app.py

import os, re, json, traceback, pathlib
from functools import lru_cache

import gradio as gr
import torch
import regex as re2  # pip install regex

from settings import SNAPSHOT_PATH, PERSIST_CONTENT
from audit_log import log_event, hash_summary
from privacy import redact_text

# ---------- Environment / cache (Spaces-safe, writable) ----------
HOME = pathlib.Path.home()  # /home/user on Spaces
HF_HOME = str(HOME / ".cache" / "huggingface")
HF_HUB_CACHE = str(HOME / ".cache" / "huggingface" / "hub")
HF_TRANSFORMERS = str(HOME / ".cache" / "huggingface" / "transformers")
ST_HOME = str(HOME / ".cache" / "sentence-transformers")
GRADIO_TMP = str(HOME / "app" / "gradio")  # you can switch to "/tmp/gradio" if preferred
GRADIO_CACHE = GRADIO_TMP

os.environ.setdefault("HF_HOME", HF_HOME)
os.environ.setdefault("HF_HUB_CACHE", HF_HUB_CACHE)
os.environ.setdefault("TRANSFORMERS_CACHE", HF_TRANSFORMERS)  # deprecated warning is harmless
os.environ.setdefault("SENTENCE_TRANSFORMERS_HOME", ST_HOME)
os.environ.setdefault("GRADIO_TEMP_DIR", GRADIO_TMP)
os.environ.setdefault("GRADIO_CACHE_DIR", GRADIO_CACHE)

# Disable experimental xet; prefer stable transfer
os.environ.setdefault("HF_HUB_ENABLE_XET", "0")
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")

for p in [HF_HOME, HF_HUB_CACHE, HF_TRANSFORMERS, ST_HOME, GRADIO_TMP, GRADIO_CACHE]:
    try:
        os.makedirs(p, exist_ok=True)
    except Exception:
        pass

# Optional Cohere
try:
    import cohere
    _HAS_COHERE = True
except Exception:
    _HAS_COHERE = False

from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login

from safety import safety_filter, refusal_reply
from retriever import init_retriever, retrieve_context
from decision_math import compute_operational_numbers
from prompt_templates import build_system_preamble
from upload_ingest import extract_text_from_files
from session_rag import SessionRAG
from mdsi_analysis import capacity_projection, cost_estimate, outcomes_summary

# ---------- Config ----------
MODEL_ID = os.getenv("MODEL_ID", "microsoft/Phi-3-mini-4k-instruct")  # local fallback
HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN") or os.getenv("HF_TOKEN")

COHERE_API_KEY = os.getenv("COHERE_API_KEY")
USE_HOSTED_COHERE = bool(COHERE_API_KEY and _HAS_COHERE)

MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "512"))

# ---------- System Master (two-phase, LLM-only behavior) ----------
SYSTEM_MASTER = """
SYSTEM ROLE (fixed, always on)
You are ClarityOps, a medical analytics system that interacts only via this chat.

Absolute rules:
- Use ONLY information provided in this conversation (scenario text + uploaded files).
- Never invent data. If something required is missing after clarifications, output the literal token: INSUFFICIENT_DATA.
- Always run in TWO PHASES:
  Phase 1: Ask up to 5 concise clarification questions, grouped by category (Prioritization, Capacity, Cost, Clinical, Recommendations). Then STOP and WAIT.
  Phase 2: After answers are provided, produce the final structured analysis exactly in the required format.

Core behavior:
- Read and synthesize any user-uploaded files (e.g., CSV/XLSX/PDF) relevant to the scenario.
- Prefer analytics/longitudinal recommendations (risk targeting, follow-up, clustering) over generic ops advice.
- Show all calculations explicitly for capacity and costs (e.g., “6 teams × 8 clients/day × 60 days = 2,880”).
- Use correct clinical units and plausible ranges.
- Include a brief “Provenance” section mapping each key output to scenario text, files, and/or clarified answers.

Medical guardrails (always apply):
- Units: BP in mmHg, A1c in %, BMI in kg/m², Total Cholesterol in mmol/L (or as provided), Percentages in %.
- Plausible ranges: A1c 3–20 %, SBP 60–250 mmHg, DBP 30–150 mmHg, BMI 10–70 kg/m², Total Chol 2–12 mmol/L.
- Privacy: avoid PHI; aggregate only; apply small-cell suppression where cohort < 10 (describe at a higher level).
- When data includes mixed or ambiguous indicators, ask to confirm preferred indicators (e.g., obesity/metabolic syndrome vs self-reported diabetes).

Formatting hard rules:
- Phase 1 output MUST include the header line: “Clarification Questions”
- Phase 2 output MUST include the header line: “Structured Analysis”
- Phase 2 MUST follow this exact section order:
  1. Prioritization
  2. Capacity
  3. Cost
  4. Clinical Benefits
  5. ClarityOps Top 3 Recommendations
  (Include a short Provenance block at the end.)
""".strip()

# ---------- Helpers ----------
def pick_dtype_and_map():
    if torch.cuda.is_available():
        return torch.float16, "auto"
    if torch.backends.mps.is_available():
        return torch.float16, {"": "mps"}
    return torch.float32, "cpu"

def is_identity_query(message, history):
    patterns = [
        r"\bwho\s+are\s+you\b", r"\bwhat\s+are\s+you\b", r"\bwhat\s+is\s+your\s+name\b",
        r"\bwho\s+is\s+this\b", r"\bidentify\s+yourself\b", r"\btell\s+me\s+about\s+yourself\b",
        r"\bdescribe\s+yourself\b", r"\band\s+you\s*\?\b", r"\byour\s+name\b",
        r"\bwho\s+am\s+i\s+chatting\s+with\b",
    ]
    def match(t): return any(re.search(p, (t or "").strip().lower()) for p in patterns)
    if match(message): return True
    if history:
        last_user = history[-1][0] if isinstance(history[-1], (list, tuple)) else None
        if match(last_user): return True
    return False

def _iter_user_assistant(history):
    for item in (history or []):
        if isinstance(item, (list, tuple)):
            u = item[0] if len(item) > 0 else ""
            a = item[1] if len(item) > 1 else ""
            yield u, a

def _sanitize_text(s: str) -> str:
    if not isinstance(s, str):
        return s
    return re2.sub(r'[\p{C}--[\n\t]]+', '', s)

def _history_to_prompt(message, history):
    parts = [f"System: {SYSTEM_MASTER}"]
    for u, a in _iter_user_assistant(history):
        if u: parts.append(f"User: {u}")
        if a: parts.append(f"Assistant: {a}")
    parts.append(f"User: {message}")
    parts.append("Assistant:")
    return "\n".join(parts)

# ---------- Cohere first ----------
def cohere_chat(message, history):
    if not USE_HOSTED_COHERE:
        return None
    try:
        client = cohere.Client(api_key=COHERE_API_KEY)
        prompt = _history_to_prompt(message, history)
        resp = client.chat(
            model="command-r7b-12-2024",
            message=prompt,
            temperature=0.3,
            max_tokens=MAX_NEW_TOKENS,
        )
        if hasattr(resp, "text") and resp.text: return resp.text.strip()
        if hasattr(resp, "reply") and resp.reply: return resp.reply.strip()
        if hasattr(resp, "generations") and resp.generations: return resp.generations[0].text.strip()
        return None
    except Exception:
        return None

# ---------- Local model (HF) ----------
@lru_cache(maxsize=1)
def load_local_model():
    if not HF_TOKEN:
        raise RuntimeError("HUGGINGFACE_HUB_TOKEN is not set.")
    login(token=HF_TOKEN, add_to_git_credential=False)
    dtype, device_map = pick_dtype_and_map()
    tok = AutoTokenizer.from_pretrained(
        MODEL_ID, token=HF_TOKEN, use_fast=True, model_max_length=8192,
        padding_side="left", trust_remote_code=True,
        cache_dir=os.environ.get("TRANSFORMERS_CACHE")
    )
    try:
        mdl = AutoModelForCausalLM.from_pretrained(
            MODEL_ID, token=HF_TOKEN, device_map=device_map,
            low_cpu_mem_usage=True, torch_dtype=dtype, trust_remote_code=True,
            cache_dir=os.environ.get("TRANSFORMERS_CACHE")
        )
    except Exception:
        mdl = AutoModelForCausalLM.from_pretrained(
            MODEL_ID, token=HF_TOKEN,
            low_cpu_mem_usage=True, torch_dtype=dtype, trust_remote_code=True,
            cache_dir=os.environ.get("TRANSFORMERS_CACHE")
        )
        mdl.to("cuda" if torch.cuda.is_available() else "cpu")
    if mdl.config.eos_token_id is None and tok.eos_token_id is not None:
        mdl.config.eos_token_id = tok.eos_token_id
    return mdl, tok

def build_inputs(tokenizer, message, history):
    msgs = [{"role": "system", "content": SYSTEM_MASTER}]
    for u, a in _iter_user_assistant(history):
        if u: msgs.append({"role": "user", "content": u})
        if a: msgs.append({"role": "assistant", "content": a})
    msgs.append({"role": "user", "content": message})
    return tokenizer.apply_chat_template(
        msgs, tokenize=True, add_generation_prompt=True, return_tensors="pt"
    )

def local_generate(model, tokenizer, input_ids, max_new_tokens=MAX_NEW_TOKENS):
    input_ids = input_ids.to(model.device)
    with torch.no_grad():
        out = model.generate(
            input_ids=input_ids, max_new_tokens=max_new_tokens,
            do_sample=True, temperature=0.3, top_p=0.9,
            repetition_penalty=1.15,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
    gen_only = out[0, input_ids.shape[-1]:]
    return tokenizer.decode(gen_only, skip_special_tokens=True).strip()

# ---------- Snapshot, retriever, RAG ----------
def _load_snapshot(path=SNAPSHOT_PATH):
    try:
        with open(path, "r", encoding="utf-8") as f:
            return json.load(f)
    except Exception:
        return {
            "timestamp": None, "beds_total": 400, "staffed_ratio": 1.0, "occupied_pct": 0.97,
            "ed_census": 62, "ed_admits_waiting": 19, "avg_ed_wait_hours": 8,
            "discharge_ready_today": 11, "discharge_barriers": {"allied_health": 7, "placement": 4},
            "rn_shortfall": {"med_ward_A": 1, "med_ward_B": 1},
            "forecast_admits_next_24h": {"respiratory": 14, "other": 9},
            "isolation_needs_waiting": {"contact": 3, "airborne": 1}, "telemetry_needed_waiting": 5
        }

init_retriever()
_session_rag = SessionRAG()

def _mdsi_block():
    base_capacity = capacity_projection(18, 48, 6)
    cons_capacity = capacity_projection(12, 48, 6)
    opt_capacity = capacity_projection(24, 48, 6)
    cost_1200 = cost_estimate(1200, 74.0, 75000.0)
    outcomes = outcomes_summary()
    return json.dumps({
        "capacity_projection": {"conservative": cons_capacity, "base": base_capacity, "optimistic": opt_capacity},
        "cost_for_1200": cost_1200,
        "outcomes_summary": outcomes
    }, indent=2)

# ---------- Core chat logic (two-phase) ----------
def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answers=False):
    """
    awaiting_answers:
      - False: Phase 1 -> generate clarification questions and WAIT
      - True:  Phase 2 -> consume clarifications and produce structured analysis
    """
    try:
        log_event("user_message", None, {"sizes": {"chars": len(user_msg or "")}})

        # Safety (input)
        safe_in, blocked_in, reason_in = safety_filter(user_msg, mode="input")
        if blocked_in:
            ans = refusal_reply(reason_in)
            return history + [(user_msg, ans)], awaiting_answers

        # Identity short-circuit
        if is_identity_query(safe_in, history):
            ans = "I am ClarityOps, your strategic decision making AI partner."
            return history + [(user_msg, ans)], awaiting_answers

        # Ingest uploads (text + artifacts like CSV headers)
        if uploaded_files_paths:
            ing = extract_text_from_files(uploaded_files_paths)
            chunks = ing.get("chunks", []) if isinstance(ing, dict) else (ing or [])
            artifacts = ing.get("artifacts", []) if isinstance(ing, dict) else []
            if chunks:
                _session_rag.add_docs(chunks)
            if artifacts:
                _session_rag.register_artifacts(artifacts)
            log_event("uploads_added", None, {"chunks": len(chunks), "artifacts": len(artifacts)})

        # Columns helper
        if re.search(r"\b(columns?|headers?)\b", (safe_in or "").lower()):
            cols = _session_rag.get_latest_csv_columns()
            if cols:
                return history + [(user_msg, "Here are the column names from your most recent CSV upload:\n\n- " + "\n- ".join(cols))], awaiting_answers

        # Session retrieval to enrich the system preamble
        session_snips = "\n---\n".join(_session_rag.retrieve(
            "diabetes screening Indigenous Métis mobile program cost throughput outcomes logistics",
            k=6
        ))

        snapshot = _load_snapshot()
        policy_context = retrieve_context(
            "mobile diabetes screening Indigenous community outreach cultural safety data governance outcomes"
        )
        computed = compute_operational_numbers(snapshot)

        user_lower = (safe_in or "").lower()
        mdsi_extra = _mdsi_block() if ("diabetes" in user_lower or "mdsi" in user_lower or "mobile screening" in user_lower) else ""

        scenario_block = safe_in if len((safe_in or "")) > 0 else ""
        system_preamble = build_system_preamble(
            snapshot=snapshot,
            policy_context=policy_context,
            computed_numbers=computed,
            scenario_text=scenario_block + (f"\n\nExecutive Pre-Computed Blocks:\n{mdsi_extra}" if mdsi_extra else ""),
            session_snips=session_snips
        )

        # Phase directive
        if not awaiting_answers:
            phase_directive = (
                "\n\n[INSTRUCTION TO MODEL]\n"
                "Produce **Phase 1** only: output a header 'Clarification Questions' and ask up to 5 concise, grouped questions "
                "(Prioritization, Capacity, Cost, Clinical, Recommendations). Then STOP and WAIT.\n"
            )
        else:
            phase_directive = (
                "\n\n[INSTRUCTION TO MODEL]\n"
                "Produce **Phase 2** only: output a header 'Structured Analysis' and follow the exact section order "
                "(Prioritization, Capacity, Cost, Clinical Benefits, ClarityOps Top 3 Recommendations). "
                "Use uploaded files + the user's latest answers as authoritative. Show calculations, units, and a brief Provenance.\n"
            )

        augmented_user = SYSTEM_MASTER + "\n\n" + system_preamble + "\n\nUser message:\n" + safe_in + phase_directive

        # Call LLM
        out = cohere_chat(augmented_user, history)
        if not out:
            model, tokenizer = load_local_model()
            inputs = build_inputs(tokenizer, augmented_user, history)
            out = local_generate(model, tokenizer, inputs, max_new_tokens=MAX_NEW_TOKENS)

        # Clean + sanitize
        if isinstance(out, str):
            for tag in ("Assistant:", "System:", "User:"):
                if out.startswith(tag):
                    out = out[len(tag):].strip()
        out = _sanitize_text(out)

        # Safety (output)
        safe_out, blocked_out, reason_out = safety_filter(out, mode="output")
        if blocked_out:
            safe_out = refusal_reply(reason_out)

        # Flip phase state based on headers
        new_awaiting = awaiting_answers
        low = safe_out.lower()
        if not awaiting_answers and "clarification questions" in low:
            new_awaiting = True
        elif awaiting_answers and "structured analysis" in low:
            new_awaiting = False

        log_event("assistant_reply", None, {
            **hash_summary("prompt", augmented_user if not PERSIST_CONTENT else ""),
            **hash_summary("reply", safe_out if not PERSIST_CONTENT else ""),
            "awaiting_next_phase": new_awaiting
        })

        return history + [(user_msg, safe_out)], new_awaiting

    except Exception as e:
        err = f"Error: {e}"
        try:
            traceback.print_exc()
        except Exception:
            pass
        return history + [(user_msg, err)], awaiting_answers

# ---------- Theme & CSS ----------
theme = gr.themes.Soft(primary_hue="teal", neutral_hue="slate", radius_size=gr.themes.sizes.radius_lg)
custom_css = """
:root { --brand-bg: #e6f7f8; --brand-accent: #0d9488; --brand-text: #0f172a; --brand-text-light: #ffffff; }
.gradio-container { background: var(--brand-bg); }
h1 { color: var(--brand-text); font-weight: 700; font-size: 28px !important; }
.chatbot header, .chatbot .label, .chatbot .label-wrap, .chatbot .top, .chatbot .header, .chatbot > .wrap > header { display: none !important; }
.message.user, .message.bot { background: var(--brand-accent) !important; color: var(--brand-text-light) !important; border-radius: 12px !important; padding: 8px 12px !important; }
textarea, input, .gr-input { border-radius: 12px !important; }

/* Centered handshake overlay */
#handshake-overlay {
  position: absolute;
  z-index: 50;
  top: 50%;
  left: 50%;
  transform: translate(-50%, -50%);
  background: rgba(13, 148, 136, 0.92);
  color: #fff;
  padding: 18px 22px;
  border-radius: 14px;
  font-size: 16px;
  max-width: 720px;
  text-align: center;
  box-shadow: 0 10px 24px rgba(0,0,0,0.2);
}
#handshake-overlay.hidden { display: none; }
#chat-container { position: relative; }
"""

# ---------- UI ----------
with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
    gr.Markdown("# ClarityOps Augmented Decision AI")

    with gr.Column(elem_id="chat-container"):
        chat = gr.Chatbot(label="", show_label=False, height=700)
        handshake = gr.HTML(
            value='<div id="handshake-overlay">ClarityOps loaded. Paste your scenario and attach files. I’ll ask up to 5 clarifications, then produce the structured analysis</div>'
        )

    with gr.Row():
        uploads = gr.Files(
            label="Upload docs/images (PDF, DOCX, CSV, PNG, JPG)",
            file_types=["file"], file_count="multiple", height=68
        )

    with gr.Row():
        msg = gr.Textbox(
            label="",
            show_label=False,
            placeholder="Paste your scenario here (attach files below). ClarityOps will ask clarifications first.",
            scale=10
        )
        send = gr.Button("Send", scale=1)
        clear = gr.Button("Clear chat", scale=1)

    state_history = gr.State(value=[])
    state_uploaded = gr.State(value=[])
    state_awaiting = gr.State(value=False)  # False -> Phase 1 next; True -> awaiting answers for Phase 2

    def _store_uploads(files, current):
        paths = []
        for f in (files or []):
            paths.append(getattr(f, "name", None) or f)
        return (current or []) + paths

    uploads.change(fn=_store_uploads, inputs=[uploads, state_uploaded], outputs=state_uploaded)

    def _on_send(user_msg, history, up_paths, awaiting):
        hide_overlay_js = gr.update(value='<div id="handshake-overlay" class="hidden"></div>')
        try:
            if not user_msg or not user_msg.strip():
                return history, "", history, awaiting, hide_overlay_js
            new_history, new_awaiting = clarityops_reply(
                user_msg.strip(), history or [], None, up_paths or [], awaiting_answers=awaiting
            )
            return new_history, "", new_history, new_awaiting, hide_overlay_js
        except Exception as e:
            err = f"Error: {e}"
            try: traceback.print_exc()
            except Exception: pass
            new_hist = (history or []) + [(user_msg or "", err)]
            return new_hist, "", new_hist, awaiting, hide_overlay_js

    send.click(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
               outputs=[chat, msg, state_history, state_awaiting, handshake],
               concurrency_limit=2, queue=True)

    msg.submit(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
               outputs=[chat, msg, state_history, state_awaiting, handshake],
               concurrency_limit=2, queue=True)

    def _on_clear():
        return [], "", [], False, '<div id="handshake-overlay">ClarityOps loaded. Paste your scenario and attach files. I’ll ask up to 5 clarifications, then produce the structured analysis</div>'

    clear.click(_on_clear, None, [chat, msg, state_history, state_awaiting, handshake])

if __name__ == "__main__":
    port = int(os.environ.get("PORT", "7860"))
    demo.launch(server_name="0.0.0.0", server_port=port, show_api=False, max_threads=8)