Refactor manual pydantics for scanpy pl agents #255
Conversation
…on 2.2.8 to superior or equal to 2.2.8. Indeed, it appears that the grpcio 1.53.0 external dependency of pymilvus version 2.2.8 is not compatible with Windows OS 11 and Python version 2.12.3, whatever it is the wheel or source version. Running pytest does not yield any errors, beyond raising deprecated warnings
method. Currently scanpy is imported when ScanpyTLQueryBuilder.parametrise_query is called.
Only includes functions which dont start with "_"
Just the generic function
Merge Dev/tl into main to avoid plenty of branches
* add scanpy_pl module with initial fields * add mocked test for module * add module to API agent __init__.py * add benchmark case * add conditional for module benchmark * downgrade httpx due to conflict 0.28 removed the proxy keyword, but openai is not aware * add back default `question_uuid` field into pydantic class * add scatter pydantic class * add sc.pl.pca * add pca benchmark case * distinguish web api and python api benchmark * change case to scatter * add tsne class * add tsne case * fix typing * add generic formatter (#233) * add formatter functions for REST and Python * make discoverable on module level * add required field * test the formatting functions * `scanpy` to `sc` to fit common usage * adjust benchmark to use the formatter --------- Co-authored-by: daniele-lucarelli <[email protected]>
* pushed starter anndata file * removed the tester * Aim of the anndata api module * Draft of the AnnDataIOParameters * added a prompt * updated the prompt * started to implement the AnndataIOQueryBuilder * added test for anndata api * pushed pydantic reader classes * Updated the anndata tool with integrated test: -> returns dict with method & args Co-authored-by: Anis Ismail <[email protected]> * added query builder * added querybuilder for anndata and its test * updated query builder * added exclude none * feat(BaseAPIModel): Add reusable base class for structured outputs • Introduced BaseAPIModel, a reusable base class to streamline the creation of Pydantic models for structured outputs. • The class includes: • uuid: An optional field (str | None) for unique identification of model instances. • method_name: A required field (str) to specify the associated function or method, ensuring consistency across models. • Configured with arbitrary_types_allowed to support flexible extensions. • Designed for use in structured output generation. This addition lays the groundwork for standardized, maintainable, and consistent API models. * update query builder to remove create_runnable * Updated the pydatic classes with the BaseAPIModel * Updated the system prompt in the runnable of the AnnDataIOQueryBuilder * fix in import of pydanticparser * added test for query builder parameterise_query * removed comments + redundant script --------- Co-authored-by: Anis Ismail <[email protected]> Co-authored-by: Anis Ismail <[email protected]>
…tem prompt is updated for the anndata query
… into biohackathon3
…en number of latent dimensions
replace with any length type (...)
… into biohackathon3
…c classes adjusts the ABC, the individual legacy classes (builder and fetcher), and the tests
… into biohackathon3
now has empty list in parameters
… into biohackathon3
… into biohackathon3
does not work yet; seems we now supply only metadata, no parameters
…her/biochatter into refactor_manual_pydantics
|
Still working on this. Haven't found much time this week but I hope to get it done before you want to merge the biohackathon branch into main.
|
… to how they are defined manually. Pydantic classes are then created through the BaseTools derived classes.
…QueryBuilder(scanpy.tl) to build a agent for scanpy.tl
…the automated approach can be applied
…ink a format as python call should be needed, as structued LLM output takes care of this.
|
I tried to resolve conflicts with the biohackathon main branch and it should be fine. But there are still difference in how tools are created, for example between the AnnData Builder and the others. I tried to unify the automatic generation of pydantic classes with the manual creation. The automated approach now doesn't directly generate pydnatic classes, but creates two dictionaries with parameters and function descriptions. In the manual approach (see scanpy_pl agent) these dictionaries are defined manually. The tool creation is then handled by BaseTools derived classes. |
|
Hi @mengerj, thanks for the great work! We don't need to merge this before we merge the main biohackathon branch, it is fine as a standalone feature. We'd need to make it robust though.
I would caution that this then is not compatible with the other solution. If we can, we should definitely find a common ground and consensus way of returning the parameterised call. Naively, I would say that returning the Pydantic class and then parsing independently is most flexible and does not cost anything. Your method is based on Pydantic, after all. Any reason why this is not possible? |
Refactored pydantic class definitions in pl modules to match new structure.
#245