new model

app.py

@@ -1,11 +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 ✅
-  - Language selector: Auto, English, German, French, Italian, Portuguese, Hindi, Spanish, Thai ✅
-  - No JSON output leakage ✅
+  - Chat tab: Blocks + Chatbot(height=...) ✅
+  - LLM: meta-llama/Meta-Llama-3.1-8B-Instruct ✅
+  - RAG: e5-base-v2 + (BM25+Vector) con fallback + Multi-Query + reranker ✅
+  - Responde en el idioma elegido (sin pasar claves extra) ✅
 =========================================================
 """
 
@@ -13,14 +12,14 @@ import os
 import json
 import shutil
 
-# --- Env / telemetry (set before imports that use them) ---
+# --- Env / telemetry (antes de imports que lo usen) ---
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 os.environ.setdefault("HF_HUB_DISABLE_TELEMETRY", "1")
 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: reproducible CPU math (silences some TF logs)
+# Opcional: resultados CPU más estables de TF
 # os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")
 
 import gradio as gr
@@ -36,7 +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 / RAG ---
+# --- LangChain / RAG ---
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_core.prompts import ChatPromptTemplate
 from langchain.chains import ConversationalRetrievalChain
@@ -46,19 +45,18 @@ from langchain_community.vectorstores import Chroma
 
 # Embeddings
 try:
-    from langchain_huggingface import HuggingFaceEmbeddings  # pip install -U langchain-huggingface
+    from langchain_huggingface import HuggingFaceEmbeddings
 except ImportError:
     from langchain_community.embeddings import HuggingFaceEmbeddings
 
 # Retrieval utilities
 from langchain.retrievers import ContextualCompressionRetriever, EnsembleRetriever
-from langchain.retrievers.document_compressors import CrossEncoderReranker
 from langchain.retrievers.multi_query import MultiQueryRetriever
-
+from langchain.retrievers.document_compressors import CrossEncoderReranker
 from langchain_community.retrievers import BM25Retriever
 from langchain_community.cross_encoders import HuggingFaceCrossEncoder
 
-# HF Hub for SavedModel
+# HF Hub
 from huggingface_hub import snapshot_download
 
 # LLM via HF Inference
@@ -128,7 +126,7 @@ base_splitter = RecursiveCharacterTextSplitter(
 )
 docs = base_splitter.split_documents(all_loaded_docs)
 
-# Embeddings (better recall)
+# Embeddings
 embeddings = HuggingFaceEmbeddings(model_name="intfloat/e5-base-v2")
 
 # Vector store
@@ -143,10 +141,10 @@ vectordb = Chroma.from_documents(
 # Vector retriever
 vec_retriever = vectordb.as_retriever(search_kwargs={"k": 8}, search_type="mmr")
 
-# BM25 + Ensemble with safe fallback if rank-bm25 isn't installed
+# BM25 + Ensemble con fallback si falta rank-bm25
 use_bm25 = True
 try:
-    bm25 = BM25Retriever.from_documents(docs)  # requires rank-bm25
+    bm25 = BM25Retriever.from_documents(docs)  # requiere rank-bm25
     bm25.k = 8
 except Exception as e:
     print(f"[RAG] BM25 unavailable ({e}). Falling back to vector-only retriever.")
@@ -159,14 +157,14 @@ else:
     base_retriever = vec_retriever
 
 # ======================================
-# 3) PROMPT (with target language variable)
+# 3) PROMPT (sin variables extra: solo {context} y {question})
+#    Instruimos al modelo a obedecer un prefijo en la propia pregunta.
 # ======================================
 SYSTEM_TEMPLATE = (
-    "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"
+    "You are Greta, a recycling & sustainability assistant. "
+    "Follow any explicit language directive at the start of the question, e.g., "
+    "‘Answer ONLY in Spanish.’ If there is no directive, detect the user's language and answer accordingly. "
+    "Be direct and practical. If the snippets are insufficient, say so and suggest actionable next steps.\n\n"
     "{context}\n\n"
     "Question: {question}"
 )
@@ -197,10 +195,10 @@ memory = ConversationBufferMemory(
     return_messages=True,
 )
 
-# Multi-Query boosts recall by generating paraphrases
+# Multi-Query (paráfrasis de la consulta)
 mqr = MultiQueryRetriever.from_llm(retriever=base_retriever, llm=llm, include_original=True)
 
-# Cross-encoder reranker (lighter/faster than large)
+# Reranker (cross-encoder base)
 cross_encoder = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-base")
 reranker = CrossEncoderReranker(model=cross_encoder, top_n=4)
 
@@ -220,13 +218,21 @@ qa_chain = ConversationalRetrievalChain.from_llm(
     return_source_documents=False,
 )
 
+# ===== Helper: construir prefijo de idioma en la propia pregunta =====
+def _lang_directive(lang: str) -> str:
+    if not lang or lang.strip().lower() == "auto":
+        return "Detect the user's language and answer in that language."
+    return f"Answer ONLY in {lang}."
+
 def chat_interface(question: str, history, target_language: str = "Auto"):
-    """Wrap the RAG chain to return a clean text answer in the requested language."""
+    """Devuelve respuesta limpia en el idioma solicitado, SIN pasar claves extra al chain."""
     try:
-        result = qa_chain.invoke({"question": question, "target_language": target_language})
+        directive = _lang_directive(target_language)
+        combined_q = f"{directive}\n\n{question}"
+        result = qa_chain.invoke({"question": combined_q})
         answer = result.get("answer", "")
         if not answer:
-            return "Sorry, I couldn't produce a useful answer from the available information."
+            return "Lo siento, no pude generar una respuesta útil con la información disponible."
         return answer
     except Exception as e:
         return (
@@ -267,7 +273,7 @@ banner_tab = gr.Markdown(banner_tab_content)
 
 SUPPORTED_LANGS = ["Auto", "English", "German", "French", "Italian", "Portuguese", "Hindi", "Spanish", "Thai"]
 
-# CSS: make chat area taller and widen app a bit
+# CSS: ampliar área de chat y ancho general
 custom_css = """
 .gradio-container { max-width: 1200px !important; }
 #greta-chat { height: 700px !important; }
@@ -276,16 +282,15 @@ custom_css = """
 """
 
 def _user_submit(user_msg, history):
-    """Append user turn; bot fills later."""
+    """Añade turno del usuario; el bot responde después."""
     if not user_msg:
         return "", history
     history = history + [[user_msg, None]]
     return "", history
 
 def _bot_respond(history, target_language):
-    """Generate bot answer for the last user turn in the requested language."""
+    """Genera la respuesta del bot en el idioma solicitado."""
     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], target_language=target_language or "Auto")
     history[-1][1] = answer
     return history
@@ -303,7 +308,7 @@ 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 (pass language value into the responder)
+    # Envío por botón o Enter (pasamos el idioma al responder)
     send.click(_user_submit, [msg, chat], [msg, chat], queue=False).then(
         _bot_respond, [chat, lang_sel], [chat]
     )
@@ -311,11 +316,11 @@ with gr.Blocks(theme=theme, css=custom_css) as chatbot_gradio_app:
         _bot_respond, [chat, lang_sel], [chat]
     )
 
-    # Utilities respect current language selection too
+    # Utilidades
     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
+        lambda h: (h[:-1] + [[h[-1][0], None]]) if h else h,  # reintenta la última pregunta
         chat, chat, queue=False
     ).then(_bot_respond, [chat, lang_sel], [chat])