app.py

#!/usr/bin/env python3
import os
import json
import logging
import re
from typing import Dict, Any
from pathlib import Path
from unstructured.partition.pdf import partition_pdf
from flask import Flask, request, jsonify
from flask_cors import CORS
from dotenv import load_dotenv
from bloatectomy import bloatectomy
from werkzeug.utils import secure_filename
from langchain_groq import ChatGroq
from typing_extensions import TypedDict, NotRequired

# --- Logging ---
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger("patient-assistant")
ALLOWED_EXTENSIONS = {"pdf"}
# --- Load environment ---
load_dotenv()
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
if not GROQ_API_KEY:
    logger.error("GROQ_API_KEY not set in environment")
    exit(1)

# --- Flask app setup ---
BASE_DIR = Path(__file__).resolve().parent
REPORTS_ROOT = Path(os.getenv("REPORTS_ROOT", str(BASE_DIR / "reports")))
static_folder = BASE_DIR / "static"

app = Flask(__name__, static_folder=str(static_folder), static_url_path="/static")
CORS(app)

# Ensure the reports directory exists
os.makedirs(REPORTS_ROOT, exist_ok=True)

# --- LLM setup ---
llm = ChatGroq(
    model=os.getenv("LLM_MODEL", "meta-llama/llama-4-scout-17b-16e-instruct"),
    temperature=0.0,
    max_tokens=1024,
    api_key=GROQ_API_KEY,
)

def clean_notes_with_bloatectomy(text: str, style: str = "remov") -> str:
    """Helper function to clean up text using the bloatectomy library."""
    try:
        b = bloatectomy(text, style=style, output="html")
        tokens = getattr(b, "tokens", None)
        if not tokens:
            return text
        return "\n".join(tokens)
    except Exception:
        logger.exception("Bloatectomy cleaning failed; returning original text")
        return text

# --- Agent prompt instructions ---
# Important change: the assistant should NOT insist on PID at the start of conversation.
# It should only ask for patient ID when the user asks for something that requires accessing previous medical records.
# Also, avoid asking for additional identity details (name/DOB) purely for "verification" — accept PID as sufficient for record access in this flow.
PATIENT_ASSISTANT_PROMPT = """
You are a patient assistant helping to analyze medical records and reports. You should be helpful for general medical questions even when no patient ID (PID) is provided.

Behavior rules (follow these strictly):
- Do NOT ask for the patient ID at the start of every conversation. Only request the PID when the user's question explicitly requires accessing prior medical records (for example: "show my previous lab report", "what does my thyroid test from last month say", "what was my doctor's note for PID 12345", etc.).
- When you do ask for a PID, be concise and ask only for the PID (e.g., "Please provide the patient ID (PID) to retrieve previous records."). Do not request name/DOB/other verification unless the user explicitly asks for an extra verification step.
- If the user supplies a PID in their message (patterns like "pid 5678", "p5678", "patient id: 5678"), accept and use it — do not ask again.
- Avoid repeating unnecessary clarifying questions. If the user has already given the PID, proceed to use it. If you previously asked for the PID and the user didn't provide it, ask once more succinctly and then offer to help with general guidance without records.

STRICT OUTPUT FORMAT (JSON ONLY):
Return a single JSON object with the following keys:
- assistant_reply: string  // a natural language reply to the user (short, helpful, always present)
- patientDetails: object  // keys may include name, problem, pid (patient ID), city, contact (update if user shared info)
- conversationSummary: string (optional)  // short summary of conversation + relevant patient docs

Rules:
- ALWAYS include `assistant_reply` as a non-empty string.
- Do NOT produce any text outside the JSON object.
- Be concise in `assistant_reply`. If you need more details, ask a targeted follow-up question.
- Do not make up information that is not present in the provided medical reports or conversation history.
"""

# --- Helper utilities ---
PID_PATTERN = re.compile(r"(?:\bpid\b|\bpatient\s*id\b|\bp\b)\s*[:#\-]?\s*(p?\d+)", re.IGNORECASE)
DIGIT_PATTERN = re.compile(r"\b(p?\d{3,})\b")  # fallback: any 3+ digit cluster

RECORD_KEYWORDS = [
    "report", "lab", "result", "results", "previous", "history", "record", "records",
    "test", "tests", "scan", "imaging", "radiology", "thyroid", "tsh", "t3", "t4",
    "prescription", "doctor", "referral", "visit", "consultation",
]


def extract_pid_from_text(text: str) -> str | None:
    """Try to extract PID from a free-form text string.
    Accepts patterns like: 'pid 5678', 'p5678', 'patient id: 5678', or a bare number if clearly intended as an id.
    """
    if not text:
        return None
    m = PID_PATTERN.search(text)
    if m:
        return m.group(1).lstrip('pP')
    # fallback: look for explicit mention like 'pid p5678' with different casing
    # final fallback: any 3+ digit cluster but only if message also contains a record keyword
    if any(k in text.lower() for k in RECORD_KEYWORDS):
        m2 = DIGIT_PATTERN.search(text)
        if m2:
            return m2.group(1).lstrip('pP')
    return None


def needs_pid_for_query(text: str) -> bool:
    """Decide whether the user's message requires looking up prior records (thus needs a PID).
    Simple heuristic: if message contains any keyword from RECORD_KEYWORDS or explicit phrases asking for previous tests/records.
    """
    if not text:
        return False
    lower = text.lower()
    # direct phrases that clearly require historical records
    phrases = ["previous report", "previous lab", "my report", "my records", "past report", "last report", "previous test", "previous results"]
    if any(p in lower for p in phrases):
        return True
    # if message contains any record-related keyword, treat as needing PID
    if any(k in lower for k in RECORD_KEYWORDS):
        return True
    return False

# --- JSON extraction helper (unchanged) ---
def extract_json_from_llm_response(raw_response: str) -> dict:
    """Safely extracts a JSON object from a string that might contain extra text or markdown."""
    default = {
        "assistant_reply": "I'm sorry — I couldn't understand that. Could you please rephrase?",
        "patientDetails": {},
        "conversationSummary": "",
    }

    if not raw_response or not isinstance(raw_response, str):
        return default

    # Find the JSON object, ignoring any markdown code fences
    m = re.search(r"```(?:json)?\s*([\s\S]*?)\s*```", raw_response)
    json_string = m.group(1).strip() if m else raw_response

    # Find the first opening brace and the last closing brace
    first = json_string.find('{')
    last = json_string.rfind('}')
    if first == -1 or last == -1 or first >= last:
        try:
            return json.loads(json_string)
        except Exception:
            logger.warning("Could not locate JSON braces in LLM output. Falling back to default.")
            return default

    candidate = json_string[first:last+1]
    # Remove trailing commas that might cause parsing issues
    candidate = re.sub(r',\s*(?=[}\]])', '', candidate)

    try:
        parsed = json.loads(candidate)
    except Exception as e:
        logger.warning("Failed to parse JSON from LLM output: %s", e)
        return default

    # Basic validation of the parsed JSON
    if isinstance(parsed, dict) and "assistant_reply" in parsed and isinstance(parsed["assistant_reply"], str) and parsed["assistant_reply"].strip():
        parsed.setdefault("patientDetails", {})
        parsed.setdefault("conversationSummary", "")
        return parsed
    else:
        logger.warning("Parsed JSON missing 'assistant_reply' or invalid format. Returning default.")
        return default

# --- Flask routes ---
@app.route("/", methods=["GET"])
def serve_frontend():
    """Serves the frontend HTML file."""
    try:
        return app.send_static_file("frontend.html")
    except Exception:
        return "<h3>frontend.html not found in static/ — please add your frontend.html there.</h3>", 404

@app.route("/upload_report", methods=["POST"])
def upload_report():
    """Handles the upload of a new PDF report for a specific patient."""
    if 'report' not in request.files:
        return jsonify({"error": "No file part in the request"}), 400

    file = request.files['report']
    patient_id = request.form.get("patient_id")

    if file.filename == '' or not patient_id:
        return jsonify({"error": "No selected file or patient ID"}), 400

    if file:
        filename = secure_filename(file.filename)
        patient_folder = REPORTS_ROOT / f"{patient_id}"
        os.makedirs(patient_folder, exist_ok=True)
        file_path = patient_folder / filename
        file.save(file_path)
        return jsonify({"message": f"File '{filename}' uploaded successfully for patient ID '{patient_id}'."}), 200

@app.route("/chat", methods=["POST"])
def chat():
    """Handles the chat conversation with the assistant."""
    data = request.get_json(force=True)
    if not isinstance(data, dict):
        return jsonify({"error": "invalid request body"}), 400

    chat_history = data.get("chat_history") or []
    patient_state = data.get("patient_state") or {}
    patient_id = patient_state.get("patientDetails", {}).get("pid")

    # --- Prepare the state for the LLM ---
    state = patient_state.copy()
    state.setdefault("asked_for_pid", False)
    state.setdefault("conversationSummary", state.get("conversationSummary", ""))
    state["lastUserMessage"] = ""
    if chat_history:
        # Find the last user message
        for msg in reversed(chat_history):
            if msg.get("role") == "user" and msg.get("content"):
                state["lastUserMessage"] = msg["content"]
                break

    # Try to infer PID from the last message (user might include it inline)
    inferred_pid = extract_pid_from_text(state.get("lastUserMessage", "") or "")
    if not patient_id and inferred_pid:
        logger.info("Inferred PID from user message: %s", inferred_pid)
        state.setdefault("patientDetails", {})["pid"] = inferred_pid
        patient_id = inferred_pid

    combined_text_parts = []

    # Decide whether this query actually needs patient records
    wants_records = needs_pid_for_query(state.get("lastUserMessage", "") or "")

    # If a PID is known, load the patient reports.
    if patient_id:
        patient_folder = REPORTS_ROOT / f"{patient_id}"
        if patient_folder.exists() and patient_folder.is_dir():
            for fname in sorted(os.listdir(patient_folder)):
                file_path = patient_folder / fname
                page_text = ""
                if partition_pdf is not None and str(file_path).lower().endswith('.pdf'):
                    try:
                        elements = partition_pdf(filename=str(file_path))
                        page_text = "\n".join([el.text for el in elements if hasattr(el, 'text') and el.text])
                    except Exception:
                        logger.exception("Failed to parse PDF %s", file_path)
                else:
                    try:
                        page_text = file_path.read_text(encoding='utf-8', errors='ignore')
                    except Exception:
                        page_text = ""

                if page_text:
                    cleaned = clean_notes_with_bloatectomy(page_text, style="remov")
                    if cleaned:
                        combined_text_parts.append(cleaned)

    # Update the conversation summary with the parsed documents
    base_summary = state.get("conversationSummary", "") or ""
    docs_summary = "\n\n".join(combined_text_parts)
    if docs_summary:
        state["conversationSummary"] = (base_summary + "\n\n" + docs_summary).strip()
    else:
        state["conversationSummary"] = base_summary

    # Build the user prompt for the LLM. We explicitly tell the LLM whether a PID is available and whether the user's message requires records.
    if patient_id:
        action_hint = f"Use the patient ID {patient_id} to retrieve and summarize any relevant reports."
    else:
        if wants_records and not state.get("asked_for_pid", False):
            action_hint = "The user is asking for information that requires prior medical records. Ask succinctly for the Patient ID (PID) if needed."
            # mark that we asked for PID so we don't repeatedly ask
            state["asked_for_pid"] = True
        elif wants_records and state.get("asked_for_pid", False):
            action_hint = "The user previously was asked for a PID but has not supplied one. Ask once more concisely for the PID; otherwise offer to help with general guidance without accessing records."
        else:
            action_hint = "No PID provided and the user's request does not need prior records. Provide helpful, general medical guidance and offer to retrieve records if the user later supplies a PID."

    user_prompt = f"""
Current patientDetails: {json.dumps(state.get("patientDetails", {}))}
Current conversationSummary: {state.get("conversationSummary", "")[:4000]}
Last user message: {state.get("lastUserMessage", "")}

SYSTEM_HINT: {action_hint}

Return ONLY valid JSON with keys: assistant_reply, patientDetails, conversationSummary.
"""

    messages = [
        {"role": "system", "content": PATIENT_ASSISTANT_PROMPT},
        {"role": "user", "content": user_prompt}
    ]

    try:
        logger.info("Invoking LLM with prepared state and prompt...")
        llm_response = llm.invoke(messages)
        raw_response = ""
        if hasattr(llm_response, "content"):
            raw_response = llm_response.content
        else:
            raw_response = str(llm_response)

        logger.info(f"Raw LLM response: {raw_response}")
        parsed_result = extract_json_from_llm_response(raw_response)

    except Exception as e:
        logger.exception("LLM invocation failed")
        return jsonify({"error": "LLM invocation failed", "detail": str(e)}), 500

    updated_state = parsed_result or {}

    # Merge patientDetails back into state (but avoid overwriting asked_for_pid)
    state.setdefault("patientDetails", {}).update(updated_state.get("patientDetails", {}))
    state["conversationSummary"] = updated_state.get("conversationSummary", state.get("conversationSummary", ""))

    assistant_reply = updated_state.get("assistant_reply")
    if not assistant_reply or not isinstance(assistant_reply, str) or not assistant_reply.strip():
        # Fallback to a polite message if the LLM response is invalid or empty
        assistant_reply = "I'm here to help — could you tell me more about your symptoms or provide a Patient ID (PID) if you want me to fetch past reports?"

    # Return the new assistant reply and the updated state so the frontend can persist it.
    response_payload = {
        "assistant_reply": assistant_reply,
        "updated_state": state,
    }

    return jsonify(response_payload)

@app.route("/upload_reports", methods=["POST"])
def upload_reports():
    """
    Upload one or more files for a patient.

    Expects multipart/form-data with:
      - patient_id (form field)
      - files (one or multiple files; use the same field name 'files' for each file)
    """
    try:
        # patient id can be in form or args (for convenience)
        patient_id = request.form.get("patient_id") or request.args.get("patient_id")
        if not patient_id:
            return jsonify({"error": "patient_id form field required"}), 400

        # get uploaded files (support both files and files[] naming)
        uploaded_files = request.files.getlist("files")
        if not uploaded_files:
            # fallback: single file under name 'file'
            single = request.files.get("file")
            if single:
                uploaded_files = [single]

        if not uploaded_files:
            return jsonify({"error": "no files uploaded (use form field 'files')"}), 400

        # create patient folder under REPORTS_ROOT/<patient_id>
        patient_folder = REPORTS_ROOT / str(patient_id)
        patient_folder.mkdir(parents=True, exist_ok=True)

        saved = []
        skipped = []

        for file_storage in uploaded_files:
            orig_name = getattr(file_storage, "filename", "") or ""
            filename = secure_filename(orig_name)
            if not filename:
                skipped.append({"filename": orig_name, "reason": "invalid filename"})
                continue

            # extension check
            ext = filename.rsplit(".", 1)[1].lower() if "." in filename else ""
            if ext not in ALLOWED_EXTENSIONS:
                skipped.append({"filename": filename, "reason": f"extension '{ext}' not allowed"})
                continue

            # avoid overwriting: if collision, add numeric suffix
            dest = patient_folder / filename
            if dest.exists():
                base, dot, extension = filename.rpartition(".")
                # if no base (e.g. ".bashrc") fallback
                base = base or filename
                i = 1
                while True:
                    candidate = f"{base}__{i}.{extension}" if extension else f"{base}__{i}"
                    dest = patient_folder / candidate
                    if not dest.exists():
                        filename = candidate
                        break
                    i += 1

            try:
                file_storage.save(str(dest))
                saved.append(filename)
            except Exception as e:
                logger.exception("Failed to save uploaded file %s: %s", filename, e)
                skipped.append({"filename": filename, "reason": f"save failed: {e}"})

        return jsonify({
            "patient_id": str(patient_id),
            "saved": saved,
            "skipped": skipped,
            "patient_folder": str(patient_folder)
        }), 200

    except Exception as exc:
        logger.exception("Upload failed: %s", exc)
        return jsonify({"error": "upload failed", "detail": str(exc)}), 500


@app.route("/ping", methods=["GET"])
def ping():
    return jsonify({"status": "ok"})

if __name__ == "__main__":
    port = int(os.getenv("PORT", 7860))
    app.run(host="0.0.0.0", port=port, debug=True)