final

prompts.py

@@ -27,13 +27,18 @@ RAG_PROMPT = ChatPromptTemplate.from_messages([
     ("human", RAG_HUMAN)
 ])
 
-REFLECTION_SYSTEM = """You are a senior editor grading an AI-generated answer. 
-Evaluate if the answer is:
-1. GROUNDED: Is the answer supported by the facts in the provided Context? 
-2. RELEVANT: Does it actually answer the User Question?
+REFLECTION_SYSTEM = """You are a balanced evaluator. Your goal is to ensure the generated answer is faithful to the context and addresses the user's question.
 
-Output exactly "yes" if the answer is both grounded and relevant.
-Output "no" if the answer contains information NOT in the context, is irrelevant, or if the assistant says it cannot answer."""
+Criteria:
+1. **Groundedness**: Are the core claims supported by the context? (Allow professional phrasing).
+2. **Relevance**: Does the answer address the question? 
+
+Evaluation Guidelines:
+- If the answer is mostly correct but uses different wording, output **yes**.
+- If the answer is an honest "I cannot answer" because info is missing, output **yes**.
+- Only output **no** if there is a CLEAR hallucination or if the answer is completely off-topic.
+
+Output exactly "yes" or "no"."""
 
 REFLECTION_HUMAN = """Context:
 {context}
@@ -41,7 +46,7 @@ REFLECTION_HUMAN = """Context:
 User Question: {question}
 Generated Answer: {generation}
 
-Current Answer Quality status:"""
+Evaluate:"""
 
 REFLECTION_PROMPT = ChatPromptTemplate.from_messages([
     ("system", REFLECTION_SYSTEM),