new model

app.py

@@ -1,10 +1,9 @@
 """
 =========================================================
- app.py — Green Greta (Gradio + TF/Keras 3 + LangChain v0.2)
+ app.py — Green Greta (Gradio + TF/Keras 3 + Local HF + LangChain v0.2)
 =========================================================
 """
 
-# ========== Imports ==========
 import os
 import json
 import shutil
@@ -17,16 +16,16 @@ from PIL import Image
 import tenacity
 from fake_useragent import UserAgent
 
-# LangChain v0.2 family
-from langchain_text_splitters import RecursiveCharacterTextSplitter, TokenTextSplitter
+# --- LangChain v0.2 family ---
+from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.output_parsers import PydanticOutputParser
-from langchain_community.document_loaders import WebBaseLoader
-from langchain_community.vectorstores import Chroma
 from langchain.chains import ConversationalRetrievalChain
 from langchain.memory import ConversationBufferMemory
+from langchain_community.document_loaders import WebBaseLoader
+from langchain_community.vectorstores import Chroma
 
-# Embeddings (use community; switch to langchain-huggingface later if desired)
+# Embeddings (prefer langchain-huggingface if installed; fallback to community)
 try:
     from langchain_huggingface import HuggingFaceEmbeddings  # pip install -U langchain-huggingface
 except ImportError:
@@ -36,45 +35,42 @@ except ImportError:
 from langchain.retrievers import ContextualCompressionRetriever
 from langchain.retrievers.document_compressors import DocumentCompressorPipeline
 
-# Pydantic schema
 from pydantic.v1 import BaseModel, Field
 
-# HF Hub for SavedModel download
+# HF Hub for downloading the SavedModel once
 from huggingface_hub import snapshot_download
 
-# Transformers local pipeline (no token needed)
+# Local transformers pipeline (no API token required)
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
 from langchain_community.llms import HuggingFacePipeline
 
-# Local theme + URLs
+# Theming + URL list
 import theme
 from url_list import URLS
 
-
-# ========== Theme ==========
 theme = theme.Theme()
 
 
 # =========================================================
-# 1) IMAGE CLASSIFICATION (Keras 3-compatible SavedModel)
+# 1) IMAGE CLASSIFICATION — Keras 3-safe SavedModel loading
 # =========================================================
 MODEL_REPO = "rocioadlc/efficientnetB0_trash"
-MODEL_SERVING_SIGNATURE = "serving_default"  # adjust if your model uses a different signature
+MODEL_SERVING_SIGNATURE = "serving_default"  # adjust if the model exposes a different endpoint
 
-# Download the SavedModel once and wrap with Keras TFSMLayer
+# Download the model snapshot and wrap it via TFSMLayer (Keras 3 compatible)
 model_dir = snapshot_download(MODEL_REPO)
-model1 = keras.layers.TFSMLayer(model_dir, call_endpoint=MODEL_SERVING_SIGNATURE)
+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 224x224 EfficientNet input and run inference."""
+    """Preprocess to EfficientNetB0 input (224x224) and run inference."""
     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)
-    x = tf.expand_dims(x, 0)  # batch
+    x = tf.expand_dims(x, 0)
 
-    outputs = model1(x)
+    outputs = image_model(x)
     if isinstance(outputs, dict) and outputs:
         preds = outputs[next(iter(outputs))]
     else:
@@ -116,13 +112,13 @@ def safe_load_all_urls(urls):
 
 all_loaded_docs = safe_load_all_urls(URLS)
 
-# Smaller base chunks to help keep prompts short
-text_splitter = RecursiveCharacterTextSplitter(
+# Smaller base chunks so downstream compression has less work
+base_splitter = RecursiveCharacterTextSplitter(
     chunk_size=700,
     chunk_overlap=80,
     length_function=len,
 )
-docs = text_splitter.split_documents(all_loaded_docs)
+docs = base_splitter.split_documents(all_loaded_docs)
 
 # Embeddings
 embeddings = HuggingFaceEmbeddings(model_name="thenlper/gte-small")
@@ -138,20 +134,27 @@ vectordb = Chroma.from_documents(
 )
 
 # Base retriever
-base_retriever = vectordb.as_retriever(search_kwargs={"k": 2}, search_type="mmr")
-
-# Hard-cap tokens in retrieved docs (~200 tokens per slice)
-token_splitter = TokenTextSplitter(chunk_size=200, chunk_overlap=30)
-compressor = DocumentCompressorPipeline(transformers=[token_splitter])
+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.
+try:
+    from langchain_text_splitters import TokenTextSplitter
+    splitter_for_compression = TokenTextSplitter(chunk_size=200, chunk_overlap=30)  # needs `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)
+
+compressor = DocumentCompressorPipeline(transformers=[splitter_for_compression])
 compression_retriever = ContextualCompressionRetriever(
-    base_retriever=base_retriever,
+    base_retriever=retriever,
     base_compressor=compressor,
 )
 
 
 # ======================================
-# 3) PROMPT & SCHEMA OUTPUT PARSING
+# 3) PROMPT & Pydantic schema parsing
 # ======================================
 class FinalAnswer(BaseModel):
     question: str = Field(description="User question")
@@ -174,7 +177,7 @@ qa_prompt = ChatPromptTemplate.from_template(
 
 
 # =============================
-# 4) LLM (token-free local model)
+# 4) LLM — local, token-free
 # =============================
 LOCAL_MODEL_ID = os.environ.get("LOCAL_LLM", "google/flan-t5-base")
 tok = AutoTokenizer.from_pretrained(LOCAL_MODEL_ID)
@@ -185,7 +188,7 @@ gen = pipeline(
     model=mdl,
     tokenizer=tok,
     max_new_tokens=512,
-    do_sample=False,  # deterministic; better for JSON adherence
+    do_sample=False,  # deterministic; helps JSON adherence
 )
 llm = HuggingFacePipeline(pipeline=gen)
 
@@ -200,7 +203,7 @@ memory = ConversationBufferMemory(
 
 qa_chain = ConversationalRetrievalChain.from_llm(
     llm=llm,
-    retriever=compression_retriever,  # <= compressed retriever to avoid 512-token overflows
+    retriever=compression_retriever,
     memory=memory,
     verbose=True,
     combine_docs_chain_kwargs={"prompt": qa_prompt},
@@ -210,7 +213,7 @@ qa_chain = ConversationalRetrievalChain.from_llm(
 )
 
 def _safe_json_extract(raw: str, question: str) -> dict:
-    """Try strict JSON; otherwise extract first {...} block; fallback to plain text."""
+    """Try strict JSON; otherwise extract first {...}; fallback to plain text."""
     raw = (raw or "").strip()
     try:
         return json.loads(raw)
@@ -279,8 +282,4 @@ app = gr.TabbedInterface(
 )
 
 app.queue()
-app.launch(
-    server_name="0.0.0.0",
-    server_port=7860,
-    share=os.environ.get("GRADIO_SHARE", "false").lower() == "true",
-)
+app.launch()

requirements.txt

@@ -28,3 +28,5 @@ fastapi==0.115.0
 starlette==0.38.2
 pydantic==2.8.2
 pydantic-core==2.20.1
+
+tiktoken>=0.5.2