0 informations

But this allows untyped functions and Any returns which essentially allows circumventing the type checker altogether. Not great imo.

For reference this is how it looks with strict:

301 errors, 0 warnings, 0 informations

The main problem is that some libraries in the python ecosystem still don't have types (noteably asyncpg and pgvector).
For mypy I can set ignore_missing_imports = true, this then means the results are treated as Any.
This causes an error when I return an Any value from a function. I can however mark that return as #type: ignore and after that mypy "trusts" my annotation.

So for Asyncpg I (or rather Claude) typed the functions as returning an asyncpg.Record and simply ignored the error that Any is not a Record.

Pyright works differently, I can ignore the error in that location but it doesn't trust my annotation afterwards it instead build a union type Record | Unknown where Unknown is equivalent to mypy untyped Any. This then throws errors in the rest of the codebase because of course I handle it as if it was a Record everywhere.

I haven't really figured out how to change this behaviour, there doesn't seem to be away to say: Trust my annotation. The only recommendation that pyright has for untyped third party libraries is to write stub files yourself, which works but honestly is a bit too much effort right now. Asyncpg e.g. has a PR to add types since 4 years here.

I get that in pgai we use pyright and would need to use the same typechecker. It just seems that once you have some sort of untyped dependencies the pyright strict mode no longer really works. Kinda sucks 🤷‍♂️

I've fixed the typing issues by providing stubs for everything in use. Took a while but with Claudes help it was doable. So now we can use pyright in strict and dont have to change anything about the config for pgai. @alejandrodnm