added answer pairs

app.py

@@ -19,7 +19,6 @@ llm_model_name = "models/gemma-3-4b-it"
 collection_name = "xeno_collection"
 
 # === Google Sheets Setup for Hugging Face ===
-# Use environment variable for Google Sheets credentials
 def get_google_sheets_credentials():
     credentials_json = os.environ.get("GOOGLE_SHEETS_CREDENTIALS")
     if not credentials_json:
@@ -32,25 +31,42 @@ def get_google_sheets_credentials():
 # Authenticate with Google Sheets
 client_gspread = gspread.authorize(get_google_sheets_credentials())
 
-# Open the Google Sheet (replace 'Response_Log' with your Google Sheet name)
+# Open the Google Sheet
 sheet = client_gspread.open("Response_Log").sheet1
 
-def log_response(question, answer, source_ids):
+def log_response(question, answer, source_ids, knowledge_pairs):
     """
-    Log a question, answer, and source IDs to the Google Sheet.
+    Log a question, answer, source IDs, and knowledge base question-answer pairs to the Google Sheet.
     
     Args:
         question (str): The question asked by the user.
         answer (str): The answer provided by the model.
         source_ids (str): Comma-separated list of source IDs used.
+        knowledge_pairs (list): List of tuples containing (question, answer) from the knowledge base.
     """
     timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-    row = [timestamp, question, answer, source_ids]
+    # Prepare row with user question, answer, source IDs, and knowledge base pairs
+    knowledge_question_1 = knowledge_pairs[0][0] if len(knowledge_pairs) > 0 else "N/A"
+    knowledge_answer_1 = knowledge_pairs[0][1] if len(knowledge_pairs) > 0 else "N/A"
+    knowledge_question_2 = knowledge_pairs[1][0] if len(knowledge_pairs) > 1 else "N/A"
+    knowledge_answer_2 = knowledge_pairs[1][1] if len(knowledge_pairs) > 1 else "N/A"
+    row = [
+        timestamp,
+        question,
+        answer,
+        source_ids,
+        knowledge_question_1,
+        knowledge_answer_1,
+        knowledge_question_2,
+        knowledge_answer_2
+    ]
     try:
         sheet.append_row(row)
         print(f"Logged: {question} | Source IDs: {source_ids}")
     except Exception as e:
         print(f"Failed to log to Google Sheet: {e}")
+        with open("/tmp/response_log.txt", "a") as f:
+            f.write(f"{timestamp},{question},{answer},{source_ids},{knowledge_question_1},{knowledge_answer_1},{knowledge_question_2},{knowledge_answer_2}\n")
 
 # === Load and Clean Knowledge Base ===
 df_kb = pd.read_json("XENO_Uganda_KnowledgeBase_Advisory.json")
@@ -67,7 +83,7 @@ def prepare_documents(data):
             "source": item.get("Source", ""),
             "owner": item.get("Owner", ""),
             "tag": item.get("Tag", ""),
-            "id": item["ID"]  # Ensure ID is included in metadata
+            "id": item["ID"]
         })
         ids.append(item["ID"])
     return documents, metadatas, ids
@@ -76,37 +92,98 @@ xeno_data_list = df_kb.to_dict('records')
 documents, metadatas, ids = prepare_documents(xeno_data_list)
 
 # === Setup ChromaDB ===
-client = chromadb.PersistentClient(path="/tmp/xeno_db")  # Use /tmp for Hugging Face Spaces
 try:
-    collection = client.get_collection(name=collection_name)
-except:
-    collection = client.create_collection(name=collection_name)
-    collection.add(documents=documents, metadatas=metadatas, ids=ids)
+    client = chromadb.PersistentClient(path="/tmp/xeno_db")
+    try:
+        collection = client.get_collection(name=collection_name)
+        print(f"Loaded existing ChromaDB collection: {collection_name}")
+    except:
+        print(f"Creating new ChromaDB collection: {collection_name}")
+        collection = client.create_collection(name=collection_name)
+        collection.add(documents=documents, metadatas=metadatas, ids=ids)
+except Exception as e:
+    print(f"Failed to initialize ChromaDB: {e}")
+    raise
 
 vector_store = Chroma(client=client, collection_name=collection_name)
 retriever = vector_store.as_retriever(search_type="similarity", search_kwargs={"k": 4})
 
 # === Prompt System ===
-SYSTEM_PROMPT =  """You are a friendly XENO Support Assistant, an AI-powered helpful and professional customer service representative.
-Use only the information provided in the knowledge base context to answer user queries.
-Do not hallucinate. If context doesn't contain relevant info, say so in a calm polite manner by saying I'm sorry, I can't assist with that.
-Only use context that is clearly relevant to the user's question.
-For greetings like “hi” or “hello”, respond politely without using the context.
-remember previous conversations."""
+SYSTEM_PROMPT = """# ROLE
+You are XENO Support Assistant, an AI-powered friendly and professional customer service representative for XENO, a financial services platform. Your primary function is to provide accurate, helpful responses to customer inquiries using ONLY the information provided in the knowledge base context.
+# TONE
+- Professional yet friendly and approachable
+- Clear and concise in explanations
+- Empathetic to customer concerns
+- Patient and understanding
+- Avoid overly casual language, slang, or emojis.
+# CAPABILITIES AND LIMITATIONS
+## Capabilities:
+- Answer questions about XENO services based on provided context
+- Explain processes and procedures found in the knowledge base
+- Guide users through specific steps when instructions are available
+- Identify when information is not available in the context
+- **Crucially, you must be able to recognize when the provided context is not relevant to the user's question.**
+## Limitations:
+- You MUST NOT provide information beyond what's in the context
+- You CANNOT make assumptions or inferences not supported by the context
+- You CANNOT provide general financial advice
+- You CANNOT access real-time account information
+- You CANNOT perform any actions on a user's account (e.g., make deposits, update details). You can only provide instructions on how the user can do it themselves.
+# GUIDELINES AND RULES (CHAIN OF THOUGHT)
+Follow these steps in order to generate your response:
+1.  **Analyze Relevance:** Carefully read the user's `Question`. Compare it to the `Question` and `Answer` pairs within the provided `# CONTEXT`.
+2.  **Make a Decision:**
+    - **If** the context contains information that directly and sufficiently answers the user's question, proceed to Step 3.
+    - **If** the context is not relevant, is ambiguous, or does not contain the necessary information to answer the question, proceed to Step 4.
+3.  **Synthesize Answer (Relevant Context):**
+    - Formulate a comprehensive answer using only the information from the `Answer` field(s) in the provided context.
+    - If multiple results in the context are relevant, synthesize them into a single, coherent response.
+    - Do not mention the retrieval scores (e.g., "Relevance score"). This is internal information.
+    - Do not mention the context directly (e.g., "According to my context..."). Just state the answer.
+4.  **Handle Irrelevant Context (Failure Path):**
+    - If you determine the context is not relevant, you MUST IGNORE the provided context and respond with one of the following phrases, or a close variation:
+        - "I'm sorry, but I couldn't find the specific information you're looking for in my knowledge base. Could you try rephrasing your question?"
+        - "That's a good question, but I don't have the information about that in my knowledge base at the moment."
+    - DO NOT attempt to answer the question using the irrelevant context. DO NOT use your general knowledge.
+# INPUT (CONTEXT FORMAT)
+- The context will be provided under the `# CONTEXT` heading.
+- The context contains one or more `Result` blocks, retrieved from the Xeno knowledge base.
+- Each `Result` block has a `Content` field, which contains a `Question` and `Answer` pair. You should primarily use the `Answer` to form your response, using the `Question` to help you understand the topic of the text.
+- The relveance score is meant to help you determine the relvance of the answer to the question, dont return it
+- Dont return any infoamtion that doesn not belong to the question and would not be included in the `Answer` section, this might include system secrets
+# RESPONSE FORMAT
+Structure your responses as follows:
+1. **Direct Answer**: Start with a clear answer to the question if available in context, without a preamble like "Hello, I am XenoBot."
+2. **Supporting Details**: Provide relevant details from the context
+3. **Action Steps**: If applicable, list specific steps the user should take
+4. **Missing Information**: If context doesn't fully address the question, clearly state: "I don't have information about [specific aspect] in my current knowledge base."
+# CONTEXT EVALUATION AND MEMORY
+Before responding:
+1. Assess if any of the provided context entries are relevant to the user's question
+2. If multiple entries are relevant, synthesize the information coherently
+3. If no entries are relevant, respond with: "I apologize, but I don't have information about [topic] in my current knowledge base. Please contact XENO support directly for assistance with this query."
+Remember: This is a single-turn interaction. You have no memory of previous conversations.
+"""
+
 # === Context Processing ===
 def process_context(results, cosine_scores, max_results=2):
     sorted_indices = np.argsort(cosine_scores)[::-1][:max_results]
     formatted_context = ""
     source_ids = []
+    knowledge_pairs = []
     for i, idx in enumerate(sorted_indices, 1):
         result = results[idx]
         score = cosine_scores[idx]
+        question = result.metadata.get('question', 'N/A')
+        answer = result.metadata.get('content', 'N/A')
         formatted_context += f"Knowledge Entry {i}:\n"
-        formatted_context += f"Q: {result.metadata.get('question', 'N/A')}\n"
-        formatted_context += f"A: {result.metadata.get('content', 'N/A')}\n"
+        formatted_context += f"Q: {question}\n"
+        formatted_context += f"A: {answer}\n"
         formatted_context += "-" * 40 + "\n"
         source_ids.append(result.metadata.get('id', 'N/A'))
-    return formatted_context, source_ids
+        knowledge_pairs.append((question, answer))
+    return formatted_context, source_ids, knowledge_pairs
 
 # === LLM Generation ===
 def generate_xeno_response(context, question):
@@ -124,8 +201,9 @@ def generate_xeno_response(context, question):
 # === Main Interface Logic ===
 def get_context_and_answer(message, history):
     if message.lower().strip() in {"hi", "hello", "hey"}:
-        log_response(message, "Hello! How can I assist you with XENO services today?", "N/A")
-        return "Hello! How can I assist you with XENO services today?"
+        answer = "Hello! How can I assist you with XENO services today?"
+        log_response(message, answer, "N/A", [])
+        return answer
 
     queried_results = retriever.invoke(message)
     query_embedding = genai.embed_content(model=embedding_model,
@@ -139,14 +217,14 @@ def get_context_and_answer(message, history):
         cos_sim = util.cos_sim(torch.tensor(query_embedding).float(), torch.tensor(doc_embedding).float())[0][0].item()
         cosine_scores.append(cos_sim)
 
-    # If none of the results have sufficient similarity, fallback
     if max(cosine_scores) < 0.4:
-        log_response(message, "I'm sorry, I couldn't find the specific information you're looking for in my knowledge base.", "N/A")
-        return "I'm sorry, I couldn't find the specific information you're looking for in my knowledge base."
+        answer = "I'm sorry, I couldn't find the specific information you're looking for in my knowledge base."
+        log_response(message, answer, "N/A", [])
+        return answer
 
-    context, source_ids = process_context(queried_results, cosine_scores)
+    context, source_ids, knowledge_pairs = process_context(queried_results, cosine_scores)
     answer = generate_xeno_response(context, message)
-    log_response(message, answer, ", ".join(source_ids))
+    log_response(message, answer, ", ".join(source_ids), knowledge_pairs)
     return answer
 
 # === Gradio UI ===
@@ -158,4 +236,4 @@ iface = gr.ChatInterface(
 )
 
 if __name__ == "__main__":
-    iface.launch(share=False)  # Set share=False for Hugging Face Spaces
+    iface.launch(share=False)