app.py

import os
import torch
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr

# ——— CONFIG ———
REPO_ID = "CodCodingCode/llama-3.1-8b-clinical"
SUBFOLDER = "checkpoint-45000"
HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN")
if not HF_TOKEN:
    raise RuntimeError("Missing HUGGINGFACE_HUB_TOKEN in env")

# ——— 1) Download the full repo ———
local_cache = snapshot_download(
    repo_id=REPO_ID,
    token=HF_TOKEN,
)
print("[DEBUG] snapshot_download → local_cache:", local_cache)
import pathlib

print(
    "[DEBUG] MODEL root contents:",
    list(pathlib.Path(local_cache).glob(f"{SUBFOLDER}/*")),
)

# ——— 2) Repo root contains tokenizer.json; model shards live in the checkpoint subfolder ———
MODEL_DIR = local_cache
MODEL_SUBFOLDER = SUBFOLDER
print("[DEBUG] MODEL_DIR:", MODEL_DIR)
print("[DEBUG] MODEL_DIR files:", os.listdir(MODEL_DIR))
print("[DEBUG] Checkpoint files:", os.listdir(os.path.join(MODEL_DIR, MODEL_SUBFOLDER)))

# ——— 3) Load tokenizer & model from disk ———
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_DIR,
    use_fast=True,
)
print("[DEBUG] Loaded fast tokenizer object:", tokenizer, "type:", type(tokenizer))
# Confirm tokenizer files are present
import os

print("[DEBUG] Files in MODEL_DIR for tokenizer:", os.listdir(MODEL_DIR))
# Inspect tokenizer's initialization arguments
try:
    print("[DEBUG] Tokenizer init_kwargs:", tokenizer.init_kwargs)
except AttributeError:
    print("[DEBUG] No init_kwargs attribute on tokenizer.")

model = AutoModelForCausalLM.from_pretrained(
    MODEL_DIR,
    subfolder=MODEL_SUBFOLDER,
    device_map="auto",
    torch_dtype=torch.float16,
)
model.eval()
print(
    "[DEBUG] Loaded model object:",
    model.__class__.__name__,
    "device:",
    next(model.parameters()).device,
)


# === Role Agent with instruction/input/output format ===
class RoleAgent:
    def __init__(self, role_instruction, tokenizer, model):
        self.tokenizer = tokenizer
        self.model = model
        self.role_instruction = role_instruction

    def act(self, input_text):
        prompt = (
            f"instruction: {self.role_instruction}\n"
            f"input: {input_text}\n"
            f"output:"
        )
        encoding = self.tokenizer(prompt, return_tensors="pt")
        inputs = {k: v.to(self.model.device) for k, v in encoding.items()}

        outputs = self.model.generate(
            **inputs,
            max_new_tokens=128,
            do_sample=True,
            temperature=0.7,
            pad_token_id=self.tokenizer.eos_token_id,
        )
        response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)

        # Extract only the new generated content after the prompt
        prompt_length = len(prompt)
        if len(response) > prompt_length:
            generated_text = response[prompt_length:].strip()
        else:
            generated_text = response.strip()

        # Clean up the response - remove any repeated instruction/input/output patterns
        lines = generated_text.split("\n")
        clean_lines = []

        for line in lines:
            line = line.strip()
            # Skip lines that look like instruction formatting
            if (
                line.startswith("instruction:")
                or line.startswith("input:")
                or line.startswith("output:")
                or line == ""
            ):
                continue
            clean_lines.append(line)

        # Join the clean lines and take the first substantial response
        if clean_lines:
            answer = clean_lines[0]
            # If there are multiple clean lines, take the first one that's substantial
            for line in clean_lines:
                if len(line) > 20:  # Arbitrary threshold for substantial content
                    answer = line
                    break
        else:
            # Fallback: try to extract after "output:" if present
            if "output:" in generated_text.lower():
                parts = generated_text.lower().split("output:")
                if len(parts) > 1:
                    answer = parts[-1].strip()
                else:
                    answer = generated_text
            else:
                answer = generated_text

        # Additional cleanup - remove any remaining instruction artifacts
        answer = (
            answer.replace("instruction:", "")
            .replace("input:", "")
            .replace("output:", "")
            .strip()
        )

        # If answer is still messy, try to extract the actual medical content
        if "patient" in answer.lower() and len(answer) > 100:
            # Look for sentences that contain medical information
            sentences = answer.split(".")
            medical_sentences = []
            for sentence in sentences:
                sentence = sentence.strip()
                if len(sentence) > 10 and any(
                    word in sentence.lower()
                    for word in [
                        "patient",
                        "pain",
                        "symptom",
                        "diagnosis",
                        "treatment",
                        "knee",
                        "reports",
                        "experiencing",
                    ]
                ):
                    medical_sentences.append(sentence)

            if medical_sentences:
                answer = ". ".join(
                    medical_sentences[:2]
                )  # Take first 2 medical sentences
                if not answer.endswith("."):
                    answer += "."

        print(
            f"[CLEANED RESPONSE] Original length: {len(response)}, Cleaned: '{answer}'"
        )

        thinking = ""  # For now, we'll focus on getting clean answers
        return {"thinking": thinking, "output": answer}


# === Agents ===
summarizer = RoleAgent(
    role_instruction="You are a clinical summarizer trained to extract structured vignettes from doctor–patient dialogues.",
    tokenizer=tokenizer,
    model=model,
)
diagnoser = RoleAgent(
    role_instruction="You are a board-certified diagnostician that diagnoses patients.",
    tokenizer=tokenizer,
    model=model,
)
questioner = RoleAgent(
    role_instruction="You are a physician asking questions to diagnose a patient.",
    tokenizer=tokenizer,
    model=model,
)
treatment_agent = RoleAgent(
    role_instruction="You are a board-certified clinician. Based on the diagnosis and patient vignette provided below, suggest a concise treatment plan that could realistically be initiated by a primary care physician or psychiatrist.",
    tokenizer=tokenizer,
    model=model,
)

"""[DEBUG] prompt: Instruction: You are a clinical summarizer trained to extract structured vignettes from doctor–patient dialogues.
Input: Doctor: What brings you in today?
Patient: I am a male. I am 15. My knee hurts. What may be the issue with my knee?

Previous Vignette: 
Output:
Instruction: You are a clinical summarizer trained to extract structured vignettes from doctor–patient dialogues.
Input: Doctor: What brings you in today?
Patient: I am a male. I am 15. My knee hurts. What may be the issue with my knee?

Previous Vignette: 
Output: The patient is a 15-year-old male presenting with knee pain."""


# === Inference State ===
conversation_history = []
summary = ""
diagnosis = ""


# === Gradio Inference ===
def simulate_interaction(user_input, conversation_history=None):
    """Single turn interaction - no iterations, uses accumulated history"""
    if conversation_history is None:
        history = [f"Doctor: What brings you in today?", f"Patient: {user_input}"]
    else:
        history = conversation_history.copy()
        history.append(f"Patient: {user_input}")

    # Summarize the full conversation history
    sum_in = "\n".join(history)
    sum_out = summarizer.act(sum_in)
    summary = sum_out["output"]

    # Diagnose based on summary
    diag_out = diagnoser.act(summary)
    diagnosis = diag_out["output"]

    # Generate next question based on current understanding
    q_in = f"Vignette: {summary}\nCurrent Estimated Diagnosis: {diagnosis}"
    q_out = questioner.act(q_in)

    # Add doctor's response to history
    history.append(f"Doctor: {q_out['output']}")

    # Generate treatment plan (but don't end conversation)
    treatment_out = treatment_agent.act(f"Diagnosis: {diagnosis}\nVignette: {summary}")

    return {
        "summary": sum_out,
        "diagnosis": diag_out,
        "question": q_out,
        "treatment": treatment_out,
        "conversation": history,  # Return full history list
    }


# === Gradio UI ===
def ui_fn(user_input):
    """Non-stateful version for testing"""
    res = simulate_interaction(user_input)
    return f"""📋 Vignette Summary:
💭 THINKING: {res['summary']['thinking']}
📝 SUMMARY: {res['summary']['output']}

🩺 Diagnosis:
💭 THINKING: {res['diagnosis']['thinking']}
🔍 DIAGNOSIS: {res['diagnosis']['output']}

❓ Follow-up Question:
💭 THINKING: {res['question']['thinking']}
👨‍⚕️ DOCTOR: {res['question']['output']}

💊 Treatment Plan:
💭 THINKING: {res['treatment']['thinking']}
📋 TREATMENT: {res['treatment']['output']}

💬 Full Conversation:
{chr(10).join(res['conversation'])}
"""


# === Stateful Gradio UI ===
def stateful_ui_fn(user_input, history):
    """Proper stateful conversation handler"""
    # Initialize history if first interaction
    if history is None:
        history = []

    # Run one turn of interaction with accumulated history
    res = simulate_interaction(user_input, history)

    # Get the updated conversation history
    updated_history = res["conversation"]

    # Format the display output
    display_output = f"""💬 Conversation:
{chr(10).join(updated_history)}

📋 Current Assessment:
🔍 Diagnosis: {res['diagnosis']['output']}
💊 Treatment Plan: {res['treatment']['output']}
"""

    # Return display text and updated history for next turn
    return display_output, updated_history


def chat_interface(user_input, history):
    """Alternative chat-style interface"""
    if history is None:
        history = []

    # Run interaction
    res = simulate_interaction(user_input, history)
    updated_history = res["conversation"]

    # Return just the doctor's latest response and updated history
    doctor_response = res["question"]["output"]

    return doctor_response, updated_history


# Create two different interfaces
demo_stateful = gr.Interface(
    fn=stateful_ui_fn,
    inputs=[
        gr.Textbox(
            label="Patient Response",
            placeholder="Describe your symptoms or answer the doctor's question...",
        ),
        gr.State(),  # holds the conversation history
    ],
    outputs=[
        gr.Textbox(label="Medical Consultation", lines=15),
        gr.State(),  # returns the updated history
    ],
    title="🧠 AI Doctor - Full Medical Consultation",
    description="Have a conversation with an AI doctor. Each response builds on the previous conversation.",
)

demo_chat = gr.Interface(
    fn=chat_interface,
    inputs=[
        gr.Textbox(label="Your Response", placeholder="Tell me about your symptoms..."),
        gr.State(),
    ],
    outputs=[
        gr.Textbox(label="Doctor", lines=5),
        gr.State(),
    ],
    title="🩺 AI Doctor Chat",
    description="Simple chat interface with the AI doctor.",
)

if __name__ == "__main__":
    # Launch the stateful version by default
    demo_stateful.launch(share=True)
    # Uncomment the line below to use the chat version instead:
    # demo_chat.launch(share=True)