feat(wren-ai-service): Add LLM-based evaluation metrics for SQL gener…

…ation (#1303)

wren-ai-service/eval/evaluation.py

@@ -15,11 +15,26 @@
 import eval.pipelines as pipelines
 import src.providers as provider
 from eval import EvalSettings
-from eval.utils import engine_config, parse_toml, trace_metadata
+from eval.utils import parse_toml, trace_metadata
 from src import utils
 
 
 def formatter(prediction: dict, meta: dict) -> dict:
+    """
+    Formats the prediction result to be used as evaluation input.
+
+    This function takes a prediction dictionary and a meta dictionary,
+    processes them to extract relevant information, and returns a formatted
+    dictionary that serves as input for evaluation. It includes details such
+    as input, actual and expected outputs, context, and additional metadata.
+
+    Args:
+        prediction (dict): A dictionary containing prediction details.
+        meta (dict): A dictionary containing metadata information.
+
+    Returns:
+        dict: A formatted dictionary containing evaluation input data.
+    """
     retrieval_context = [str(context) for context in prediction["retrieval_context"]]
     context = [str(context) for context in prediction["context"]]
     enable_spider_metrics = "spider" in meta.get("evaluation_dataset", "").lower()
@@ -33,6 +48,7 @@ def formatter(prediction: dict, meta: dict) -> dict:
         "expected_output": prediction["expected_output"],
         "retrieval_context": retrieval_context,
         "context": context,
+        "reasoning": prediction.get("reasoning", ""),
         "additional_metadata": {
             "trace_id": prediction["trace_id"],
             "trace_url": prediction["trace_url"],
@@ -82,7 +98,9 @@ def eval(self, meta: dict, predictions: list) -> None:
 
             try:
                 test_case = LLMTestCase(**formatter(prediction, meta))
-                result = evaluate([test_case], self._metrics, ignore_errors=True).test_results[0]
+                result = evaluate(
+                    [test_case], self._metrics, ignore_errors=True
+                ).test_results[0]
                 self._score_metrics(test_case, result)
                 [metric.collect(test_case, result) for metric in self._post_metrics]
             except Exception:
@@ -152,8 +170,9 @@ def _average_score(self, meta: dict) -> None:
     predictions = predicted_file["predictions"]
 
     dataset = parse_toml(meta["evaluation_dataset"])
-    engine_info = engine_config(dataset["mdl"], pipe_components)
-    metrics = pipelines.metrics_initiator(meta["pipeline"], engine_info, args.semantics)
+    metrics = pipelines.metrics_initiator(
+        meta["pipeline"], dataset, pipe_components, args.semantics
+    )
 
     evaluator = Evaluator(**metrics)
     evaluator.eval(meta, predictions)

wren-ai-service/eval/metrics/__init__.py

@@ -4,6 +4,11 @@
 from .context_recall import ContextualRecallMetric
 from .context_relevancy import ContextualRelevancyMetric
 from .faithfulness import FaithfulnessMetric
+from .llm import (
+    QuestionToReasoningJudge,
+    ReasoningToSqlJudge,
+    SqlSemanticsJudge,
+)
 from .spider.exact_match import ExactMatchAccuracy
 from .spider.exec_match import ExecutionAccuracy
 
@@ -17,4 +22,7 @@
     "FaithfulnessMetric",
     "ExactMatchAccuracy",
     "ExecutionAccuracy",
+    "QuestionToReasoningJudge",
+    "ReasoningToSqlJudge",
+    "SqlSemanticsJudge",
 ]

wren-ai-service/eval/metrics/llm/__init__.py

@@ -0,0 +1,173 @@
+import asyncio
+
+from deepeval.metrics import BaseMetric
+from deepeval.test_case import LLMTestCase
+from haystack.components.builders.prompt_builder import PromptBuilder
+from pydantic import BaseModel
+
+from src.providers import LLMProvider
+
+
+class EvalResult(BaseModel):
+    score: float
+    reason: str
+
+
+_MODEL_KWARGS = {
+    "response_format": {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "eval_result",
+            "schema": EvalResult.model_json_schema(),
+        },
+    }
+}
+
+
+def format(response: dict) -> EvalResult:
+    reply = response.get("replies", [])[0]
+    return EvalResult.model_validate_json(reply)
+
+
+class QuestionToReasoningJudge(BaseMetric):
+    _system_prompt = """
+    You are an expert evaluator. Your task is to analyze the reasoning provided for a given question and determine if it makes sense. 
+    Provide a score in the range 0.0~1.0 and a detailed explanation for your evaluation.
+    """
+    _test_case_prompt = """
+    Question: 
+    {{ question }}
+    
+    Reasoning:
+    {{ reasoning }}
+    """
+
+    def __init__(self, llm_provider: LLMProvider, **_):
+        self.threshold = 0
+        self.score = 0
+        self.llm_provider = llm_provider
+        self.llm = llm_provider.get_generator(
+            system_prompt=self._system_prompt,
+            generation_kwargs=_MODEL_KWARGS,
+        )
+        self.prompt_builder = PromptBuilder(template=self._test_case_prompt)
+
+    def measure(self, test_case: LLMTestCase):
+        return asyncio.run(self.a_measure(test_case))
+
+    async def a_measure(self, test_case: LLMTestCase, *args, **kwargs):
+        prompt = self.prompt_builder.run(
+            question=test_case.input,
+            reasoning=test_case.reasoning,
+        )
+        response = await self.llm(prompt.get("prompt"))
+        result = format(response)
+
+        self.score = result.score
+        self.reason = result.reason
+
+        self.success = self.score >= self.threshold
+        return self.score
+
+    def is_successful(self):
+        return self.success
+
+    @property
+    def __name__(self):
+        return "QuestionToReasoningJudge"
+
+
+class ReasoningToSqlJudge(BaseMetric):
+    _system_prompt = """
+    You are an expert evaluator. Your task is to analyze the reasoning provided for a given SQL query and determine if it makes sense. 
+    Provide a score in the range 0.0~1.0 and a detailed explanation for your evaluation.
+    """
+    _test_case_prompt = """
+    Actual Output: 
+    {{ actual_output }}
+
+    Reasoning:
+    {{ reasoning }}
+    """
+
+    def __init__(self, llm_provider: LLMProvider, **_):
+        self.threshold = 0
+        self.score = 0
+        self.llm_provider = llm_provider
+        self.llm = llm_provider.get_generator(
+            system_prompt=self._system_prompt,
+            generation_kwargs=_MODEL_KWARGS,
+        )
+        self.prompt_builder = PromptBuilder(template=self._test_case_prompt)
+
+    def measure(self, test_case: LLMTestCase):
+        return asyncio.run(self.a_measure(test_case))
+
+    async def a_measure(self, test_case: LLMTestCase, *args, **kwargs):
+        prompt = self.prompt_builder.run(
+            actual_output=test_case.actual_output,
+            reasoning=test_case.reasoning,
+        )
+        response = await self.llm(prompt.get("prompt"))
+        result = format(response)
+
+        self.score = result.score
+        self.reason = result.reason
+
+        self.success = self.score >= self.threshold
+        return self.score
+
+    def is_successful(self):
+        return self.success
+
+    @property
+    def __name__(self):
+        return "ReasoningToSqlJudge"
+
+
+class SqlSemanticsJudge(BaseMetric):
+    _system_prompt = """
+    You are an expert evaluator. Your task is to analyze the actual SQL query and the expected SQL query and determine if they are semantically equivalent. 
+    Provide a score in the range 0.0~1.0 and a detailed explanation for your evaluation.
+    """
+    _test_case_prompt = """
+    Actual SQL: 
+    {{ actual_sql }}
+
+    Expected SQL: 
+    {{ expected_sql }}
+    """
+
+    def __init__(self, llm_provider: LLMProvider, **_):
+        self.threshold = 0
+        self.score = 0
+        self.llm_provider = llm_provider
+        self.llm = llm_provider.get_generator(
+            system_prompt=self._system_prompt,
+            generation_kwargs=_MODEL_KWARGS,
+        )
+        self.prompt_builder = PromptBuilder(template=self._test_case_prompt)
+
+    def measure(self, test_case: LLMTestCase):
+        return asyncio.run(self.a_measure(test_case))
+
+    async def a_measure(self, test_case: LLMTestCase, *args, **kwargs):
+        prompt = self.prompt_builder.run(
+            actual_sql=test_case.actual_output,
+            expected_sql=test_case.expected_output,
+        )
+        response = await self.llm(prompt.get("prompt"))
+        result = format(response)
+
+        self.score = result.score
+        self.reason = result.reason
+
+        self.success = self.score >= self.threshold
+        return self.score
+
+    def is_successful(self):
+        return self.success
+
+    @property
+    def __name__(self):
+        return "SqlSemanticsJudge"

wren-ai-service/eval/pipelines.py

@@ -11,6 +11,8 @@
 from langfuse.decorators import langfuse_context, observe
 from tqdm.asyncio import tqdm_asyncio
 
+from src.core.pipeline import PipelineComponent
+
 sys.path.append(f"{Path().parent.resolve()}")
 
 from eval import EvalSettings
@@ -23,6 +25,9 @@
     ExactMatchAccuracy,
     ExecutionAccuracy,
     FaithfulnessMetric,
+    QuestionToReasoningJudge,
+    ReasoningToSqlJudge,
+    SqlSemanticsJudge,
 )
 from eval.utils import (
     engine_config,
@@ -290,7 +295,11 @@ async def __call__(self, query: str, **_):
         ]
 
     @staticmethod
-    def metrics(engine_info: dict, enable_semantics_comparison: bool) -> dict:
+    def metrics(
+        engine_info: dict,
+        enable_semantics_comparison: bool,
+        component: PipelineComponent,
+    ) -> dict:
         return {
             "metrics": [
                 AccuracyMetric(
@@ -302,6 +311,9 @@ def metrics(engine_info: dict, enable_semantics_comparison: bool) -> dict:
                 # this is for spider dataset, rn we temporarily disable it
                 ExactMatchAccuracy(),
                 ExecutionAccuracy(),
+                QuestionToReasoningJudge(**component),
+                ReasoningToSqlJudge(**component),
+                SqlSemanticsJudge(**component),
             ],
             "post_metrics": [],
         }
@@ -402,7 +414,11 @@ async def __call__(self, query: str, **_):
         ]
 
     @staticmethod
-    def metrics(engine_info: dict, enable_semantics_comparison: bool) -> dict:
+    def metrics(
+        engine_info: dict,
+        enable_semantics_comparison: bool,
+        component: PipelineComponent,
+    ) -> dict:
         return {
             "metrics": [
                 AccuracyMetric(
@@ -417,6 +433,9 @@ def metrics(engine_info: dict, enable_semantics_comparison: bool) -> dict:
                 # this is for spider dataset, rn we temporarily disable it
                 ExactMatchAccuracy(),
                 ExecutionAccuracy(),
+                QuestionToReasoningJudge(**component),
+                ReasoningToSqlJudge(**component),
+                SqlSemanticsJudge(**component),
             ],
             "post_metrics": [],
         }
@@ -449,13 +468,20 @@ def init(
 
 def metrics_initiator(
     pipeline: str,
-    engine_info: dict,
+    dataset: dict,
+    pipe_components: dict[str, PipelineComponent],
     enable_semantics_comparison: bool = True,
 ) -> dict:
+    engine_info = engine_config(dataset["mdl"], pipe_components)
+    component = pipe_components["evaluation"]
     match pipeline:
         case "retrieval":
             return RetrievalPipeline.metrics(engine_info)
         case "generation":
-            return GenerationPipeline.metrics(engine_info, enable_semantics_comparison)
+            return GenerationPipeline.metrics(
+                engine_info, enable_semantics_comparison, component
+            )
         case "ask":
-            return AskPipeline.metrics(engine_info, enable_semantics_comparison)
+            return AskPipeline.metrics(
+                engine_info, enable_semantics_comparison, component
+            )

wren-ai-service/tools/config/config.example.yaml

@@ -145,6 +145,8 @@ pipes:
   - name: sql_regeneration
     llm: litellm_llm.gpt-4o-mini-2024-07-18
     engine: wren_ui
+  - name: evaluation
+    llm: litellm_llm.gpt-4o-mini-2024-07-18
 
 ---
 settings:

wren-ai-service/tools/config/config.full.yaml

@@ -145,6 +145,8 @@ pipes:
   - name: sql_regeneration
     llm: litellm_llm.gpt-4o-mini-2024-07-18
     engine: wren_ui
+  - name: evaluation
+    llm: litellm_llm.gpt-4o-mini-2024-07-18
 
 ---
 settings: