Creating Encodable For Callable Type #456
Conversation
effectful/handlers/llm/encoding.py
Outdated
| @classmethod | ||
| @abstractmethod | ||
| def decode(cls, vl: U) -> T: | ||
| def decode(cls, vl: U, template: typing.Any = None) -> T: |
There was a problem hiding this comment.
encode and decode are intended mainly to deal with types that have special meaning for the LLM api. We shouldn't extend the interface in this way.
effectful/handlers/llm/synthesis.py
Outdated
|
|
||
| # Start with provided context or empty dict | ||
| # Include collections module for type hints in synthesized code | ||
| exec_globals: dict[str, typing.Any] = {"collections": collections} |
There was a problem hiding this comment.
The synthesized code should have imports for types or they should come from the lexical context.
effectful/handlers/llm/synthesis.py
Outdated
| _decode_counter: typing.ClassVar[int] = 0 | ||
|
|
||
| @classmethod | ||
| def decode(cls, vl: SynthesizedFunction, template: typing.Any = None) -> Callable: |
There was a problem hiding this comment.
This could be a decode_response handler, rather than changing the decode interface.
There was a problem hiding this comment.
Thanks! it makes sense, so we add an indirection between the LLM response and decode?
There was a problem hiding this comment.
I implemented it, but I'm not sure it's the cleanest implementation, mainly in how we would pass context to decode, because context has to be bound before executing the returned code.
effectful/handlers/llm/synthesis.py
Outdated
| context: LexicalContext | None = None | ||
| if template is not None and hasattr(template, "__context__"): | ||
| ctx = template.__context__ | ||
| context = ctx if isinstance(ctx, LexicalContext) else LexicalContext(ctx) |
There was a problem hiding this comment.
We should assume there is always a lexical context.
Summarize changes:
Templateparameter to decodeTemplateto gather lexical context for decoding synthesized function