Feat: evaluate retrieved context against golden context [cog-1481] (#619)

<!-- .github/pull_request_template.md -->

## Description
- Compare retrieved context to golden context using deepeval's
summarization metric
- Display relevant fields to each metric on metrics dashboard

Example output:

![image](https://github.com/user-attachments/assets/9facf716-b2ab-4573-bfdf-7b343d2a57c5)


## DCO Affirmation
I affirm that all code in every commit of this pull request conforms to
the terms of the Topoteretes Developer Certificate of Origin


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **New Features**
- Enhanced context handling in answer generation and corpus building to
include extended details.
- Introduced a new context coverage metric for deeper evaluation
insights.
- Upgraded the evaluation dashboard with dynamic presentation of metric
details.
- Added a new parameter to support loading golden context in corpus
loading methods.

- **Bug Fixes**
- Improved clarity in how answers are structured and appended in the
answer generation process.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: alekszievr

cognee/eval_framework/answer_generation/answer_generation_executor.py

@@ -29,13 +29,16 @@ async def question_answering_non_parallel(
             retrieval_context = await retriever.get_context(query_text)
             search_results = await retriever.get_completion(query_text, retrieval_context)
 
-            answers.append(
-                {
-                    "question": query_text,
-                    "answer": search_results[0],
-                    "golden_answer": correct_answer,
-                    "retrieval_context": retrieval_context,
-                }
-            )
+            answer = {
+                "question": query_text,
+                "answer": search_results[0],
+                "golden_answer": correct_answer,
+                "retrieval_context": retrieval_context,
+            }
+
+            if "golden_context" in instance:
+                answer["golden_context"] = instance["golden_context"]
+
+            answers.append(answer)
 
         return answers

cognee/eval_framework/benchmark_adapters/dummy_adapter.py

@@ -5,18 +5,21 @@
 
 class DummyAdapter(BaseBenchmarkAdapter):
     def load_corpus(
-        self, limit: Optional[int] = None, seed: int = 42
+        self, limit: Optional[int] = None, seed: int = 42, load_golden_context: bool = False
     ) -> tuple[list[str], list[dict[str, Any]]]:
         corpus_list = [
             "The cognee is an AI memory engine that supports different vector and graph databases",
             "Neo4j is a graph database supported by cognee",
         ]
-        question_answer_pairs = [
-            {
-                "answer": "Yes",
-                "question": "Is Neo4j supported by cognee?",
-                "type": "dummy",
-            }
-        ]
+        qa_pair = {
+            "answer": "Yes",
+            "question": "Is Neo4j supported by cognee?",
+            "type": "dummy",
+        }
+
+        if load_golden_context:
+            qa_pair["golden_context"] = "Cognee supports Neo4j and NetworkX"
+
+        question_answer_pairs = [qa_pair]
 
         return corpus_list, question_answer_pairs

cognee/eval_framework/corpus_builder/corpus_builder_executor.py

@@ -28,14 +28,22 @@ def __init__(
         self.questions = None
         self.task_getter = task_getter
 
-    def load_corpus(self, limit: Optional[int] = None) -> Tuple[List[Dict], List[str]]:
-        self.raw_corpus, self.questions = self.adapter.load_corpus(limit=limit)
+    def load_corpus(
+        self, limit: Optional[int] = None, load_golden_context: bool = False
+    ) -> Tuple[List[Dict], List[str]]:
+        self.raw_corpus, self.questions = self.adapter.load_corpus(
+            limit=limit, load_golden_context=load_golden_context
+        )
         return self.raw_corpus, self.questions
 
     async def build_corpus(
-        self, limit: Optional[int] = None, chunk_size=1024, chunker=TextChunker
+        self,
+        limit: Optional[int] = None,
+        chunk_size=1024,
+        chunker=TextChunker,
+        load_golden_context: bool = False,
     ) -> List[str]:
-        self.load_corpus(limit=limit)
+        self.load_corpus(limit=limit, load_golden_context=load_golden_context)
         await self.run_cognee(chunk_size=chunk_size, chunker=chunker)
         return self.questions
 

cognee/eval_framework/corpus_builder/run_corpus_builder.py

@@ -47,7 +47,10 @@ async def run_corpus_builder(params: dict, chunk_size=1024, chunker=TextChunker)
             task_getter=task_getter,
         )
         questions = await corpus_builder.build_corpus(
-            limit=params.get("number_of_samples_in_corpus"), chunk_size=chunk_size, chunker=chunker
+            limit=params.get("number_of_samples_in_corpus"),
+            chunk_size=chunk_size,
+            chunker=chunker,
+            load_golden_context=params.get("evaluating_contexts"),
         )
         with open(params["questions_path"], "w", encoding="utf-8") as f:
             json.dump(questions, f, ensure_ascii=False, indent=4)

cognee/eval_framework/evaluation/deep_eval_adapter.py

@@ -4,6 +4,7 @@
 from cognee.eval_framework.evaluation.base_eval_adapter import BaseEvalAdapter
 from cognee.eval_framework.evaluation.metrics.exact_match import ExactMatchMetric
 from cognee.eval_framework.evaluation.metrics.f1 import F1ScoreMetric
+from cognee.eval_framework.evaluation.metrics.context_coverage import ContextCoverageMetric
 from typing import Any, Dict, List
 from deepeval.metrics import ContextualRelevancyMetric
 
@@ -15,6 +16,7 @@ def __init__(self):
             "EM": ExactMatchMetric(),
             "f1": F1ScoreMetric(),
             "contextual_relevancy": ContextualRelevancyMetric(),
+            "context_coverage": ContextCoverageMetric(),
         }
 
     async def evaluate_answers(
@@ -32,6 +34,7 @@ async def evaluate_answers(
                 actual_output=answer["answer"],
                 expected_output=answer["golden_answer"],
                 retrieval_context=[answer["retrieval_context"]],
+                context=[answer["golden_context"]] if "golden_context" in answer else None,
             )
             metric_results = {}
             for metric in evaluator_metrics:

cognee/eval_framework/evaluation/evaluation_executor.py

@@ -23,5 +23,6 @@ def __init__(
     async def execute(self, answers: List[Dict[str, str]], evaluator_metrics: Any) -> Any:
         if self.evaluate_contexts:
             evaluator_metrics.append("contextual_relevancy")
+            evaluator_metrics.append("context_coverage")
         metrics = await self.eval_adapter.evaluate_answers(answers, evaluator_metrics)
         return metrics

cognee/eval_framework/evaluation/metrics/context_coverage.py

@@ -0,0 +1,50 @@
+from deepeval.metrics import SummarizationMetric
+from deepeval.test_case import LLMTestCase
+from deepeval.metrics.summarization.schema import ScoreType
+from deepeval.metrics.indicator import metric_progress_indicator
+from deepeval.utils import get_or_create_event_loop
+
+
+class ContextCoverageMetric(SummarizationMetric):
+    def measure(
+        self,
+        test_case,
+        _show_indicator: bool = True,
+    ) -> float:
+        mapped_test_case = LLMTestCase(
+            input=test_case.context[0],
+            actual_output=test_case.retrieval_context[0],
+        )
+        self.assessment_questions = None
+        self.evaluation_cost = 0 if self.using_native_model else None
+        with metric_progress_indicator(self, _show_indicator=_show_indicator):
+            if self.async_mode:
+                loop = get_or_create_event_loop()
+                return loop.run_until_complete(
+                    self.a_measure(mapped_test_case, _show_indicator=False)
+                )
+            else:
+                self.coverage_verdicts = self._generate_coverage_verdicts(mapped_test_case)
+                self.alignment_verdicts = []
+                self.score = self._calculate_score(ScoreType.COVERAGE)
+                self.reason = self._generate_reason()
+                self.success = self.score >= self.threshold
+                return self.score
+
+    async def a_measure(
+        self,
+        test_case,
+        _show_indicator: bool = True,
+    ) -> float:
+        self.evaluation_cost = 0 if self.using_native_model else None
+        with metric_progress_indicator(
+            self,
+            async_mode=True,
+            _show_indicator=_show_indicator,
+        ):
+            self.coverage_verdicts = await self._a_generate_coverage_verdicts(test_case)
+            self.alignment_verdicts = []
+            self.score = self._calculate_score(ScoreType.COVERAGE)
+            self.reason = await self._a_generate_reason()
+            self.success = self.score >= self.threshold
+            return self.score

cognee/eval_framework/metrics_dashboard.py

@@ -3,6 +3,12 @@
 from typing import Dict, List, Tuple
 from collections import defaultdict
 
+metrics_fields = {
+    "contextual_relevancy": ["question", "retrieval_context"],
+    "context_coverage": ["question", "retrieval_context", "golden_context"],
+}
+default_metrics_fields = ["question", "answer", "golden_answer"]
+
 
 def create_distribution_plots(metrics_data: Dict[str, List[float]]) -> List[str]:
     """Create distribution histogram plots for each metric."""
@@ -59,38 +65,30 @@ def generate_details_html(metrics_data: List[Dict]) -> List[str]:
         for metric, values in entry["metrics"].items():
             if metric not in metric_details:
                 metric_details[metric] = []
+            current_metrics_fields = metrics_fields.get(metric, default_metrics_fields)
             metric_details[metric].append(
-                {
-                    "question": entry["question"],
-                    "answer": entry["answer"],
-                    "golden_answer": entry["golden_answer"],
+                {key: entry[key] for key in current_metrics_fields}
+                | {
                     "reason": values.get("reason", ""),
                     "score": values["score"],
                 }
             )
 
     for metric, details in metric_details.items():
+        formatted_column_names = [key.replace("_", " ").title() for key in details[0].keys()]
         details_html.append(f"<h3>{metric} Details</h3>")
-        details_html.append("""
+        details_html.append(f"""
             <table class="metric-table">
                 <tr>
-                    <th>Question</th>
-                    <th>Answer</th>
-                    <th>Golden Answer</th>
-                    <th>Reason</th>
-                    <th>Score</th>
+                    {"".join(f"<th>{col}</th>" for col in formatted_column_names)}
                 </tr>
         """)
         for item in details:
-            details_html.append(
-                f"<tr>"
-                f"<td>{item['question']}</td>"
-                f"<td>{item['answer']}</td>"
-                f"<td>{item['golden_answer']}</td>"
-                f"<td>{item['reason']}</td>"
-                f"<td>{item['score']}</td>"
-                f"</tr>"
-            )
+            details_html.append(f"""
+                <tr>
+                    {"".join(f"<td>{value}</td>" for value in item.values())}
+                </tr>
+            """)
         details_html.append("</table>")
     return details_html
 

cognee/tests/unit/eval_framework/deepeval_adapter_test.py

@@ -5,7 +5,12 @@
 
 with patch.dict(
     sys.modules,
-    {"deepeval": MagicMock(), "deepeval.metrics": MagicMock(), "deepeval.test_case": MagicMock()},
+    {
+        "deepeval": MagicMock(),
+        "deepeval.metrics": MagicMock(),
+        "deepeval.test_case": MagicMock(),
+        "cognee.eval_framework.evaluation.metrics.context_coverage": MagicMock(),
+    },
 ):
     from cognee.eval_framework.evaluation.deep_eval_adapter import DeepEvalAdapter