Overview

Ragas is an evaluation framework for retrieval augmented generation (RAG) applications. Datadog’s Ragas integration enables you to evaluate your production application with scores for faithfulness, answer relevancy, and context precision. You can use these scores to find traces that have a high likelihood of inaccurate answers and review them to improve your RAG pipeline.

For a simplified setup guide, see Ragas Quickstart.

Evaluations

Faithfulness

The Faithfulness score evaluates how consistent an LLM’s generation is against the provided ground truth context data.

This score is generated through three steps:

1. Creating statements: Asking another LLM to break down an answer into individual statements
2. Creating verdicts: For each statement, determining if it is unfaithful to the provided context
3. Computing a score: Dividing the number of contradicting statements over the total number of statements

For more information, see Ragas’s Faithfulness documentation.

Answer Relevancy

The Answer Relevancy (or Response Relevancy) score assesses how pertinent the generated answer is to the given prompt. A lower score is assigned to answers that are incomplete or contain redundant information, and higher scores indicate better relevancy. This metric is computed using the question, the retrieved contexts, and the answer.

The Answer Relevancy score is defined as the mean cosine similarity of the original question to a number of artificial questions, which are generated (reverse engineered) based on the response.

For more information, see Ragas’s Answer Relevancy documentation.

Context Precision

The Context Precision score assesses if the context was useful in arriving at the given answer.

This score is modified from Ragas’s original Context Precision metric, which computes the mean of the Precision@k for each chunk in the context. Precision@k is the ratio of the number of relevant chunks at rank k to the total number of chunks at rank k.

Datadog’s Context Precision score is computed by dividing the number of relevant contexts by the total number of contexts.

For more information, see Ragas’s Context Precision documentation.

Setup

Datadog’s Ragas evaluations require ragas v0.1+ and ddtrace v3.0.0+.

1. Install dependencies. Run the following command:

   Copy

      pip install ragas==0.1.21 openai ddtrace>=3.0.0
      

   The Ragas integration automatically runs evaluations in the background of your application. By default, Ragas uses OpenAI’s GPT-4 model for evaluations, which requires you to set an OPENAI_API_KEY in your environment. You can also customize Ragas to use a different LLM.

2. Instrument your LLM calls with RAG context information. Datadog’s Ragas integration attempts to extract context information from the prompt variables attached to a span.

   Examples:

   from ddtrace.llmobs import LLMObs
   from ddtrace.llmobs.utils import Prompt
   
   with LLMOBs.annotation_context(
       prompt=Prompt(
           variables={"context": "rag context here"},
           rag_context_variable_keys = ["context"], # defaults to ['context']
           rag_query_variable_keys = ["question"], # defaults to ['question']
       ),
       name="generate_answer",
   ):
       oai_client.chat.completions.create(...) 
   

   from ddtrace.llmobs import LLMObs
   from ddtrace.llmobs.utils import Prompt
   
   @llm(model = "llama")
   def generate_answer():
       ...
       LLMObs.annotate(
           prompt=Prompt(variables={'context': "rag context..."})
       )
   

3. (Optional, but recommended) Enable sampling. Datadog traces Ragas score generation. These traces contain LLM spans, which may affect your LLM Observability billing. See Sampling.

4. Run your script and specify enabled Ragas evaluators. Use the environment variable DD_LLMOBS_EVALUATORS to provide a comma-separated list of Ragas evaluators you wish to enable. These evaluators are ragas_faithfulness, ragas_context_precision, and ragas_answer_relevancy.

   For example, to run your script with all Ragas evaluators enabled:

   DD_LLMOBS_EVALUATORS="ragas_faithfulness,ragas_context_precision,ragas_answer_relevancy" \
   DD_ENV=dev \
   DD_API_KEY=<YOUR-DATADOG-API-KEY> \
   DD_SITE= \
   python driver.py
   

Configuration

Sampling

To enable Ragas scoring for a sampled subset of LLM calls, use the DD_LLMOBS_EVALUATORS_SAMPLING_RULES environment variable. Pass in a list of arrays, each containing the following fields:

In the following example, Ragas Faithfulness scoring is enabled for 50% of all answer_question spans. Ragas evaluations are disabled for all other spans ("sample_rate": 0).

export DD_LLMOBS_EVALUATORS_SAMPLING_RULES='[
  {
    "sample_rate": 0.5,
    "evaluator_label": "ragas_faithfulness",
    "span_name": "answer_question"
  },
  {
    "sample_rate": 0
  }
]'

Customization

Ragas supports customizations. For example, the following snippet configures the Faithfulness evaluator to use gpt-4 and adds custom instructions for the prompt to the evaluating LLM:

from langchain_openai import ChatOpenAI
from ragas.metrics import faithfulness
from ragas.llms.base import LangchainLLMWrapper

faithfulness.llm = LangchainLLMWrapper(ChatOpenAI(model="gpt-4"))
faithfulness.statement_prompt.instruction += "\nMake sure text containing code instructions are grouped with contextual information on how to run that code."

Any customizations you make to the global Ragas instance are automatically applied to Datadog’s Ragas evaluators. No action is required.

View Ragas evaluations in Datadog

Ragas scores are sent to Datadog as evaluation metrics. When you view a scored trace in LLM Observability, Ragas scores appear under Custom Evaluations.

You can also configure your LLM Traces page to display Ragas scores.

1. Go to the LLM Observability Traces page in Datadog.
2. Search for @meta.span.kind:llm in All Spans to view only LLM spans.
3. Add -runner.integration:ragas to the search field. Datadog automatically traces the generation of Ragas scores. Use this exclusion term to filter out these traces.
4. Select Custom Evaluations and enable your desired Ragas evaluations. These scores are then displayed in additional columns on the LLM Observability Traces page.

Troubleshooting

Missing evaluations

Your LLM inference could be missing an evaluation for the following reasons:

• The LLM span was not sampled for an evaluation because you have implemented sampling.
• An error occurred during the Ragas evaluation. Search for runner.integration:ragas to see traces for the Ragas evaluation itself.

Flushing

Use the LLMObs.flush() command to guarantee all traces and evaluations are flushed to Datadog. Note: This is a blocking function.

Further reading

추가 유용한 문서, 링크 및 기사:

Submit EvaluationsDOCUMENTATION