Merge pull request #1990 from Giskard-AI/GSK-3609-Avoid-redundant-questions-in-data-generation

GSK-3609  Avoid redundant questions in data generation

Author: henchaves

giskard/rag/knowledge_base.py

@@ -300,6 +300,17 @@ def get_failure_plot(self, question_evaluation: Sequence[dict] = None):
     def get_random_document(self):
         return self._rng.choice(self._documents)
 
+    def get_random_documents(self, n: int, with_replacement=False):
+        if with_replacement:
+            return list(self._rng.choice(self._documents, n, replace=True))
+
+        docs = list(self._rng.choice(self._documents, min(n, len(self._documents)), replace=False))
+
+        if len(docs) <= n:
+            docs.extend(self._rng.choice(self._documents, n - len(docs), replace=True))
+
+        return docs
+
     def get_neighbors(self, seed_document: Document, n_neighbors: int = 4, similarity_threshold: float = 0.2):
         seed_embedding = seed_document.embeddings
 

giskard/rag/question_generators/base.py

@@ -52,9 +52,11 @@ class GenerateFromSingleQuestionMixin:
     _question_type: str
 
     def generate_questions(self, knowledge_base: KnowledgeBase, num_questions: int, *args, **kwargs) -> Iterator[Dict]:
-        for _ in range(num_questions):
+        docs = knowledge_base.get_random_documents(num_questions)
+
+        for doc in docs:
             try:
-                yield self.generate_single_question(knowledge_base, *args, **kwargs)
+                yield self.generate_single_question(knowledge_base, *args, **kwargs, seed_document=doc)
             except Exception as e:  # @TODO: specify exceptions
                 logger.error(f"Encountered error in question generation: {e}. Skipping.")
                 logger.exception(e)

giskard/rag/question_generators/double_questions.py

@@ -92,9 +92,9 @@ class DoubleQuestionsGenerator(GenerateFromSingleQuestionMixin, _LLMBasedQuestio
     _question_type = "double"
 
     def generate_single_question(
-        self, knowledge_base: KnowledgeBase, agent_description: str, language: str
+        self, knowledge_base: KnowledgeBase, agent_description: str, language: str, seed_document=None
     ) -> QuestionSample:
-        seed_document = knowledge_base.get_random_document()
+        seed_document = seed_document or knowledge_base.get_random_document()
         context_documents = knowledge_base.get_neighbors(
             seed_document, self._context_neighbors, self._context_similarity_threshold
         )

giskard/rag/question_generators/oos_questions.py

@@ -68,7 +68,7 @@ class OutOfScopeGenerator(GenerateFromSingleQuestionMixin, _LLMBasedQuestionGene
     _question_type = "out of scope"
 
     def generate_single_question(
-        self, knowledge_base: KnowledgeBase, agent_description: str, language: str
+        self, knowledge_base: KnowledgeBase, agent_description: str, language: str, seed_document=None
     ) -> QuestionSample:
         """
         Generate a question from a list of context documents.
@@ -87,7 +87,7 @@ def generate_single_question(
         Tuple[dict, dict]
             The generated question and the metadata of the question.
         """
-        seed_document = knowledge_base.get_random_document()
+        seed_document = seed_document or knowledge_base.get_random_document()
 
         context_documents = knowledge_base.get_neighbors(
             seed_document, self._context_neighbors, self._context_similarity_threshold

giskard/rag/question_generators/simple_questions.py

@@ -66,7 +66,13 @@ class SimpleQuestionsGenerator(GenerateFromSingleQuestionMixin, _LLMBasedQuestio
 
     _question_type = "simple"
 
-    def generate_single_question(self, knowledge_base: KnowledgeBase, agent_description: str, language: str) -> dict:
+    def generate_single_question(
+        self,
+        knowledge_base: KnowledgeBase,
+        agent_description: str,
+        language: str,
+        seed_document=None,
+    ) -> dict:
         """
         Generate a question from a list of context documents.
 
@@ -80,7 +86,8 @@ def generate_single_question(self, knowledge_base: KnowledgeBase, agent_descript
         QuestionSample
             The generated question and the metadata of the question.
         """
-        seed_document = knowledge_base.get_random_document()
+        seed_document = seed_document or knowledge_base.get_random_document()
+
         context_documents = knowledge_base.get_neighbors(
             seed_document, self._context_neighbors, self._context_similarity_threshold
         )

tests/rag/test_knowledge_base.py

@@ -8,6 +8,32 @@
 from giskard.rag.knowledge_base import KnowledgeBase
 
 
+def test_knowledge_base_get_random_documents():
+    llm_client = Mock()
+    embeddings = Mock()
+    embeddings.embed.side_effect = [np.random.rand(5, 10), np.random.rand(3, 10)]
+
+    kb = KnowledgeBase.from_pandas(
+        df=pd.DataFrame({"text": ["This is a test string"] * 5}), llm_client=llm_client, embedding_model=embeddings
+    )
+
+    # Test when k is smaller than the number of documents
+    docs = kb.get_random_documents(3)
+    assert len(docs) == 3
+    # Check that all document IDs are unique
+    assert len(set(doc.id for doc in docs)) == len(docs)
+
+    # Test when k is equal to the number of documents
+    docs = kb.get_random_documents(5)
+    assert len(docs) == 5
+    assert all([doc == kb[doc.id] for doc in docs])
+
+    # Test when k is larger than the number of documents
+    docs = kb.get_random_documents(10)
+    assert len(docs) == 10
+    assert all([doc == kb[doc.id] for doc in docs])
+
+
 def test_knowledge_base_creation_from_df():
     dimension = 8
     df = pd.DataFrame(["This is a test string"] * 5)

tests/rag/test_question_generators.py

@@ -33,6 +33,7 @@ def test_simple_question_generation():
             Document(dict(content="Milk is produced by cows, goats or sheep.")),
         ]
     knowledge_base.get_random_document = Mock(return_value=documents[0])
+    knowledge_base.get_random_documents = Mock(return_value=documents)
     knowledge_base.get_neighbors = Mock(return_value=documents)
 
     question_generator = SimpleQuestionsGenerator(llm_client=llm_client)
@@ -212,6 +213,7 @@ def test_double_question_generation():
             Document(dict(content="Milk is produced by cows, goats or sheep.")),
         ]
     knowledge_base.get_random_document = Mock(return_value=documents[0])
+    knowledge_base.get_random_documents = Mock(return_value=documents)
     knowledge_base.get_neighbors = Mock(return_value=documents)
 
     question_generator = DoubleQuestionsGenerator(llm_client=llm_client)
@@ -304,6 +306,7 @@ def test_oos_question_generation():
             dict(content="Paul Graham liked to buy a baguette every day at the local market."), doc_id="1"
         )
     )
+    knowledge_base.get_random_documents = Mock(return_value=documents)
     knowledge_base.get_neighbors = Mock(return_value=documents)
 
     question_generator = OutOfScopeGenerator(llm_client=llm_client)

tests/rag/test_testset_generator.py

@@ -177,6 +177,8 @@ def test_question_generation_fail(caplog):
     knowledge_base.__getitem__ = lambda obj, idx: documents[0]
     knowledge_base.topics = ["Cheese", "Ski"]
 
+    knowledge_base.get_random_documents = Mock(return_value=documents)
+
     simple_gen = Mock()
     simple_gen.generate_questions.return_value = [q1, q2]
     failing_gen = SimpleQuestionsGenerator(llm_client=Mock())