changes to llm chain #6328
changes to llm chain #6328
Conversation
hwchase17
commented
Jun 17, 2023
hwchase17
commented
- return raw and full output (but keep run shortcut method functional)
- change output parser to take in generations (good for working with messages)
- add output parser to base class, always run (default to same as current)
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎ 1 Ignored Deployment
|
langchain/chains/llm.py
Outdated
| {self.output_key: generation[0].text} | ||
| { | ||
| self.output_key: self.output_parser.parse_result(generation), | ||
| "result": generation, |
There was a problem hiding this comment.
Can this be a breaking change when used as input to another chain etc (because was 1 key before and now 2)
There was a problem hiding this comment.
i think only minorly
cause if they are using call it returns a dictionary, so if they were doing something weird like list(return)[0] or something like that then ya. but if they were using run i changed it so i think it still works as intended, and i imagine most people either use run or use the dictionary as a whole
There was a problem hiding this comment.
actually thats wrong will mess things up with memory, will update to add a flag
langchain/chains/llm.py
Outdated
| @@ -43,6 +48,8 @@ def lc_serializable(self) -> bool: | |||
| """Prompt object to use.""" | |||
| llm: BaseLanguageModel | |||
| output_key: str = "text" #: :meta private: | |||
| output_parser: BaseLLMOutputParser = Field(default_factory=DefaultOutputParser) | |||
There was a problem hiding this comment.
how does this play with
def apply_and_parse(
self, input_list: List[Dict[str, Any]], callbacks: Callbacks = None
) -> Sequence[Union[str, List[str], Dict[str, str]]]:
"""Call apply and then parse the results."""
result = self.apply(input_list, callbacks=callbacks)
return self._parse_result(result)
There was a problem hiding this comment.
ill add a deprecation warning
langchain/chains/llm.py
Outdated
| def _run_output_key(self) -> str: | ||
| return self.output_key | ||
|
|
||
| def create_outputs(self, response: LLMResult) -> List[Dict[str, Any]]: |
There was a problem hiding this comment.
I don't see a way to propagate llm_output to the end user because top level response is a list rather than dict and llm_output is across all inputs if I understand correctly.
But if there's a way to do it, we should do it, so users don't have to implement callbacks as a work-around to raw information missing.
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
…gchain into harrison/llm-chain-changes
use functions in llm chain
