Updated model.

app.py

@@ -1,52 +1,103 @@
 import gradio as gr
-import pandas as pd, faiss, torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+import pandas as pd
+import faiss
+import torch
+import numpy as np
+
 from sentence_transformers import SentenceTransformer
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-# —— Load data & embedding model ——
+# ===============================
+# Load Retrieval Components
+# ===============================
+print("Loading corpus and FAISS index...")
 df = pd.read_csv("retrieval_corpus.csv")
 index = faiss.read_index("faiss_index.bin")
+
+print("Loading embedding model...")
 embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
 
-# —— Quantized BioMedLM with CPU offload ——
-model_id = "stanford-crfm/BioMedLM"
-bnb_config = BitsAndBytesConfig(
-    load_in_8bit=True,
-    llm_int8_threshold=6.0,
-    llm_int8_enable_fp32_cpu_offload=True,
-)
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-tokenizer.pad_token = tokenizer.eos_token
-generation_model = AutoModelForCausalLM.from_pretrained(
+# ===============================
+# Load LLM on CPU
+# ===============================
+model_id = "PrunaAI/BioMistral-7B-bnb-8bit-smashed"
+tokenizer = AutoTokenizer.from_pretrained("BioMistral/BioMistral-7B")
+model = AutoModelForCausalLM.from_pretrained(
     model_id,
-    quantization_config=bnb_config,
-    device_map={"": "cpu"},
+    trust_remote_code=True,
+    device_map=None,  # CPU only
 )
+tokenizer.pad_token = tokenizer.eos_token
+
+# ===============================
+# RAG Pipeline
+# ===============================
+def get_top_k_chunks(query, k=5):
+    query_embedding = embedding_model.encode([query])
+    scores, indices = index.search(np.array(query_embedding).astype("float32"), k)
+    return df.iloc[indices[0]]["text"].tolist()
 
-def retrieve_top_k(q, k=5):
-    emb = embedding_model.encode([q]).astype("float32")
-    D,I = index.search(emb, k)
-    res = df.iloc[I[0]].copy(); res["score"]=D[0]; return res
-
-def build_prompt(q, docs):
-    ctx = "\n".join(f"- {d['text']}" for _,d in docs.iterrows())
-    return f"""[INST] <<SYS>>…[/INST]"""  # your existing template
-
-def generate_local_answer(prompt, max_new_tokens=512):
-    import time
-    device = torch.device("cpu")
-    start = time.time()
-    inputs = tokenizer(prompt, return_tensors="pt", padding=True).to(device)
-    out = generation_model.generate(
-        input_ids=inputs.input_ids,
-        attention_mask=inputs.attention_mask,
-        max_new_tokens=max_new_tokens,
-        do_sample=False,
-        num_beams=1,
+def build_prompt(query, chunks):
+    context = "\n".join(f"{i+1}. {chunk}" for i, chunk in enumerate(chunks))
+    prompt = (
+        "You are a clinical reasoning assistant. Based on the following medical information, "
+        "answer the query with a detailed explanation.\n\n"
+        f"Context:\n{context}\n\n"
+        f"Query: {query}\n"
+        "Answer:"
     )
-    print(f"Gen time: {time.time()-start:.2f}s")
-    return tokenizer.decode(out[0], skip_special_tokens=True)
+    return prompt
+
+def generate_diagnosis(query):
+    chunks = get_top_k_chunks(query)
+    prompt = build_prompt(query, chunks)
+
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=2048)
+    input_ids = inputs.input_ids.to("cpu")
+
+    with torch.no_grad():
+        output = model.generate(
+            input_ids=input_ids,
+            max_new_tokens=256,
+            do_sample=True,
+            top_k=50,
+            top_p=0.95,
+            temperature=0.7,
+            pad_token_id=tokenizer.eos_token_id
+        )
+
+    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
+    answer = generated_text.split("Answer:")[-1].strip()
+    return answer, "\n\n".join(chunks)
+
+# ===============================
+# Gradio UI
+# ===============================
+def run_interface():
+    with gr.Blocks(theme=gr.themes.Soft()) as demo:
+        gr.Markdown("## 🧠 Clinical Diagnosis Assistant (RAG)")
+        gr.Markdown("Enter a clinical query. The assistant retrieves relevant medical facts and generates a diagnostic explanation.")
+
+        with gr.Row():
+            query_input = gr.Textbox(label="Clinical Query", placeholder="e.g. 65-year-old male with shortness of breath...")
+            generate_btn = gr.Button("Generate Diagnosis")
+
+        with gr.Accordion("📄 Retrieved Context", open=False):
+            context_output = gr.Textbox(label="Top-5 Retrieved Chunks", lines=10, interactive=False)
+
+        answer_output = gr.Textbox(label="Generated Diagnosis", lines=8)
+
+        generate_btn.click(
+            fn=generate_diagnosis,
+            inputs=query_input,
+            outputs=[answer_output, context_output]
+        )
+
+    return demo
 
-iface = gr.Interface(fn=lambda q: generate_local_answer(build_prompt(q, retrieve_top_k(q))),
-                     inputs="text", outputs="text")
-iface.launch()
+# ===============================
+# Launch App
+# ===============================
+if __name__ == "__main__":
+    demo = run_interface()
+    demo.launch()