Add multi‐backend LLM support and audio‐driven medical agent pipeline

- Integrate Whisper ASR for speech‐to‐text symptom input
- Unify agent logic in `transcribe_and_respond()` using `get_llm_predictor()` (OpenAI, Mistral, or local pipeline)
- Enable environment flags `USE_LOCAL_GPU` and `USE_MISTRAL` to switch models dynamically
- Update Gradio `app.py` to launch audio/chat interface with MCP support

.gitignore

@@ -1,2 +1,4 @@
 venv
-.venv
+.venv
+__pycache__
+gpt2-medium

app.py

@@ -2,4 +2,4 @@
 from src.app import demo
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(server_name="0.0.0.0", server_port=7860, mcp_server=True)

requirements.txt

@@ -1,4 +1,7 @@
 gradio[full]
+gradio[mcp]
 llama-index==0.6.9
 openai==0.27.0
-transformers
+transformers
+torch
+accelerate

src/app.py

@@ -1,40 +1,74 @@
-# app.py
-import json
+import os
 import gradio as gr
+from transformers import pipeline
+from llama_index import SimpleDirectoryReader, GPTVectorStoreIndex
+from llama_index.llm_predictor import HuggingFaceLLMPredictor, LLMPredictor
 
-# Load the merged knowledge base
-with open("data/knowledge_base.json", encoding="utf-8") as f:
-    kb = json.load(f)
-
-symptom_to_icd    = kb["symptom_to_icd"]
-icd_to_description = kb["icd_to_description"]
-
-def map_symptoms(raw_input):
-    terms = [t.strip().lower() for t in raw_input.split(",") if t.strip()]
-    icd_counts = {}
-    for term in terms:
-        for code in symptom_to_icd.get(term, []):
-            icd_counts[code] = icd_counts.get(code, 0) + 1
-    if not icd_counts:
-        return {"diagnoses": [], "confidences": []}
-    total = sum(icd_counts.values())
-    # sort codes by frequency descending
-    sorted_items = sorted(icd_counts.items(), key=lambda x: x[1], reverse=True)
-    diagnoses = []
-    confidences = []
-    for code, count in sorted_items:
-        desc = icd_to_description.get(code, "Unknown")
-        diagnoses.append(f"{code}: {desc}")
-        confidences.append(round(count / total, 2))
-    return {"diagnoses": diagnoses, "confidences": confidences}
-
-# Use Blocks so that mcp_server=True is accepted
+# Optional OpenAI import remains for default predictor
+import openai
+
+# --- Whisper ASR setup ---
+asr = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-small",
+    device=0,
+    chunk_length_s=30,
+)
+
+# --- LlamaIndex utils import ---
+from utils.llama_index_utils import get_llm_predictor, build_index, query_symptoms
+
+# --- System prompt ---
+SYSTEM_PROMPT = """
+You are a medical assistant helping a user narrow down to the most likely ICD-10 code.
+At each turn, EITHER ask one focused clarifying question (e.g. “Is your cough dry or productive?”)
+or, if you have enough info, output a final JSON with fields:
+{"diagnoses":[…], "confidences":[…]}.
+"""
+
+
+def transcribe_and_respond(audio, history):
+    # 1) Transcribe audio → text
+    user_text = asr(audio)["text"]
+    history = history or []
+    history.append(("user", user_text))
+
+    # 2) Build unified prompt for LLM
+    messages = [("system", SYSTEM_PROMPT)] + history
+    prompt = "\n".join(f"{role.capitalize()}: {text}" for role, text in messages)
+    prompt += "\nAssistant:"
+
+    # 3) Select predictor (OpenAI or Mistral/local)
+    predictor = get_llm_predictor()
+    resp = predictor.predict(prompt)
+
+    # 4) If JSON-style output, treat as final
+    if resp.strip().startswith("{"):
+        result = query_symptoms(resp)
+        history.append(("assistant", f"Here is your diagnosis: {result}"))
+        return "", history
+
+    # 5) Otherwise, it's a follow-up question
+    history.append(("assistant", resp))
+    return "", history
+
+
+# --- Build Gradio app ---
 with gr.Blocks() as demo:
-    gr.Markdown("## Symptom to ICD‐10 Code Lookup")
-    inp = gr.Textbox(label="Enter symptoms (comma‐separated)")
-    out = gr.JSON(label="Result")
-    # Wire the submit event
-    inp.submit(fn=map_symptoms, inputs=inp, outputs=out)
+    gr.Markdown("## Symptom to ICD-10 Diagnoser (audio & chat)")
+    chatbot = gr.Chatbot(label="Conversation")
+    mic     = gr.Microphone(label="Describe your symptoms")
+    state   = gr.State([])
+
+    mic.submit(
+        fn=transcribe_and_respond,
+        inputs=[mic, state],
+        outputs=[mic, chatbot, state]
+    )
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, mcp_server=True)
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        mcp_server=True
+    )

utils/llama_index_utils.py

@@ -1,10 +1,15 @@
 import os
-
+import json
 from transformers import pipeline
 from llama_index import SimpleDirectoryReader, GPTVectorStoreIndex, LLMPredictor, OpenAI
 
 _index = None
 
+def query_symptoms_tool(prompt_json: str):
+    # parse “prompt_json” into Python dict and call your existing query_symptoms()
+    data = json.loads(prompt_json)
+    return query_symptoms(data["raw_input"])
+
 def get_llm_predictor():
     """
     Return an LLMPredictor configured for local GPU (transformers) if USE_LOCAL_GPU=1,
@@ -41,3 +46,4 @@ def query_symptoms(prompt: str, top_k: int = 5):
     predictor = get_llm_predictor()
     query_engine = idx.as_query_engine(similarity_top_k=top_k, llm_predictor=predictor)
     return query_engine.query(prompt)
+