LLM Wrapper Error: ValueError: OPENAI_API_KEY not found in the environmen(t).

[antoninoLorenzo]

Code
I tried to use evaluate with a LangchainLLMWrapper, however for some it still requires an OpenAI key, here is the code:

from ragrank import evaluate
from ragrank.evaluation import EvalResult
from ragrank.integrations.langchain import LangchainLLMWrapper
from ragrank.dataset import from_dataframe
from ragrank.metric import (
    context_relevancy,
    response_relevancy
)
from langchain_community.chat_models import ChatOllama

rr_dataset = from_dataframe(df)

ollama_llm = ChatOllama(model='gemma:2b')
ragrank_llm = LangchainLLMWrapper(llm=ollama_llm)

result: EvalResult = evaluate(
    dataset=rr_dataset,
    llm=ragrank_llm,
    metrics=[
        response_relevancy,
        context_relevancy,
    ],
)

PS: I do not usally work with Langchain so there is a chance I done something wrong with it, however by the error raised, even if this is the case, it is unclear if that's a bug in the library

Problem
By further inspecting the problem I found that it is happening here:

File D:\...\.venv\Lib\site-packages\ragrank\evaluation\base.py:74, in <listcomp>(.0)
     70     metrics = [metrics]
     72 dt = time()
     73 scores = [
---> 74     [
     75         metric.score(datanode).score
     76         for datanode in dataset.with_progress("Evaluating")
     77     ]
     78     for metric in metrics
     79 ]
     
     ...
     
     File D:\...\.venv\Lib\site-packages\ragrank\metric\_response_related\relevancy.py:74, in ResponseRelevancy.score(self, data)
     72 prompt_str = self.prompt.to_string()
     73 prompt_dt = prompt_str.format(**data.model_dump())
---> 74 response = self.llm.generate_text(
     75     prompt_dt,
     76 )
     77 try:
     78     score = float(response.response)

Solution Attempt
I read trough the source code and tried to find a work around, this was my attempt:

...
from ragrank.metric._response_related.relevancy import ResponseRelevancy
from ragrank.metric._context_related.relevancy import ContextRevevancy

resp_r = ResponseRelevancy()
resp_r.llm = ragrank_llm

cont_r = ContextRevevancy()
cont_r.llm = ragrank_llm

result: EvalResult = evaluate(
    dataset=rr_dataset,
    llm=ragrank_llm,
    metrics=[
        resp_r,
        cont_r,
    ],
)

However it yielded another error KeyError: 'token_usage':

File D:\...\.venv\Lib\site-packages\ragrank\metric\_response_related\relevancy.py:74, in ResponseRelevancy.score(self, data)
     72 prompt_str = self.prompt.to_string()
     73 prompt_dt = prompt_str.format(**data.model_dump())
---> 74 response = self.llm.generate_text(
     75     prompt_dt,
     76 )
     77 try:
     78     score = float(response.response)

File D:\...\.venv\Lib\site-packages\ragrank\integrations\langchain\langchain_llm_wrapper.py:97, in LangchainLLMWrapper.generate_text(self, text)
     93 langchain_result: LangchainLLMResult = (
     94     self.llm.generate_prompt(prompts=[prompt])
     95 )
     96 message = langchain_result.generations[0][0].text
---> 97 response_tokens = langchain_result.llm_output["token_usage"][
     98     "completion_tokens"
     99 ]
    100 response_time = time() - start_time
    102 result = LLMResult(
    103     response=message,
    104     response_time=response_time,
   (...)
    107     llm_config=self.llm_config,
    108 )

At that point I don't know what I can try going deeper in the source code (i.e I can't propose a solution), so I point out that problem in hope that you can improve the library. (or tell me what I am doing wrong)

*PS: 'ValueError: OPENAI_API_KEY not found in the environmen.' has a typo, it misses the 't'; *

[github-actions]

Thanks a lot for the first issue posting

[izam-mohammed]

Thanks for pointing this out @antoninoLorenzo. I appreciate that you tried to solve it on your own. It is actually a bug in the langchain integration part. it will be resolved in the next release (0.0.8).

I assume that the first problem arose because, in the file evaluation/base.py, there is an import default_llm. it needs to change. and also we are accessing the token_usage property which is only available for a few llms in langchain. Needed to remove the token_usage attribute in langchain integration.

If possible, please raise a PR as a solution.