Library for evaluating RAG using nuclia's models
Its evaluation follows the RAG triad as proposed by TruLens:
In summary, the metrics nuclia-eval provides for a RAG Experience involving a question, an answer and N pieces of context are:
- Answer Relevance: Answer relevance refers to the directness and appropriateness of the response in addressing the specific question asked, providing accurate, complete, and contextually suitable information.
- score: A number between 0 and 5 indicating the score of the relevance of the answer to the question.
- reason: A string explaining the reason for the score.
- For each of the N pieces of context:
- Context Relevance Score: The context relevance is the relevance of the context to the question, on a scale of 0 to 5.
- Groundedness Score: Groundedness is defined as the degree of information overlap to which the answer contains information that is substantially similar or identical to that in the context piece. The score is between 0 and 5.
nuclia-eval is only supported on Linux-based systems.
Installing the package
pip install nuclia-evalRequirements for downloading the models
To download the models, you must have a Hugging Face account and be logged in. You can create an account here. You also need to authenticate your session with the Hugging Face API by running huggingface-cli login or any other method described here.
Then, you need to have access to both the base model and the adapter model, you can easily request access to each model by clicking the button on the model's page (once logged in). More information here.
For example, for REMi-v0, you need access to the Mistral-7B-Instruct-v0.3 base model and the REMi-v0 adapter model.
If this authentication and authorization process is not completed, you will see a message like this when trying to instantiate an evaluator for the first time:
Access to model __model_name__ is restricted. You must be authenticated to access it.
REMi-v0 (RAG Evaluation MetrIcs) is a LoRa adapter for the Mistral-7B-Instruct-v0.3 model.
It has been finetuned by the team at nuclia to evaluate the quality of all parts of the RAG experience.
Note: The REMi-v0 model requires a GPU with at least 24GB of memory.
from nuclia_eval import REMi
evaluator = REMi()
query = "By how many Octaves can I shift my OXYGEN PRO 49 keyboard?"
context1 = """\
* Oxygen Pro 49's keyboard can be shifted 3 octaves down or 4 octaves up.
* Oxygen Pro 61's keyboard can be shifted 3 octaves down or 3 octaves up.
To change the transposition of the keyboard, press and hold Shift, and then use the Key Octave –/+ buttons to lower or raise the keybed by one one, respectively.
The display will temporarily show TRANS and the current transposition (-12 to 12)."""
context2 ="""\
To change the octave of the keyboard, use the Key Octave –/+ buttons to lower or raise the octave, respectively
The display will temporarily show OCT and the current octave shift.\n\nOxygen Pro 25's keyboard can be shifted 4 octaves down or 5 octaves up"""
context3 = """\
If your DAW does not automatically configure your Oxygen Pro series keyboard, please follow the setup steps listed in the Oxygen Pro DAW Setup Guides.
To set the keyboard to operate in Preset Mode, press the DAW/Preset Button (on the Oxygen Pro 25) or Preset Button (on the Oxygen Pro 49 and 61).
On the Oxygen Pro 25 the DAW/Preset button LED will be off to show that Preset Mode is selected.
On the Oxygen Pro 49 and 61 the Preset button LED will be lit to show that Preset Mode is selected."""
answer = "Based on the context provided, The Oxygen Pro 49's keyboard can be shifted 3 octaves down or 4 octaves up."
result = evaluator.evaluate_rag(query=query, answer=answer, contexts=[context1, context2, context3])
answer_relevance, context_relevances, groundednesses = result
print(f"{answer_relevance.score}, {answer_relevance.reason}")
# 5, The response directly answers the query by specifying the range of octave shifts for the Oxygen Pro 49 keyboard.
print([cr.score for cr in context_relevances]) # [5, 1, 0]
print([g.score for g in groundednesses]) # [2, 0, 0]The REMi evaluator provides a fine-grained and strict evaluation of the RAG triad. For instance if we slightly modify the answer to the query:
- answer = "Based on the context provided, The Oxygen Pro 49's keyboard can be shifted 3 octaves down or 4 octaves up."
+ answer = "Based on the context provided, the Oxygen Pro 49's keyboard can be shifted 4 octaves down or 4 octaves up."
...
print([g.score for g in groundednesses]) # [0, 0, 0]As the information provided in the answer is not present in any of the contexts, the groundedness score is 0 for all contexts.
What if the information in the answer does not answer the question?
- answer = "Based on the context provided, The Oxygen Pro 49's keyboard can be shifted 3 octaves down or 4 octaves up."
+ answer = "Based on the context provided, the Oxygen Pro 61's keyboard can be shifted 3 octaves down or 4 octaves up."
...
print(f"{answer_relevance.score}, {answer_relevance.reason}")
# 1, The response is relevant to the entire query but incorrectly mentions the Oxygen Pro 61 instead of the Oxygen Pro 49We can also compute each metric separately:
...
answer_relevance = evaluator.answer_relevance(query=query, answer=answer)
context_relevances = evaluator.context_relevance(query=query, contexts=[context1, context2, context3])
groundednesses = evaluator.groundedness(answer=answer, contexts=[context1, context2, context3])
...By default, the models are downloaded to the ~/.nuclia-model-cache/ directory. You can specify a different location in two ways:
- By setting the
NUCLIA_MODEL_CACHEenvironment variable to the desired location. - By overriding the default settings when instantiating the evaluator:
from nuclia_eval import REMi
from nuclia_eval.settings import Settings
# Create custom settings
settings = Settings(
nuclia_model_cache="my_cache/",
)
# Instantiate the evaluator
evaluator = REMi(settings=settings)For feedback, questions, or to get in touch with the nuclia team, we are available on our community Slack channel.