new model

app.py

@@ -25,7 +25,7 @@ from langchain.memory import ConversationBufferMemory
 from langchain_community.document_loaders import WebBaseLoader
 from langchain_community.vectorstores import Chroma
 
-# Embeddings (prefer langchain-huggingface if installed; fallback to community)
+# Embeddings (prefer langchain-huggingface if installed; fallback a community)
 try:
     from langchain_huggingface import HuggingFaceEmbeddings  # pip install -U langchain-huggingface
 except ImportError:
@@ -37,12 +37,11 @@ from langchain.retrievers.document_compressors import DocumentCompressorPipeline
 
 from pydantic.v1 import BaseModel, Field
 
-# HF Hub for downloading the SavedModel once
+# HF Hub for downloading the SavedModel once (image classifier)
 from huggingface_hub import snapshot_download
 
-# Local transformers pipeline (no API token required)
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline  # (still imported; not used in Hub mode)
-from langchain_community.llms import HuggingFacePipeline, HuggingFaceHub  # <-- ADDED: HuggingFaceHub import
+# === LLM endpoint moderno (compatible con huggingface_hub>=0.23) ===
+from langchain_huggingface import HuggingFaceEndpoint  # Opción 1
 
 # Theming + URL list
 import theme
@@ -50,21 +49,27 @@ from url_list import URLS
 
 theme = theme.Theme()
 
+# (Opcional) reducir telemetría/ruido en logs de Space
+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")
+
 
 # =========================================================
 # 1) IMAGE CLASSIFICATION — Keras 3-safe SavedModel loading
 # =========================================================
 MODEL_REPO = "rocioadlc/efficientnetB0_trash"
-MODEL_SERVING_SIGNATURE = "serving_default"  # adjust if the model exposes a different endpoint
+MODEL_SERVING_SIGNATURE = "serving_default"  # ajusta si el modelo expone otra firma
 
-# Download the model snapshot and wrap it via TFSMLayer (Keras 3 compatible)
+# Descarga el snapshot y envuélvelo con TFSMLayer (compatible Keras 3)
 model_dir = snapshot_download(MODEL_REPO)
 image_model = keras.layers.TFSMLayer(model_dir, call_endpoint=MODEL_SERVING_SIGNATURE)
 
 class_labels = ["cardboard", "glass", "metal", "paper", "plastic", "trash"]
 
 def predict_image(input_image: Image.Image):
-    """Preprocess to EfficientNetB0 input (224x224) and run inference."""
+    """Preprocess a EfficientNetB0 (224x224) y ejecuta inferencia."""
     img = input_image.convert("RGB").resize((224, 224))
     x = tf.keras.preprocessing.image.img_to_array(img)
     x = tf.keras.applications.efficientnet.preprocess_input(x)
@@ -112,7 +117,7 @@ def safe_load_all_urls(urls):
 
 all_loaded_docs = safe_load_all_urls(URLS)
 
-# Smaller base chunks so downstream compression has less work
+# Chunks base pequeños para que el compresor downstream trabaje menos
 base_splitter = RecursiveCharacterTextSplitter(
     chunk_size=700,
     chunk_overlap=80,
@@ -136,13 +141,11 @@ vectordb = Chroma.from_documents(
 # Base retriever
 retriever = vectordb.as_retriever(search_kwargs={"k": 2}, search_type="mmr")
 
-# --- Context compression to keep inputs under FLAN-T5 512-token limit ---
-# Prefer token-aware splitter; fall back to char splitter if `tiktoken` isn't installed.
+# --- Compresión de contexto para entradas ≤ ~512 tokens (t5/…); útil igual con Mixtral ---
 try:
     from langchain_text_splitters import TokenTextSplitter
-    splitter_for_compression = TokenTextSplitter(chunk_size=200, chunk_overlap=30)  # needs `tiktoken`
+    splitter_for_compression = TokenTextSplitter(chunk_size=200, chunk_overlap=30)  # requiere tiktoken
 except Exception:
-    # Fallback that doesn't require tiktoken
     from langchain_text_splitters import RecursiveCharacterTextSplitter as FallbackSplitter
     splitter_for_compression = FallbackSplitter(chunk_size=300, chunk_overlap=50)
 
@@ -170,8 +173,6 @@ SYSTEM_TEMPLATE = (
     "{format_instructions}"
 )
 
-# NOTE: Your original pattern kept; if you prefer, you can also do:
-# ChatPromptTemplate.from_template(SYSTEM_TEMPLATE).partial(format_instructions=parser.get_format_instructions())
 qa_prompt = ChatPromptTemplate.from_template(
     SYSTEM_TEMPLATE,
     partial_variables={"format_instructions": parser.get_format_instructions()},
@@ -179,11 +180,10 @@ qa_prompt = ChatPromptTemplate.from_template(
 
 
 # =============================
-# 4) LLM — HuggingFace Hub (Mixtral)
+# 4) LLM — Hugging Face Inference API (Mixtral)
 # =============================
-# REQUIREMENT: set env var HUGGINGFACEHUB_API_TOKEN
-# (Settings → Variables & secrets in your Space)
-llm = HuggingFaceHub(
+# Requiere el secreto HUGGINGFACEHUB_API_TOKEN en el Space
+llm = HuggingFaceEndpoint(
     repo_id="mistralai/Mixtral-8x7B-v0.1",
     task="text-generation",
     model_kwargs={
@@ -191,8 +191,10 @@ llm = HuggingFaceHub(
         "top_k": 30,
         "temperature": 0.1,
         "repetition_penalty": 1.03,
-        # You may also pass: "return_full_text": False
+        "return_full_text": False,
     },
+    huggingfacehub_api_token=os.getenv("HUGGINGFACEHUB_API_TOKEN"),
+    timeout=120,  # opcional
 )
 
 
@@ -216,7 +218,7 @@ qa_chain = ConversationalRetrievalChain.from_llm(
 )
 
 def _safe_json_extract(raw: str, question: str) -> dict:
-    """Try strict JSON; otherwise extract first {...}; fallback to plain text."""
+    """Intenta JSON estricto; si falla, extrae el primer {...}; si no, texto plano."""
     raw = (raw or "").strip()
     try:
         return json.loads(raw)