new model

app.py

@@ -1,28 +1,35 @@
 """
 =========================================================
- Fixed app.py — Green Greta (Gradio + HF + LangChain v0.2)
- Notes:
- - Uses HuggingFaceEndpoint with a public router model (Zephyr) by default.
- - Robust JSON parsing (no fragile string slicing).
- - EfficientNet input size fixed (224x224).
- - LangChain v0.2 import layout (core/community/text-splitters).
- - Safer memory for ConversationalRetrievalChain; better error handling.
- - Gradio binds to 0.0.0.0:7860 for Docker.
+ app.py — Green Greta (Gradio + HF + LangChain v0.2 + Keras 3)
+ - Keras 3: load SavedModel via keras.layers.TFSMLayer (not load_model)
+ - LLM: HuggingFaceEndpoint with router-friendly Zephyr by default
+ - LangChain v0.2 import layout (core/community/text-splitters)
+ - Robust JSON parsing for schema-shaped output
+ - EfficientNet input size fix (224x224)
+ - Gradio binds to 0.0.0.0:7860 (Docker-friendly)
 =========================================================
 """
 
+# =========================
+# Imports (grouped together)
+# =========================
 import os
 import json
 import shutil
 
+# UI / web
 import gradio as gr
+
+# TensorFlow / Keras / image
 import tensorflow as tf
+from tensorflow import keras
 from PIL import Image
 
-import tenacity  # for retrying failed requests
+# Networking / retry
+import tenacity
 from fake_useragent import UserAgent
 
-# LangChain (v0.2+ layout)
+# LangChain v0.2 family
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.output_parsers import PydanticOutputParser
@@ -33,43 +40,60 @@ from langchain_community.vectorstores import Chroma
 from langchain.chains import ConversationalRetrievalChain
 from langchain.memory import ConversationBufferMemory
 
+# Pydantic (for typed schema in prompt)
 from pydantic.v1 import BaseModel, Field
 
-# Theming
+# Hugging Face Hub helpers
+from huggingface_hub import snapshot_download
+
+# Local theming + URLs list
 import theme
+from url_list import URLS
+
+
+# =========================
+# Theme instance
+# =========================
 theme = theme.Theme()
 
-# Import URL list
-from url_list import URLS
 
-"""
-=========================================================
- 1) IMAGE CLASSIFICATION MODEL SETUP
-=========================================================
-"""
-from huggingface_hub import from_pretrained_keras
+# =========================================================
+# 1) IMAGE CLASSIFICATION MODEL SETUP (Keras 3-compatible)
+# =========================================================
+# The HF repo is a TensorFlow SavedModel; with Keras 3 we must use TFSMLayer.
+MODEL_REPO = "rocioadlc/efficientnetB0_trash"
+MODEL_SERVING_SIGNATURE = "serving_default"  # adjust if your repo uses another signature
 
-# Load a Keras model from HuggingFace Hub
-model1 = from_pretrained_keras("rocioadlc/efficientnetB0_trash")
+# Download SavedModel locally
+model_dir = snapshot_download(MODEL_REPO)
 
-# Define class labels for the trash classification
+# Wrap SavedModel as a Keras layer
+model1 = keras.layers.TFSMLayer(model_dir, call_endpoint=MODEL_SERVING_SIGNATURE)
+
+# Class labels
 class_labels = ["cardboard", "glass", "metal", "paper", "plastic", "trash"]
 
 
 def predict_image(input_image: Image.Image):
     """
     Resize the user-uploaded image and preprocess it for EfficientNetB0.
-    Returns a dict of class probabilities.
+    Works with a TFSMLayer (SavedModel) that returns a dict of tensors.
     """
-    # Correct size for EfficientNetB0 is 224x224
-    image_array = tf.keras.preprocessing.image.img_to_array(
-        input_image.resize((224, 224))
-    )
+    img = input_image.convert("RGB").resize((224, 224))  # EfficientNetB0 expects 224x224
+    image_array = tf.keras.preprocessing.image.img_to_array(img)
     image_array = tf.keras.applications.efficientnet.preprocess_input(image_array)
-    image_array = tf.expand_dims(image_array, 0)  # batch dim
-
-    predictions = model1.predict(image_array)
-    probs = predictions[0].tolist()
+    image_array = tf.expand_dims(image_array, 0)  # [1, 224, 224, 3]
+
+    # TFSMLayer returns a dict for SavedModel; select the first output
+    outputs = model1(image_array)
+    if isinstance(outputs, dict) and outputs:
+        first_key = next(iter(outputs.keys()))
+        preds = outputs[first_key]
+    else:
+        preds = outputs
+
+    preds_np = preds.numpy() if hasattr(preds, "numpy") else preds
+    probs = preds_np[0].tolist()
     return {label: float(probs[i]) for i, label in enumerate(class_labels)}
 
 
@@ -81,12 +105,10 @@ image_gradio_app = gr.Interface(
     theme=theme,
 )
 
-"""
-=========================================================
- 2) KNOWLEDGE LOADING (RAG)
-=========================================================
-"""
-# 2.1) Define user agent to avoid blocking, etc.
+
+# ============================================
+# 2) KNOWLEDGE LOADING (RAG: loader + splitter)
+# ============================================
 user_agent = UserAgent().random
 header_template = {"User-Agent": user_agent}
 
@@ -104,7 +126,6 @@ def safe_load_all_urls(urls):
             docs = load_url(link)
             all_docs.extend(docs)
         except Exception as e:
-            # If load_url fails after all retries, skip that URL
             print(f"Skipping URL due to error: {link}\nError: {e}\n")
     return all_docs
 
@@ -119,9 +140,10 @@ text_splitter = RecursiveCharacterTextSplitter(
 
 docs = text_splitter.split_documents(all_loaded_docs)
 
-# Small + high-quality general embedding
+# Embeddings
 embeddings = HuggingFaceEmbeddings(model_name="thenlper/gte-small")
 
+# Vector store (Chroma)
 persist_directory = "docs/chroma/"
 shutil.rmtree(persist_directory, ignore_errors=True)
 
@@ -133,12 +155,10 @@ vectordb = Chroma.from_documents(
 
 retriever = vectordb.as_retriever(search_kwargs={"k": 3}, search_type="mmr")
 
-"""
-=========================================================
- 3) PROMPT & PARSER
-=========================================================
-"""
 
+# ======================================
+# 3) PROMPT & SCHEMA OUTPUT PARSING
+# ======================================
 class FinalAnswer(BaseModel):
     question: str = Field(description="User question")
     answer: str = Field(description="Direct answer")
@@ -166,16 +186,10 @@ qa_prompt = ChatPromptTemplate.from_template(
     partial_variables={"format_instructions": parser.get_format_instructions()},
 )
 
-"""
-=========================================================
- 4) LLM SETUP (no router 404s)
-=========================================================
-"""
-# IMPORTANT:
-# The old route "mistralai/Mixtral-8x7B-Instruct-v0.1" often 404s on the public HF router.
-# Use a router-available model OR your own paid Inference Endpoint.
 
-# ---- Option A (DEFAULT): public, free router model that works out-of-the-box
+# =============================
+# 4) LLM (router-friendly HF)
+# =============================
 DEFAULT_REPO = os.environ.get("HF_REPO_ID", "HuggingFaceH4/zephyr-7b-beta")
 
 llm = HuggingFaceEndpoint(
@@ -186,11 +200,11 @@ llm = HuggingFaceEndpoint(
     top_k=50,
     repetition_penalty=1.05,
     do_sample=True,
-    # Set env: HUGGINGFACEHUB_API_TOKEN=hf_xxx
+    # Requires env: HUGGINGFACEHUB_API_TOKEN=hf_xxx
 )
 
-# ---- Option B (MIXTRAL): your paid Inference Endpoint
-# MIXTRAL_ENDPOINT_URL = os.environ.get("HF_ENDPOINT_URL")  # e.g. https://xyz.aws.endpoints.huggingface.cloud
+# If you deploy a paid Inference Endpoint (e.g., for Mixtral), use:
+# MIXTRAL_ENDPOINT_URL = os.environ.get("HF_ENDPOINT_URL")
 # if MIXTRAL_ENDPOINT_URL:
 #     llm = HuggingFaceEndpoint(
 #         endpoint_url=MIXTRAL_ENDPOINT_URL,
@@ -202,12 +216,10 @@ llm = HuggingFaceEndpoint(
 #         do_sample=True,
 #     )
 
-"""
-=========================================================
- 5) CHAIN (with safer memory + error handling)
-=========================================================
-"""
 
+# ===========================================
+# 5) Chain (memory + robust JSON extraction)
+# ===========================================
 memory = ConversationBufferMemory(
     memory_key="chat_history",
     return_messages=True,
@@ -219,7 +231,7 @@ qa_chain = ConversationalRetrievalChain.from_llm(
     memory=memory,
     verbose=True,
     combine_docs_chain_kwargs={"prompt": qa_prompt},
-    get_chat_history=lambda h: h,  # memory already returns messages
+    get_chat_history=lambda h: h,
     rephrase_question=False,
     output_key="output",
 )
@@ -227,18 +239,18 @@ qa_chain = ConversationalRetrievalChain.from_llm(
 
 def chat_interface(question, history):
     """
-    Processes the user's question through the qa_chain,
-    and robustly parses the JSON output per schema.
+    Run the QA chain and return the 'answer' field from a JSON payload.
+    Falls back safely if the LLM returns non-JSON text.
     """
     try:
         result = qa_chain.invoke({"question": question})
         raw = result.get("output", "").strip()
 
-        # Try strict JSON first
+        # Strict JSON first
         try:
             payload = json.loads(raw)
         except json.JSONDecodeError:
-            # If the model returned extra text around JSON, try to extract the first JSON object
+            # Try extracting first {...} block
             start = raw.find("{")
             end = raw.rfind("}")
             if start != -1 and end != -1 and end > start:
@@ -249,12 +261,10 @@ def chat_interface(question, history):
             else:
                 payload = {"question": question, "answer": raw}
 
-        # Enforce schema
-        answer_out = payload.get("answer", raw)
-        return answer_out
+        # Return the schema field
+        return payload.get("answer", raw)
 
     except Exception as e:
-        # Fallback: return a friendly error + no crash
         return (
             "Lo siento, tuve un problema procesando tu pregunta. "
             "Intenta de nuevo en un momento o formula la consulta de otra manera.\n\n"
@@ -267,12 +277,10 @@ chatbot_gradio_app = gr.ChatInterface(
     title="<span style='color: rgb(243, 239, 224);'>Green Greta</span>",
 )
 
-"""
-=========================================================
- 6) BANNER / WELCOME TAB
-=========================================================
-"""
 
+# ============================
+# 6) Banner / Welcome content
+# ============================
 banner_tab_content = """
 <div style="background-color: #d3e3c3; text-align: center; padding: 20px; display: flex; flex-direction: column; align-items: center;">
     <img src="https://huggingface.co/spaces/ALVHB95/TFM_DataScience_APP/resolve/main/front_4.jpg" alt="Banner Image" style="width: 50%; max-width: 500px; margin: 0 auto;">
@@ -297,21 +305,18 @@ banner_tab_content = """
 
 banner_tab = gr.Markdown(banner_tab_content)
 
-"""
-=========================================================
- 7) GRADIO FINAL APP: TABS
-=========================================================
-"""
 
+# ============================
+# 7) Gradio app (tabs + run)
+# ============================
 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,
 )
 
-# Enable queue() for concurrency and launch the Gradio app
+# Concurrency queue + launch (Docker-friendly binding)
 app.queue()
-# Tip: set GRADIO_SHARE=true in env if you want a public link
 app.launch(
     server_name="0.0.0.0",
     server_port=7860,