new model

app.py

@@ -1,9 +1,9 @@
 """
 =========================================================
- 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)
+ app.py — Green Greta (Gradio + TF/Keras 3 + LangChain v0.2)
+ - Image model: load TF SavedModel via keras.layers.TFSMLayer (Keras 3 safe)
+ - LLM: local transformers pipeline (no HF API token required)
+ - LangChain v0.2 imports (text_splitters/core/community)
  - Robust JSON parsing for schema-shaped output
  - EfficientNet input size fix (224x224)
  - Gradio binds to 0.0.0.0:7860 (Docker-friendly)
@@ -33,17 +33,20 @@ from fake_useragent import UserAgent
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.output_parsers import PydanticOutputParser
-from langchain_community.embeddings import HuggingFaceEmbeddings
 from langchain_community.document_loaders import WebBaseLoader
-from langchain_community.llms import HuggingFaceEndpoint
 from langchain_community.vectorstores import Chroma
 from langchain.chains import ConversationalRetrievalChain
 from langchain.memory import ConversationBufferMemory
 
-# Pydantic (for typed schema in prompt)
+# Embeddings (community version works; you can switch to langchain-huggingface later)
+from langchain_community.embeddings import HuggingFaceEmbeddings
+# If you prefer to silence deprecation warnings in the future:
+# from langchain_huggingface import HuggingFaceEmbeddings  # pip install -U langchain-huggingface
+
+# Pydantic for schema in prompt
 from pydantic.v1 import BaseModel, Field
 
-# Hugging Face Hub helpers
+# Hugging Face Hub helper for SavedModel
 from huggingface_hub import snapshot_download
 
 # Local theming + URLs list
@@ -77,18 +80,17 @@ 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.
-    Works with a TFSMLayer (SavedModel) that returns a dict of tensors.
+    Works with a TFSMLayer (SavedModel) that may return a dict of tensors.
     """
     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)  # [1, 224, 224, 3]
+    x = tf.keras.preprocessing.image.img_to_array(img)
+    x = tf.keras.applications.efficientnet.preprocess_input(x)
+    x = tf.expand_dims(x, 0)  # [1, 224, 224, 3]
 
-    # TFSMLayer returns a dict for SavedModel; select the first output
-    outputs = model1(image_array)
+    outputs = model1(x)
     if isinstance(outputs, dict) and outputs:
-        first_key = next(iter(outputs.keys()))
-        preds = outputs[first_key]
+        key = next(iter(outputs))
+        preds = outputs[key]
     else:
         preds = outputs
 
@@ -188,33 +190,26 @@ qa_prompt = ChatPromptTemplate.from_template(
 
 
 # =============================
-# 4) LLM (router-friendly HF)
+# 4) LLM (token-free local model)
 # =============================
-DEFAULT_REPO = os.environ.get("HF_REPO_ID", "HuggingFaceH4/zephyr-7b-beta")
-
-llm = HuggingFaceEndpoint(
-    repo_id=DEFAULT_REPO,
-    task="text-generation",
-    max_new_tokens=1024,
-    temperature=0.2,
-    top_k=50,
-    repetition_penalty=1.05,
-    do_sample=True,
-    # Requires env: HUGGINGFACEHUB_API_TOKEN=hf_xxx
+# Avoids HF Endpoint auth + deprecated .post path. Good defaults for CPU.
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+from langchain_community.llms import HuggingFacePipeline
+
+LOCAL_MODEL_ID = os.environ.get("LOCAL_LLM", "google/flan-t5-base")
+
+tok = AutoTokenizer.from_pretrained(LOCAL_MODEL_ID)
+mdl = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_ID)
+
+gen = pipeline(
+    task="text2text-generation",
+    model=mdl,
+    tokenizer=tok,
+    max_new_tokens=512,
+    do_sample=False,
 )
 
-# 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,
-#         task="text-generation",
-#         max_new_tokens=1024,
-#         temperature=0.2,
-#         top_k=50,
-#         repetition_penalty=1.05,
-#         do_sample=True,
-#     )
+llm = HuggingFacePipeline(pipeline=gen)
 
 
 # ===========================================