app.py

# app.py — ClimaMind on Hugging Face Spaces (Gradio)
# Providers:
#   PROVIDER=hf_model (default) -> calls HF Inference API (tries MODEL_ID then ALT_MODEL_ID)
#   PROVIDER=local              -> loads model in Space (requires GPU)
#   PROVIDER=stub               -> offline canned answers

import os, time, json, random
import requests
import gradio as gr

# -------- Config --------
PROVIDER     = os.getenv("PROVIDER", "hf_model").strip()
MODEL_ID     = os.getenv("MODEL_ID", "LLM360/K2-Think").strip()              # default = public K2
ALT_MODEL_ID = os.getenv("ALT_MODEL_ID", "Qwen/Qwen2.5-7B-Instruct").strip() # fallback that works on serverless
HF_TOKEN     = os.getenv("HF_TOKEN", "").strip()

# -------- HTTP helper --------
def _get(url, params=None, headers=None, timeout=12, retries=2, backoff=1.6):
    for i in range(retries + 1):
        try:
            r = requests.get(url, params=params, headers=headers, timeout=timeout)
            r.raise_for_status()
            return r
        except Exception:
            if i == retries:
                raise
            time.sleep((backoff ** i) + random.random() * 0.25)

# -------- Geocode (free) --------
def geocode_city(city:str):
    r = _get("https://nominatim.openstreetmap.org/search",
             params={"q": city, "format": "json", "limit": 1},
             headers={"User-Agent": "climamind-space"})
    j = r.json()
    if not j:
        raise RuntimeError("City not found")
    return {"lat": float(j[0]["lat"]), "lon": float(j[0]["lon"]), "name": j[0]["display_name"]}

# -------- Weather (Open-Meteo, free) --------
def fetch_open_meteo(lat, lon):
    r = _get("https://api.open-meteo.com/v1/forecast", params={
        "latitude": lat, "longitude": lon,
        "current": "temperature_2m,relative_humidity_2m,wind_speed_10m,precipitation,uv_index",
        "hourly": "temperature_2m,relative_humidity_2m,wind_speed_10m,precipitation_probability,uv_index",
        "timezone": "auto"
    })
    return r.json()

# -------- PM2.5 (Open-Meteo Air-Quality, free) --------
def fetch_pm25(lat, lon):
    try:
        r = _get("https://air-quality-api.open-meteo.com/v1/air-quality", params={
            "latitude": lat, "longitude": lon, "hourly": "pm2_5", "timezone": "auto"
        }, headers={"User-Agent": "climamind-space"})
        j = r.json()
        hourly = j.get("hourly", {})
        values = hourly.get("pm2_5") or []
        if values:
            return values[-1]
    except Exception:
        pass
    return None

def fetch_factors(lat, lon):
    wx = fetch_open_meteo(lat, lon)
    cur = wx.get("current", {}) or {}
    factors = {
        "temp_c": cur.get("temperature_2m"),
        "rh": cur.get("relative_humidity_2m"),
        "wind_kmh": cur.get("wind_speed_10m"),
        "precip_mm": cur.get("precipitation"),
        "uv": cur.get("uv_index"),
        "pm25": fetch_pm25(lat, lon),
    }
    return {"factors": factors, "raw": wx}

# -------- Indices --------
def drying_index(temp_c, rh, wind_kmh, cloud_frac=None):
    base = (temp_c or 0) * 1.2 + (wind_kmh or 0) * 0.8 - (rh or 0) * 0.9
    if cloud_frac is not None:
        base -= 20 * cloud_frac
    return max(0, min(100, round(base)))

def heat_stress_index(temp_c, rh, wind_kmh):
    hs = (temp_c or 0) * 1.1 + (rh or 0) * 0.3 - (wind_kmh or 0) * 0.2
    return max(0, min(100, round(hs)))

# -------- Prompt (escape literal braces in JSON) --------
PROMPT = """You are ClimaMind, a climate reasoning assistant. Use ONLY the observations provided and return STRICT JSON.

Location: {loc} (lat={lat}, lon={lon}), local time: {t_local}
Observations: temp={temp_c}°C, rh={rh}%, wind={wind_kmh} km/h, precip={precip_mm} mm, uv={uv}, pm25={pm25}
Derived: drying_index={d_idx}, heat_stress_index={hs_idx}

Task: Answer the user’s query: "{query}" for the next 24 hours.
Steps:
 1) Identify the relevant factors.
 2) Reason causally (2–3 steps).
 3) Give a concise recommendation with time window(s) and a confidence.
 4) Output a short WHY-TRACE (3 bullets).
Return JSON ONLY:
{{
 "answer": "...",
 "why_trace": ["...", "...", "..."],
 "risk_badge": "Low"|"Moderate"|"High"
}}
"""

# -------- Reasoning providers --------
def call_stub(_prompt:str)->str:
    return json.dumps({
        "answer": "Based on 32°C, 50% RH and 12 km/h wind, cotton dries in ~2–3h (faster after 2pm).",
        "why_trace": [
            "Higher temperature and wind increase evaporation rate",
            "Moderate humidity slightly slows drying",
            "Lower afternoon cloud cover speeds it up"
        ],
        "risk_badge": "Low"
    })

# Try HF Inference (MODEL_ID -> ALT_MODEL_ID), return (text, model_used)
def call_hf_model(prompt:str) -> tuple[str, str]:
    from huggingface_hub import InferenceClient
    attempts = [m for m in [MODEL_ID, ALT_MODEL_ID] if m]
    for mid in attempts:
        try:
            client = InferenceClient(model=mid, token=(HF_TOKEN or None))
            out = client.text_generation(
                prompt,
                max_new_tokens=200,
                temperature=0.1,
                repetition_penalty=1.05,
                do_sample=False,
            )
            return str(out), mid
        except Exception as e:
            print(f"[HF_MODEL] Failed on {mid}: {repr(e)}")
            continue
    # If all failed, raise so we can stub
    raise RuntimeError(f"No serverless provider available. Tried: {attempts}")

_local_loaded = False
def _ensure_local_loaded():
    # Optional local load — requires GPU Space
    global _local_loaded, tokenizer, model
    if _local_loaded:
        return
    from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
    bnb_cfg = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype="bfloat16",
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",
    )
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        trust_remote_code=True,
        device_map="auto",           # allows CPU offload if needed
        quantization_config=bnb_cfg,
        low_cpu_mem_usage=True,
    )
    _local_loaded = True

def call_local(prompt:str)->tuple[str, str]:
    _ensure_local_loaded()
    import torch  # import here to avoid dependency if not used
    if hasattr(tokenizer, "apply_chat_template"):
        messages = [{"role":"user","content":prompt}]
        inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt").to(model.device)
    else:
        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        out = model.generate(
            **inputs,
            max_new_tokens=200,
            temperature=0.1,
            do_sample=False,
            repetition_penalty=1.05,
            eos_token_id=tokenizer.eos_token_id,
        )
    return tokenizer.decode(out[0], skip_special_tokens=True), MODEL_ID

def reason_answer(loc, coords, factors, query):
    d_idx = drying_index(factors.get("temp_c"), factors.get("rh"), factors.get("wind_kmh"))
    hs_idx = heat_stress_index(factors.get("temp_c"), factors.get("rh"), factors.get("wind_kmh"))
    t_local = time.strftime("%Y-%m-%d %H:%M")
    prompt = PROMPT.format(
        loc=loc, lat=coords["lat"], lon=coords["lon"], t_local=t_local,
        temp_c=factors.get("temp_c"), rh=factors.get("rh"), wind_kmh=factors.get("wind_kmh"),
        precip_mm=factors.get("precip_mm"), uv=factors.get("uv"), pm25=factors.get("pm25"),
        d_idx=d_idx, hs_idx=hs_idx, query=query
    )

    if PROVIDER == "hf_model":
        try:
            raw, model_used = call_hf_model(prompt)
        except Exception as e:
            print("[HF_MODEL] Falling back to stub:", repr(e))
            raw, model_used = call_stub(prompt), "stub"
    elif PROVIDER == "local":
        raw, model_used = call_local(prompt)
    else:
        raw, model_used = call_stub(prompt), "stub"

    # Extract JSON
    start, end = raw.find("{"), raw.rfind("}")
    if start == -1 or end == -1:
        parsed = {
            "answer": "The reasoning service returned non-JSON text. Please try again.",
            "why_trace": ["Response formatting issue", "Keep temperature low", "Retry once"],
            "risk_badge": "Low"
        }
    else:
        try:
            parsed = json.loads(raw[start:end+1])
        except Exception:
            parsed = {
                "answer": "Failed to parse JSON from model output.",
                "why_trace": ["JSON parsing error", "Reduce tokens/temperature", "Retry once"],
                "risk_badge": "Low"
            }
    parsed["_model_used"] = model_used
    return parsed

# -------- Gradio UI --------
def app(city, question):
    geo = geocode_city(city)
    data = fetch_factors(geo["lat"], geo["lon"])
    ans = reason_answer(
        geo["name"], {"lat": geo["lat"], "lon": geo["lon"]},
        data["factors"], question
    )
    fx = ", ".join([f"{k}={v}" for k, v in data["factors"].items()])
    why_list = ans.get("why_trace") or []
    why = "\n• " + "\n• ".join(why_list) if why_list else "\n• (no trace returned)"
    model_used = ans.pop("_model_used", "unknown")
    md = (
        f"**Answer:** {ans.get('answer','(no answer)')}\n\n"
        f"**Why-trace:**{why}\n\n"
        f"**Risk:** {ans.get('risk_badge','N/A')}\n\n"
        f"**Factors:** {fx}\n\n"
        f"<sub>Provider: {PROVIDER} • Model: `{model_used}`</sub>"
    )
    return md

demo = gr.Interface(
    fn=app,
    inputs=[
        gr.Textbox(label="City", value="New Delhi"),
        gr.Dropdown(
            choices=[
                "If I wash clothes now, when will they dry?",
                "Should I water my plants today or wait?",
                "What is the heat/wildfire risk today? Explain briefly."
            ],
            label="Question",
            value="If I wash clothes now, when will they dry?"
        )
    ],
    outputs=gr.Markdown(label="ClimaMind"),
    title="ClimaMind — K2-Think + Live Climate Data",
    description="Serverless tries K2, falls back to Qwen if needed; or run locally on GPU Space. Stub as last resort.",
    flagging_mode="never",
    concurrency_limit=2,
)

demo.queue(max_size=8)

if __name__ == "__main__":
    demo.launch()