main.py

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import Optional, Dict, Any
import pandas as pd
import numpy as np
import os
import json
from catboost import CatBoostRegressor
from difflib import SequenceMatcher

app = FastAPI(
    title="Districtmaps.ai API",
    description="District-level NCD risk intelligence for India. Pre-computed scores + live inference.",
    version="2.0.0"
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# ── GLOBALS ───────────────────────────────────────────────────────────────────
df            = None
model         = None
feature_names = None
feature_medians = None
feature_map   = None

# ── STARTUP ───────────────────────────────────────────────────────────────────
@app.on_event("startup")
def load_all():
    global df, model, feature_names, feature_medians, feature_map

    # Pre-computed CSV
    DATA_PATH = os.getenv("DATA_PATH", "india_all_districts_risk.csv")
    df = pd.read_csv(DATA_PATH)
    df.columns = [c.strip().lower().replace(" ", "_") for c in df.columns]
    df["district_lower"] = df["district"].str.lower().str.strip()
    df["state_lower"]    = df["state"].str.lower().str.strip()
    print(f"✅ Loaded {len(df)} districts")

    # Live inference model
    model = CatBoostRegressor()
    model.load_model("model_clean_inference.cbm")
    print("✅ Model loaded")

    with open("feature_names.json")   as f: feature_names    = json.load(f)
    with open("feature_medians.json") as f: feature_medians  = json.load(f)
    with open("feature_map.json")     as f: feature_map      = json.load(f)
    print(f"✅ Feature map loaded: {len(feature_map)} mappings")

# ── HELPERS ───────────────────────────────────────────────────────────────────
def safe_float(val):
    try:
        f = float(val)
        return round(f, 4) if not np.isnan(f) else None
    except:
        return None

def format_district(row):
    return {
        "district":       row.get("district", ""),
        "state":          row.get("state", ""),
        "risk_scores": {
            "diabetes":        safe_float(row.get("diabetes_risk")),
            "blood_pressure":  safe_float(row.get("blood_pressure_risk")),
            "obesity":         safe_float(row.get("obesity_risk")),
            "anaemia":         safe_float(row.get("anaemia_risk")),
        },
        "composite_risk":  safe_float(row.get("composite_risk")),
        "risk_percentile": safe_float(row.get("diabetes_risk_norm")),
    }

def fuzzy_match_feature(input_key: str, threshold: float = 0.6):
    """Match an input column name to a model feature name."""
    input_clean = input_key.lower().strip().replace("_", " ").replace("-", " ")

    # Direct match in feature map
    if input_clean in feature_map:
        return feature_map[input_clean], 1.0

    # Partial match in feature map keys
    best_score, best_match = 0, None
    for map_key, feat_name in feature_map.items():
        score = SequenceMatcher(None, input_clean, map_key).ratio()
        if score > best_score:
            best_score = score
            best_match = feat_name

    if best_score >= threshold:
        return best_match, best_score

    # Direct match against raw feature names
    for feat in feature_names:
        feat_clean = feat.lower().replace("_", " ")
        score = SequenceMatcher(None, input_clean, feat_clean).ratio()
        if score > best_score:
            best_score = score
            best_match = feat

    if best_score >= threshold:
        return best_match, best_score

    return None, best_score

# ── ROUTES ────────────────────────────────────────────────────────────────────

@app.get("/", tags=["Info"])
def root():
    return {
        "product":     "Districtmaps.ai",
        "version":     "2.0.0",
        "description": "District-level NCD risk intelligence for India",
        "districts":   len(df) if df is not None else 0,
        "conditions":  ["diabetes", "blood_pressure", "obesity", "anaemia"],
        "validation": {
            "cross_sectional_r2": 0.7132,
            "temporal_r2":        0.6279,
            "temporal_gap":       "4 years (NFHS-4 2015-16 → NFHS-5 2019-21)",
            "features":           len(feature_names) if feature_names else 0,
        },
        "endpoints": {
            "GET  /risk":          "Pre-computed risk scores for a named district",
            "GET  /top":           "Top N highest risk districts",
            "GET  /state/{state}": "All districts within a state",
            "GET  /districts":     "Full ranked list",
            "POST /predict":       "Live inference — send your own district data",
            "GET  /features":      "List all supported input features",
        }
    }

# ── PRE-COMPUTED ENDPOINTS ────────────────────────────────────────────────────

@app.get("/risk", tags=["Pre-computed"])
def get_district_risk(district: str, state: str = None):
    mask = df["district_lower"] == district.lower().strip()
    if state:
        mask &= df["state_lower"] == state.lower().strip()
    results = df[mask]
    if results.empty:
        mask2 = df["district_lower"].str.contains(district.lower().strip(), na=False)
        if state:
            mask2 &= df["state_lower"].str.contains(state.lower().strip(), na=False)
        results = df[mask2]
    if results.empty:
        raise HTTPException(status_code=404, detail=f"District '{district}' not found.")
    return {"query": district, "matches": [format_district(row) for _, row in results.iterrows()]}

@app.get("/districts", tags=["Pre-computed"])
def get_all_districts(sort_by: str = "composite_risk", order: str = "desc", limit: int = 708):
    col = sort_by if sort_by in df.columns else "composite_risk"
    sorted_df = df.sort_values(col, ascending=(order == "asc")).head(limit)
    return {"total": len(sorted_df), "sorted_by": col,
            "districts": [format_district(row) for _, row in sorted_df.iterrows()]}

@app.get("/top", tags=["Pre-computed"])
def get_top_districts(n: int = 10, condition: str = "composite_risk"):
    col = condition if condition in df.columns else "composite_risk"
    top = df.nlargest(n, col)
    return {"condition": col, "top_n": n,
            "districts": [format_district(row) for _, row in top.iterrows()]}

@app.get("/state/{state}", tags=["Pre-computed"])
def get_state_districts(state: str):
    mask    = df["state_lower"].str.contains(state.lower().strip(), na=False)
    results = df[mask].sort_values("composite_risk", ascending=False)
    if results.empty:
        raise HTTPException(status_code=404, detail=f"State '{state}' not found.")
    return {"state": state, "districts": len(results),
            "ranked": [format_district(row) for _, row in results.iterrows()]}

# ── LIVE INFERENCE ────────────────────────────────────────────────────────────

class PredictRequest(BaseModel):
    data: Dict[str, Any]
    district_name: Optional[str] = "Unknown"
    fill_missing: Optional[bool] = True

@app.post("/predict", tags=["Live Inference"])
def predict(request: PredictRequest):
    """
    Live inference endpoint. Send any district-level health indicators
    in your own column naming convention. We fuzzy-match to our 78 features,
    fill missing values with national medians, and return a live prediction.

    Example:
    {
      "district_name": "My District",
      "data": {
        "obesity": 32.1,
        "tobacco": 18.4,
        "literacy": 89.2,
        "insurance": 45.0,
        "anaemia": 52.3
      }
    }
    """
    input_data   = request.data
    matched      = {}
    unmatched    = []
    match_report = []

    # Fuzzy match each input column to a model feature
    for input_key, input_val in input_data.items():
        feat_name, score = fuzzy_match_feature(str(input_key))
        if feat_name:
            matched[feat_name] = float(input_val)
            match_report.append({
                "input_column":    input_key,
                "mapped_to":       feat_name,
                "confidence":      round(score, 3),
                "value":           float(input_val)
            })
        else:
            unmatched.append(input_key)

    if len(matched) == 0:
        raise HTTPException(
            status_code=422,
            detail=f"None of your columns could be matched to model features. "
                   f"Call GET /features to see supported inputs."
        )

    # Build full feature vector
    feature_vector = {}
    filled_with_median = []

    for feat in feature_names:
        if feat in matched:
            feature_vector[feat] = matched[feat]
        elif request.fill_missing:
            feature_vector[feat] = feature_medians.get(feat, 0)
            filled_with_median.append(feat)
        else:
            feature_vector[feat] = 0

    # Run prediction
    X_input = pd.DataFrame([feature_vector])[feature_names]
    prediction = float(model.predict(X_input)[0])
    prediction = round(max(0, min(100, prediction)), 4)

    # Risk band
    if prediction > 15:   risk_band = "VERY HIGH"
    elif prediction > 10: risk_band = "HIGH"
    elif prediction > 7:  risk_band = "MODERATE"
    elif prediction > 4:  risk_band = "LOW"
    else:                 risk_band = "VERY LOW"

    return {
        "district":              request.district_name,
        "prediction": {
            "diabetes_risk_pct": prediction,
            "risk_band":         risk_band,
            "model":             "CatBoost · R²=0.7132 · MAE=2.06%",
        },
        "input_summary": {
            "columns_received":      len(input_data),
            "columns_matched":       len(matched),
            "columns_unmatched":     len(unmatched),
            "features_filled_median": len(filled_with_median),
            "coverage_pct":          round(len(matched) / len(feature_names) * 100, 1),
        },
        "match_report":   match_report,
        "unmatched_cols": unmatched,
        "note": f"Prediction based on {len(matched)} matched features out of {len(feature_names)} total. "
                f"{len(filled_with_median)} features filled with national district medians."
    }

@app.get("/features", tags=["Live Inference"])
def list_features():
    """Returns all supported feature names and common aliases for /predict."""
    return {
        "total_features": len(feature_names),
        "model_feature_names": feature_names,
        "common_aliases": {
            "obesity / overweight / bmi":    "Women overweight or obese (BMI ≥25)",
            "tobacco / smoking":             "Men age 15+ who use tobacco",
            "literacy / education":          "Female population who attended school",
            "insurance / health insurance":  "Households with health insurance",
            "anaemia / anemia":              "Women age 15-49 who are anaemic",
            "sanitation":                    "Households with improved sanitation",
            "teen pregnancy":                "Women 15-19 already mothers or pregnant",
            "children overweight":           "Children under 5 who are overweight",
        },
        "tip": "Column names are fuzzy-matched — send whatever naming convention you use."
    }