feat: add support for complex types in component processing #9305
Conversation
haystack/tools/component_tool.py
Outdated
| f"component's run method." | ||
| ) | ||
| elif python_type is dict or (origin is dict): | ||
| schema = {"type": "object", "description": description, "additionalProperties": True} |
There was a problem hiding this comment.
@anakin87 do you happen to know if by adding additionalProperties here we run the risk of breaking ChatGenerators that don't support it?
There was a problem hiding this comment.
I'm not sure - it might be better to avoid it if not strictly necessary
I found this: https://platform.openai.com/docs/guides/structured-outputs/additionalproperties-false-must-always-be-set-in-objects
Anyway, it might be a good idea to involve @vblagoje who worked much on this, if we are not in a rush
There was a problem hiding this comment.
+1 on the structured output or if you set the JSON response to strict. But that might be a separate issue we need to open where if structured output or strict is set in OpenAIChatGenerator then it should probably go in and update all additionalProperties to false. I think this would need to be in the ChatGenerator since the ComponenTool (or any Tool) can't know yet if additionalProperties must be set to False.
haystack/tools/component_tool.py
Outdated
| f"Pydantic models (e.g. {python_type.__name__}) are not supported as input types for " | ||
| f"component's run method." | ||
| ) | ||
| elif python_type is dict or (origin is dict): |
There was a problem hiding this comment.
Since we are handling dict here now could we also remove dict from the _create_basic_type_schema function?
| # Check the parameter schema | ||
| assert "meta" in tool.parameters["properties"] | ||
| assert "extraction_kwargs" in tool.parameters["properties"] | ||
|
|
||
| # Meta should be oneOf with both object and array options | ||
| meta_schema = tool.parameters["properties"]["meta"] | ||
| assert meta_schema["description"].startswith("Optional metadata") | ||
| assert "oneOf" in meta_schema | ||
|
|
||
| # extraction_kwargs should be an object | ||
| kwargs_schema = tool.parameters["properties"]["extraction_kwargs"] | ||
| assert kwargs_schema["type"] == "object" | ||
| assert "additionalProperties" in kwargs_schema | ||
| assert kwargs_schema["additionalProperties"] is True |
There was a problem hiding this comment.
I appreciate the additional comments and individual asserts, but could we move update this to a full dict comparison. So
assert tool.parameters["properties"] == {...}I think that would make it easier to judge at a glance.
There was a problem hiding this comment.
I think it is better for readability. But no problem, I will do it
Co-authored-by: Sebastian Husch Lee <sjrl@users.noreply.github.com>
| if origin is dict and args and args[0] is str and args[1] is Any: | ||
| if description and "meta" in description.lower(): | ||
| schema = {"type": "string", "description": description} | ||
| else: | ||
| schema = {"type": "object", "description": description, "additionalProperties": True} |
There was a problem hiding this comment.
Out of curiosity why do we need this special case?
There was a problem hiding this comment.
Good question! The reason we single out pure Dict[str, Any] parameters whose description mentions “meta” is that metadata blobs by definition have an arbitrary, unknown shape, and we don’t want to bake an open-ended object schema (with unpredictable keys) into our function spec.
- If I treated metadata as a normal object with
additionalProperties: true, undertools_strictI would have to enumerate every possible metadata key (which we can’t know ahead of time), and LLMs often struggle to fill such free-form nested schemas. - All other dicts (that aren’t metadata) still fall back to an object schema with
additionalProperties: true.
|
@sjrl — Could you please take a look and let me know if this now aligns with your requirements? I’m happy to tweak it further if needed. |
|
Hey @YassinNouh21 thanks for your work on this! However, in the end we opted to go with a different approach outlined here #9342 which uses pydantic to handle the automatic schema generation rather than reimplementing this ourselves. |
3 participants
Related Issues
parametersinComponentTooldoes not work for more complex types #9292Proposed Changes:
The
ComponentToolclass was incorrectly handling complex types like dictionaries and unions when generating parameter schemas. Specifically:Fixed the schema generation for dictionary types (
Dict[str, Any]) by:"type": "object"schema"additionalProperties": trueto allow arbitrary key-value pairsAdded proper handling for Union types by:
oneOfschema generation for Union typesUnion[Dict[str, Any], List[Dict[str, Any]]]type hintsThese changes ensure that tools like HTMLToDocument's parameters are correctly represented in the schema, allowing LLMs to properly understand and use the parameters.
How did you test it?
Added a new comprehensive unit test
test_from_component_with_complex_typesthat verifies:Ran all existing tests to ensure no regressions
Verified pre-commit hooks pass including:
Notes for the reviewer
Key areas to review:
_create_property_schemamethod incomponent_tool.pywhere the main changes were madetest_component_tool.pythat verifies the fixoneOfschema generationChecklist
fix: improve ComponentTool parameter schema for complex types