log the data

app.py

@@ -5,7 +5,10 @@ from langchain.embeddings import HuggingFaceEmbeddings
 from langchain.vectorstores import FAISS
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 import json
+from utils.helpers import retrieve_context, upload_log_to_hf, log_interaction_hf
 
+turn_counter = 0
+UPLOAD_INTERVAL = 5
 
 with open("config.json") as f:
     config = json.load(f)
@@ -13,86 +16,12 @@ with open("config.json") as f:
 # load the results
 token = config["hf"] + config["token"] 
 
-
-# === Step 0: Download FAISS index files if not present ===
-
-def download_faiss_index(repo_id="kaburia/epic-a-embeddings", local_folder="faiss_index"):
-    os.makedirs(local_folder, exist_ok=True)
-
-    index_faiss_path = os.path.join(local_folder, "index.faiss")
-    index_pkl_path = os.path.join(local_folder, "index.pkl")
-
-    if not os.path.exists(index_faiss_path):
-        print("Downloading index.faiss from Hugging Face Dataset...")
-        hf_hub_download(
-            repo_id=repo_id,
-            filename="index.faiss",
-            repo_type="dataset",          # 🛑 MUST add this line
-            local_dir=local_folder,
-            local_dir_use_symlinks=False,
-        )
-
-    if not os.path.exists(index_pkl_path):
-        print("Downloading index.pkl from Hugging Face Dataset...")
-        hf_hub_download(
-            repo_id=repo_id,
-            filename="index.pkl",
-            repo_type="dataset",          # 🛑 MUST add this line
-            local_dir=local_folder,
-            local_dir_use_symlinks=False,
-        )
-
-# === Step 1: Load Vectorstore ===
-
-def load_vectorstore(index_path="faiss_index"):
-    embedding_model = HuggingFaceEmbeddings(
-        model_name="sentence-transformers/all-MiniLM-L6-v2"
-    )
-    db = FAISS.load_local(
-        index_path,
-        embeddings=embedding_model,
-        allow_dangerous_deserialization=True
-    )
-    return db
-
-# Download FAISS index if needed
-download_faiss_index()
-
-vectorstore = load_vectorstore()
-
-
 # === Step 2: Setup HuggingFace Inference API ===
 
 client = InferenceClient("HuggingFaceH4/zephyr-7b-beta", 
                          token=token)
 
 
-# === Step 3: Build Retrieval-Augmented Response Function ===
-
-# def retrieve_context(question, k=5):
-#     retriever = vectorstore.as_retriever(search_kwargs={"k": k})
-#     docs = retriever.get_relevant_documents(question)
-#     context = "\n\n".join(doc.page_content for doc in docs)
-#     return context
-
-def cosine_to_prob(score):
-    # Convert cosine similarity from [-1, 1] to [0, 1]
-    return (score + 1) / 2
-
-def retrieve_context(question, p=5, threshold=0.5):
-    # Get docs with raw scores
-    results = vectorstore.similarity_search_with_score(question, k=50)  # get more than needed
-
-    # Filter for "probability" above threshold
-    filtered = [(doc, score) for doc, score in results if cosine_to_prob(score) > threshold]
-
-    # Sort by score descending and take top-p
-    top_p_docs = sorted(filtered, key=lambda x: x[1], reverse=True)[:p]
-
-    # Join content for prompt
-    context = "\n\n".join(doc.page_content for doc, _ in top_p_docs)
-    return context
-
 def detect_intent(message: str) -> str:
     """Classify the message as 'small_talk' or 'info_query' using the model."""
     prompt = f"""You are a classifier. Categorize the following user message as either 'small_talk' or 'info_query'.
@@ -107,39 +36,43 @@ def detect_intent(message: str) -> str:
     return "info_query"
 
 def respond(message, history, system_message, max_tokens, temperature, top_p):
+    global turn_counter
     intent = detect_intent(message)
 
+    original_user_message = message  # preserve the real user input for logging
     messages = []
+
     if intent == "small_talk":
-        # Free conversation
         messages.append({"role": "system", "content": "You are a friendly assistant. Talk naturally and helpfully."})
     else:
-        # Retrieval + system constraints
-        context = retrieve_context(message, p=5, threshold=0.5)
+        context = retrieve_context(original_user_message, p=5, threshold=0.5)
         messages.append({"role": "system", "content": system_message})
-        message = f"""Use the following context to answer the question.
+        prompt = f"""Use the following context to answer the question.
 
-                - ONLY quote exact text from the context.
-                - Do NOT summarize, paraphrase, or infer anything.
-                - If no answer is found, respond: "The answer is not in the provided context."
+        - ONLY quote exact text from the context.
+        - Do NOT summarize, paraphrase, or infer anything.
+        - If no answer is found, respond: "The answer is not in the provided context."
 
-                Context:
-                {context}
+        Context:
+        {context}
 
-                Question: {message}"""
+        Question: {original_user_message}
+        """
+        original_user_message = prompt  # what gets sent to the model
 
-    # Add chat history
+    # Load prior chat memory
     for user, assistant in history:
         if user:
             messages.append({"role": "user", "content": user})
         if assistant:
             messages.append({"role": "assistant", "content": assistant})
 
-    # Final user input
-    messages.append({"role": "user", "content": message})
+    # Add current turn
+    messages.append({"role": "user", "content": original_user_message})
 
-    # Stream response
+    # Generate and stream response
     response = ""
+    full_response = ""
     for chunk in client.chat_completion(
         messages=messages,
         max_tokens=max_tokens,
@@ -150,8 +83,67 @@ def respond(message, history, system_message, max_tokens, temperature, top_p):
         token = chunk.choices[0].delta.content
         if token:
             response += token
+            full_response += token
             yield response
 
+    # Log this interaction
+    log_interaction_hf(message, full_response)
+
+    # Upload logs to Hugging Face every N turns
+    turn_counter += 1
+    if turn_counter % UPLOAD_INTERVAL == 0:
+        try:
+            upload_log_to_hf(token)
+        except Exception as e:
+            print(f"❌ Log upload failed: {e}")
+
+
+# def respond(message, history, system_message, max_tokens, temperature, top_p):
+#     intent = detect_intent(message)
+
+#     messages = []
+#     if intent == "small_talk":
+#         # Free conversation
+#         messages.append({"role": "system", "content": "You are a friendly assistant. Talk naturally and helpfully."})
+#     else:
+#         # Retrieval + system constraints
+#         context = retrieve_context(message, p=5, threshold=0.5)
+#         messages.append({"role": "system", "content": system_message})
+#         message = f"""Use the following context to answer the question.
+
+#                 - ONLY quote exact text from the context.
+#                 - Do NOT summarize, paraphrase, or infer anything.
+#                 - If no answer is found, respond: "The answer is not in the provided context."
+
+#                 Context:
+#                 {context}
+
+#                 Question: {message}"""
+
+#     # Add chat history
+#     for user, assistant in history:
+#         if user:
+#             messages.append({"role": "user", "content": user})
+#         if assistant:
+#             messages.append({"role": "assistant", "content": assistant})
+
+#     # Final user input
+#     messages.append({"role": "user", "content": message})
+
+#     # Stream response
+#     response = ""
+#     for chunk in client.chat_completion(
+#         messages=messages,
+#         max_tokens=max_tokens,
+#         temperature=temperature,
+#         top_p=top_p,
+#         stream=True,
+#     ):
+#         token = chunk.choices[0].delta.content
+#         if token:
+#             response += token
+#             yield response
+
 
 # def respond(message, history, 
 #             system_message, max_tokens,