Update app.py

app.py

@@ -1,14 +1,17 @@
 import os
 import gradio as gr
+import torch
 from llama_index.core import (
     SimpleDirectoryReader,
     VectorStoreIndex,
     StorageContext,
     Settings
 )
-from llama_index.embeddings.openai import OpenAIEmbedding
-from llama_index.core.node_parser import SentenceSplitter
+from llama_index.embeddings.huggingface import HuggingFaceEmbedding
 from llama_index.vector_stores.faiss import FaissVectorStore
+from llama_index.llms.huggingface import HuggingFaceLLM
+from llama_index.core.node_parser import SentenceSplitter
+from transformers import AutoTokenizer
 import faiss
 
 # ====== Configuration ======
@@ -16,11 +19,23 @@ PDF_DIR = "./data"
 INDEX_SAVE_PATH = "./saved_index"
 CHUNK_SIZE = 512
 CHUNK_OVERLAP = 50
+EMBED_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
+LLM_MODEL = "microsoft/Phi-3-mini-4k-instruct"  # 3.8B parameter model
+
+# ====== Initialize Local Models ======
+# Embedding model (runs offline)
+Settings.embed_model = HuggingFaceEmbedding(model_name=EMBED_MODEL)
 
-# ====== Initialize OpenAI Embeddings ======
-Settings.embed_model = OpenAIEmbedding(
-    model="text-embedding-3-small",  # Cheaper and faster than large
-    api_key=os.getenv("OPENAI_API_KEY")  # Set your key in environment
+# Local LLM with 4-bit quantization
+tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
+Settings.llm = HuggingFaceLLM(
+    model_name=LLM_MODEL,
+    tokenizer_name=LLM_MODEL,
+    device_map="auto",
+    model_kwargs={
+        "torch_dtype": torch.float16,
+        "trust_remote_code": True
+    }
 )
 
 # ====== Node Parser ======
@@ -37,32 +52,46 @@ if os.path.exists(INDEX_SAVE_PATH):
     index = VectorStoreIndex.load(storage_context=storage_context)
 else:
     # Create new index
+    if not os.path.exists(PDF_DIR):
+        raise FileNotFoundError(f"Add medical PDFs to {PDF_DIR} directory first")
+
     documents = SimpleDirectoryReader(PDF_DIR).load_data()
     nodes = parser.get_nodes_from_documents(documents)
     
     # Create FAISS index
-    dimension = 1536  # text-embedding-3-small dimension
+    dimension = 384  # Match MiniLM embedding size
     faiss_index = faiss.IndexFlatL2(dimension)
     vector_store = FaissVectorStore(faiss_index=faiss_index)
     
     storage_context = StorageContext.from_defaults(vector_store=vector_store)
     index = VectorStoreIndex(nodes, storage_context=storage_context)
     
-    # Save for future use
+    # Save for offline use
     index.storage_context.persist(persist_dir=INDEX_SAVE_PATH)
 
+# ====== Safety Layers ======
+def validate_response(response: str) -> str:
+    """Implements WHO protocol constraints"""
+    if len(response.split('\n')) > 6:
+        return "⚠️ Protocol too complex - must be <6 steps\n\n" + response
+    uncertainty_phrases = ["I think", "maybe", "not sure", "غير متأكد"]
+    if any(phrase in response for phrase in uncertainty_phrases):
+        return "⚠️ Consult supervisor - uncertain response\n\n" + response
+    return response
+
 # ====== Query Engine ======
 query_engine = index.as_query_engine()
 
 # ====== Gradio Interface ======
 def ask_question(query):
-    response = query_engine.query(query)
-    return str(response)
+    response = str(query_engine.query(query))
+    return validate_response(response)
 
-gr.Interface(
-    fn=ask_question,
-    inputs=gr.Textbox(lines=2, placeholder="Ask a medical question..."),
-    outputs="text",
-    title="🩺 Medical Knowledge Assistant",
-    description="Answers based on embedded documents (OpenAI embeddings + local FAISS index)"
-).launch()
+if __name__ == "__main__":
+    gr.Interface(
+        fn=ask_question,
+        inputs=gr.Textbox(lines=2, placeholder="Ask a medical question..."),
+        outputs="text",
+        title="🩺 Gaza Field Medic Assistant (Offline)",
+        description="WHO protocols • No internet required • Arabic/English"
+    ).launch(server_name="0.0.0.0")