new model

app.py

@@ -4,6 +4,7 @@
   - 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 ✅
+  - Language selector: Auto, English, German, French, Italian, Portuguese, Hindi, Spanish, Thai ✅
   - No JSON output leakage ✅
 =========================================================
 """
@@ -19,7 +20,7 @@ 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)")
-# Optional: more reproducible CPU math (silences some TF logs)
+# Optional: reproducible CPU math (silences some TF logs)
 # os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")
 
 import gradio as gr
@@ -51,7 +52,7 @@ except ImportError:
 
 # Retrieval utilities
 from langchain.retrievers import ContextualCompressionRetriever, EnsembleRetriever
-from langchain.retrievers.document_compressors import DocumentCompressorPipeline, CrossEncoderReranker
+from langchain.retrievers.document_compressors import CrossEncoderReranker
 from langchain.retrievers.multi_query import MultiQueryRetriever
 
 from langchain_community.retrievers import BM25Retriever
@@ -157,26 +158,17 @@ if use_bm25:
 else:
     base_retriever = vec_retriever
 
-# Fine-grained compressor (splitter)
-try:
-    from langchain_text_splitters import TokenTextSplitter
-    splitter_for_compression = TokenTextSplitter(chunk_size=220, chunk_overlap=30)  # needs tiktoken
-except Exception:
-    from langchain_text_splitters import RecursiveCharacterTextSplitter as FallbackSplitter
-    splitter_for_compression = FallbackSplitter(chunk_size=300, chunk_overlap=50)
-
-compressor_pipeline = DocumentCompressorPipeline(transformers=[splitter_for_compression])
-
 # ======================================
-# 3) PROMPT (NO JSON INSTRUCTIONS)
+# 3) PROMPT (with target language variable)
 # ======================================
 SYSTEM_TEMPLATE = (
-    "Eres Greta, una asistente bilingüe (ES/EN) experta en reciclaje y sostenibilidad. "
-    "Responde en el idioma del usuario, de forma directa, práctica y basada en los fragmentos. "
-    "Si la información no está en los fragmentos, dilo claramente y sugiere pasos útiles. "
-    "No inventes datos ni menciones la palabra 'fragmentos'.\n\n"
+    "You are Greta, a bilingual recycling & sustainability assistant.\n"
+    "- Always answer in the *target language*: {target_language}.\n"
+    "- If target_language is 'Auto', detect the user's language and answer in that language.\n"
+    "- Be direct, practical, and base your answer only on the snippets below; if they are insufficient, say so and propose actionable next steps.\n"
+    "- Do not reveal or mention 'snippets' or internal tools.\n\n"
     "{context}\n\n"
-    "Pregunta: {question}"
+    "Question: {question}"
 )
 qa_prompt = ChatPromptTemplate.from_template(SYSTEM_TEMPLATE)
 
@@ -205,10 +197,10 @@ memory = ConversationBufferMemory(
     return_messages=True,
 )
 
-# Multi-Query to boost recall
+# Multi-Query boosts recall by generating paraphrases
 mqr = MultiQueryRetriever.from_llm(retriever=base_retriever, llm=llm, include_original=True)
 
-# Cross-encoder reranker (lighter)
+# Cross-encoder reranker (lighter/faster than large)
 cross_encoder = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-base")
 reranker = CrossEncoderReranker(model=cross_encoder, top_n=4)
 
@@ -225,16 +217,16 @@ 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
+    return_source_documents=False,
 )
 
-def chat_interface(question, history):
-    """Wrap the RAG chain to return a clean text answer."""
+def chat_interface(question: str, history, target_language: str = "Auto"):
+    """Wrap the RAG chain to return a clean text answer in the requested language."""
     try:
-        result = qa_chain.invoke({"question": question})
+        result = qa_chain.invoke({"question": question, "target_language": target_language})
         answer = result.get("answer", "")
         if not answer:
-            return "Lo siento, no pude generar una respuesta útil con los fragmentos disponibles."
+            return "Sorry, I couldn't produce a useful answer from the available information."
         return answer
     except Exception as e:
         return (
@@ -270,9 +262,11 @@ banner_tab_content = """
 banner_tab = gr.Markdown(banner_tab_content)
 
 # ============================
-# 7) Chat tab (Blocks + Chatbot with height) — OPTION A
+# 7) Chat tab (Blocks + Chatbot with height + language selector)
 # ============================
 
+SUPPORTED_LANGS = ["Auto", "English", "German", "French", "Italian", "Portuguese", "Hindi", "Spanish", "Thai"]
+
 # CSS: make chat area taller and widen app a bit
 custom_css = """
 .gradio-container { max-width: 1200px !important; }
@@ -288,16 +282,18 @@ def _user_submit(user_msg, history):
     history = history + [[user_msg, None]]
     return "", history
 
-def _bot_respond(history):
-    """Generate bot answer for the last user turn."""
+def _bot_respond(history, target_language):
+    """Generate bot answer for the last user turn in the requested language."""
     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])
+    answer = chat_interface(user_msg, history[:-1], target_language=target_language or "Auto")
     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>")
+    with gr.Row():
+        lang_sel = gr.Dropdown(SUPPORTED_LANGS, value="Auto", label="Answer language")
     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)
@@ -307,21 +303,21 @@ with gr.Blocks(theme=theme, css=custom_css) as chatbot_gradio_app:
         undo = gr.Button("↩︎ Undo")
         clear = gr.Button("🗑 Clear")
 
-    # Submit via button or Enter
+    # Submit via button or Enter (pass language value into the responder)
     send.click(_user_submit, [msg, chat], [msg, chat], queue=False).then(
-        _bot_respond, [chat], [chat]
+        _bot_respond, [chat, lang_sel], [chat]
     )
     msg.submit(_user_submit, [msg, chat], [msg, chat], queue=False).then(
-        _bot_respond, [chat], [chat]
+        _bot_respond, [chat, lang_sel], [chat]
     )
 
-    # Utilities
+    # Utilities respect current language selection too
     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])
+    ).then(_bot_respond, [chat, lang_sel], [chat])
 
 # ============================
 # 8) Tabs + launch