new model

app.py

@@ -1,6 +1,10 @@
 """
 =========================================================
  app.py — Green Greta (Gradio + TF/Keras 3 + LangChain 0.3)
+  - Chat tab uses Blocks + Chatbot(height=...)  ✅
+  - LLM: meta-llama/Meta-Llama-3.1-8B-Instruct  ✅
+  - RAG: e5-base-v2 + (BM25+Vector) with safe fallback + Multi-Query + reranker ✅
+  - No JSON output leakage ✅
 =========================================================
 """
 
@@ -15,7 +19,7 @@ os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "False")
 os.environ.setdefault("ANONYMIZED_TELEMETRY", "false")
 os.environ.setdefault("CHROMA_TELEMETRY_ENABLED", "FALSE")
 os.environ.setdefault("USER_AGENT", "green-greta/1.0 (+contact-or-repo)")
-# If you want deterministic CPU math from TF (optional):
+# Optional: more reproducible CPU math (silences some TF logs)
 # os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")
 
 import gradio as gr
@@ -31,7 +35,7 @@ except Exception:
     user_agent = "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36"
 header_template = {"User-Agent": user_agent}
 
-# --- LangChain core ---
+# --- LangChain core / RAG ---
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_core.prompts import ChatPromptTemplate
 from langchain.chains import ConversationalRetrievalChain
@@ -221,16 +225,14 @@ qa_chain = ConversationalRetrievalChain.from_llm(
     combine_docs_chain_kwargs={"prompt": qa_prompt},
     get_chat_history=lambda h: h,
     rephrase_question=False,
-    return_source_documents=False,  # <- we only need the final answer
-    # Use default output key "answer" so we don't need to parse JSON
+    return_source_documents=False,  # we only need the final answer
 )
 
 def chat_interface(question, history):
+    """Wrap the RAG chain to return a clean text answer."""
     try:
         result = qa_chain.invoke({"question": question})
-        # ConversationalRetrievalChain returns {"answer": "...", ...}
         answer = result.get("answer", "")
-        # Safety fallback: if empty, return a friendly default
         if not answer:
             return "Lo siento, no pude generar una respuesta útil con los fragmentos disponibles."
         return answer
@@ -268,26 +270,66 @@ banner_tab_content = """
 banner_tab = gr.Markdown(banner_tab_content)
 
 # ============================
-# 7) Gradio app (tabs + run)
+# 7) Chat tab (Blocks + Chatbot with height) — OPTION A
 # ============================
+
+# CSS: make chat area taller and widen app a bit
 custom_css = """
-/* Make the chat area taller without using the height arg */
-.gr-chatbot { min-height: 700px !important; }
-.gr-chatbot > div { min-height: 700px !important; }
 .gradio-container { max-width: 1200px !important; }
+#greta-chat { height: 700px !important; }
+#greta-chat .gr-chatbot { height: 700px !important; min-height: 700px !important; }
+#greta-chat .overflow-y-auto { height: 660px !important; max-height: 660px !important; }
 """
 
-chatbot_gradio_app = gr.ChatInterface(
-    fn=chat_interface,
-    title="<span style='color: rgb(243, 239, 224);'>Green Greta</span>",
-    theme=theme,
-)
+def _user_submit(user_msg, history):
+    """Append user turn; bot fills later."""
+    if not user_msg:
+        return "", history
+    history = history + [[user_msg, None]]
+    return "", history
+
+def _bot_respond(history):
+    """Generate bot answer for the last user turn."""
+    user_msg = history[-1][0]
+    # Pass previous history to our RAG function (excluding the current empty bot turn)
+    answer = chat_interface(user_msg, history[:-1])
+    history[-1][1] = answer
+    return history
+
+with gr.Blocks(theme=theme, css=custom_css) as chatbot_gradio_app:
+    gr.Markdown("<h1 style='text-align:center;color:#f3efe0;'>Green Greta</h1>")
+    chat = gr.Chatbot(label="Chatbot", height=700, elem_id="greta-chat", show_copy_button=True)
+    with gr.Row():
+        msg = gr.Textbox(placeholder="Type a message…", scale=9)
+        send = gr.Button("Submit", scale=1)
+    with gr.Row():
+        retry = gr.Button("↻ Retry")
+        undo = gr.Button("↩︎ Undo")
+        clear = gr.Button("🗑 Clear")
+
+    # Submit via button or Enter
+    send.click(_user_submit, [msg, chat], [msg, chat], queue=False).then(
+        _bot_respond, [chat], [chat]
+    )
+    msg.submit(_user_submit, [msg, chat], [msg, chat], queue=False).then(
+        _bot_respond, [chat], [chat]
+    )
+
+    # Utilities
+    clear.click(lambda: [], None, chat, queue=False)
+    undo.click(lambda h: h[:-1] if h else h, chat, chat, queue=False)
+    retry.click(
+        lambda h: (h[:-1] + [[h[-1][0], None]]) if h else h,  # re-ask last user msg
+        chat, chat, queue=False
+    ).then(_bot_respond, [chat], [chat])
 
+# ============================
+# 8) Tabs + launch
+# ============================
 app = gr.TabbedInterface(
     [banner_tab, image_gradio_app, chatbot_gradio_app],
     tab_names=["Welcome to Green Greta", "Green Greta Image Classification", "Green Greta Chat"],
     theme=theme,
-    css=custom_css,
 )
 
 app.queue()