#!/opt/homebrew/bin/python3.11
"""
Momentry Eye — Multi-model vision detection agent
Models: grounding-dino (default), paligemma
Usage:
    python3 scripts/vision_agent.py
    curl localhost:5052/health
    curl localhost:5052/detect -d '{"time":5461,"prompt":"gun","model":"grounding-dino"}'
    curl localhost:5052/search -d '{"query":"find the gun","model":"paligemma"}'
"""
import json, os, sys, time, cv2, torch, re, psycopg2, threading
from PIL import Image, ImageDraw
from flask import Flask, request, jsonify, send_file

app = Flask(__name__)

DB_URL = "postgresql://accusys@localhost:5432/momentry?host=/tmp"
BASE_DIR = "/Users/accusys/momentry/output_dev"
SHOTS_DIR = os.path.join(BASE_DIR, "vision_shots")
os.makedirs(SHOTS_DIR, exist_ok=True)
PORT = int(os.environ.get("VISION_AGENT_PORT", 5052))

DEVICE = "mps" if torch.backends.mps.is_available() else "cpu"

VIDEO_PATHS = {
    "aeed71342a899fe4b4c57b7d41bcb692":
        "/Users/accusys/momentry/var/sftpgo/data/demo/Charade (1963) Cary Grant & Audrey Hepburn \uff5c Comedy Mystery Romance Thriller \uff5c Full Movie.mp4",
}

# ======================== Model Registry ========================
MODELS = {}  # name -> {"model": obj, "processor": obj, "info": dict}

def load_gdino():
    """Load Grounding DINO Base."""
    from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
    print("[GDINO] Loading...")
    t0 = time.time()
    proc = AutoProcessor.from_pretrained("IDEA-Research/grounding-dino-base")
    model = AutoModelForZeroShotObjectDetection.from_pretrained("IDEA-Research/grounding-dino-base").to(DEVICE)
    print(f"[GDINO] Loaded in {time.time()-t0:.1f}s")
    return {
        "model": model, "processor": proc,
        "info": {
            "name": "grounding-dino", "params_m": 232, "size_mb": 891,
            "resolution": 384, "has_confidence": True,
            "license": "Apache 2.0",
        }
    }

def load_paligemma():
    """Load PaliGemma 3B mix-224."""
    from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
    print("[PaliGemma] Loading...")
    t0 = time.time()
    proc = AutoProcessor.from_pretrained("google/paligemma-3b-mix-224")
    model = PaliGemmaForConditionalGeneration.from_pretrained(
        "google/paligemma-3b-mix-224", dtype=torch.bfloat16
    ).to(DEVICE)
    print(f"[PaliGemma] Loaded in {time.time()-t0:.1f}s")
    return {
        "model": model, "processor": proc,
        "info": {
            "name": "paligemma", "params_m": 2923, "size_mb": 3000,
            "resolution": 224, "has_confidence": False,
            "license": "Gemma license",
        }
    }

MODEL_REGISTRY = {
    "grounding-dino": load_gdino,
    "paligemma": load_paligemma,
}

def get_model(name):
    """Lazy-load and cache a model by name."""
    if name not in MODELS:
        if name not in MODEL_REGISTRY:
            return None
        MODELS[name] = MODEL_REGISTRY[name]()
    return MODELS[name]

# ======================== Inference ========================
def infer_gdino(img, prompt, threshold=0.1):
    """Grounding DINO inference. Returns [{bbox, score, label}]."""
    m = get_model("grounding-dino")
    inputs = m["processor"](images=img, text=f"{prompt}.", return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        outputs = m["model"](**inputs)
    dets = m["processor"].post_process_grounded_object_detection(
        outputs, threshold=threshold, target_sizes=[img.size[::-1]])[0]
    results = []
    for i in range(len(dets["boxes"])):
        results.append({
            "bbox": [round(v, 1) for v in dets["boxes"][i].tolist()],
            "score": round(dets["scores"][i].item(), 3),
            "label": prompt,
        })
    return results

def infer_paligemma(img, prompt, threshold=0.1):
    """PaliGemma inference. Returns [{bbox, label}] — no confidence scores."""
    m = get_model("paligemma")
    inputs = m["processor"](text=f"detect {prompt}", images=img, return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        outputs = m["model"].generate(**inputs, max_new_tokens=100)
    result = m["processor"].decode(outputs[0], skip_special_tokens=True)
    # Parse PaliGemma output format: <locXXXX><locXXXX><locXXXX><locXXXX> label
    locs = re.findall(r'<loc(\d+)>', result)
    results = []
    if len(locs) >= 4:
        n_dets = len(locs) // 4
        # Extract labels (text between bbox tokens)
        labels = re.findall(r'>\s*(\w+)\s*<|>\s*(\w+)$', result.replace('detect ' + prompt, ''))
        for i in range(n_dets):
            idx = i * 4
            # Convert PaliGemma loc tokens to image coordinates (0-1024 range)
            img_w, img_h = img.size
            x1 = int(locs[idx]) / 1024 * img_w
            y1 = int(locs[idx+1]) / 1024 * img_h
            x2 = int(locs[idx+2]) / 1024 * img_w
            y2 = int(locs[idx+3]) / 1024 * img_h
            results.append({
                "bbox": [round(x1, 1), round(y1, 1), round(x2, 1), round(y2, 1)],
                "score": 1.0,
                "label": prompt,
            })
    return results

INFERENCE = {
    "grounding-dino": infer_gdino,
    "paligemma": infer_paligemma,
}

# ======================== Utilities ========================
def find_video(uuid):
    if uuid in VIDEO_PATHS: return VIDEO_PATHS[uuid]
    import glob
    base = "/Users/accusys/momentry/var/sftpgo/data/demo"
    for f in glob.glob(f"{base}/**/Charade*", recursive=True):
        if f.endswith((".mp4", ".mov", ".avi")): VIDEO_PATHS[uuid] = f; return f
    for f in glob.glob(f"{base}/**/*{uuid[:8]}*", recursive=True):
        if f.endswith((".mp4", ".mov", ".avi")): VIDEO_PATHS[uuid] = f; return f
    return None

def parse_query(query):
    query = query.lower().strip()
    prefixes = ["find ", "show ", "search ", "where is ", "where are ",
                "looking for ", "detect ", "locate ", "spot ", "scan for "]
    for p in prefixes:
        if query.startswith(p):
            query = query[len(p):]
    for a in ["a ", "an ", "the ", "some ", "any "]:
        if query.startswith(a):
            query = query[len(a):]
    query = query.rstrip(".?!,")
    for s in [" in the image", " in this scene", " in the picture",
              " being held", " in hand", " in frame", " please"]:
        if query.endswith(s):
            query = query[: -len(s)]
    return query.strip()

def resolve_target(target_str):
    if not target_str or ":" not in target_str:
        return None
    parts = target_str.split(":", 1)
    if len(parts) != 2: return None
    uuid, identifier = parts
    conn = psycopg2.connect(DB_URL)
    cur = conn.cursor()
    cur.execute("SELECT start_time, end_time FROM dev.chunks WHERE file_uuid=%s AND chunk_id=%s LIMIT 1", (uuid, identifier))
    row = cur.fetchone()
    if row: cur.close(); conn.close(); return (uuid, float(row[0]), float(row[1]))
    if identifier.isdigit():
        cid = f"{uuid}_{identifier}"
        cur.execute("SELECT start_time, end_time FROM dev.chunks WHERE file_uuid=%s AND chunk_id=%s LIMIT 1", (uuid, cid))
        row = cur.fetchone()
        if row: cur.close(); conn.close(); return (uuid, float(row[0]), float(row[1]))
    tid = identifier.replace("trace_", "")
    cur.execute("SELECT MIN(start_time), MAX(end_time) FROM dev.chunks WHERE file_uuid=%s AND chunk_type='trace' AND chunk_id LIKE %s", (uuid, f"%_trace_{tid}"))
    row = cur.fetchone()
    if row and row[0] is not None: cur.close(); conn.close(); return (uuid, float(row[0]), float(row[1]))
    cur.close(); conn.close()
    return None

def register_resource(resource_id, name, info):
    try:
        conn = psycopg2.connect(DB_URL)
        cur = conn.cursor()
        cur.execute("""
            INSERT INTO dev.resources (resource_id, resource_type, category, capabilities, config, metadata, status, last_heartbeat)
            VALUES (%s, %s, %s, %s::jsonb, %s::jsonb, %s::jsonb, %s, NOW())
            ON CONFLICT (resource_id) DO UPDATE SET status=%s, last_heartbeat=NOW(), config=EXCLUDED.config
        """, (
            resource_id, "vision_model", "object_detection",
            json.dumps({"detect": "Single-frame detection", "search": "Range search with NL query",
                        "has_confidence": info.get("has_confidence", True)}),
            json.dumps({"name": name, "port": PORT, "device": DEVICE, "params_m": info.get("params_m"),
                        "resolution": info.get("resolution"), "license": info.get("license")}),
            json.dumps({"version": "2.0", "docs": "/health"}),
            "online", "online"))
        conn.commit(); cur.close(); conn.close()
        print(f"[Resource] Registered '{resource_id}'")
    except Exception as e:
        print(f"[Resource] Register '{resource_id}' failed: {e}")

def heartbeat_loop(resource_ids):
    while True:
        try:
            conn = psycopg2.connect(DB_URL)
            cur = conn.cursor()
            for rid in resource_ids:
                cur.execute("UPDATE dev.resources SET last_heartbeat = NOW() WHERE resource_id = %s", (rid,))
            conn.commit(); cur.close(); conn.close()
        except: pass
        time.sleep(60)

# ======================== Annotate ========================
def annotate_image(img, detections, prompt):
    draw = ImageDraw.Draw(img)
    for d in detections:
        b = d["bbox"]
        score = d.get("score", 1.0)
        draw.rectangle(b, outline="lime", width=3)
        draw.text((b[0], b[1]-18), f"{prompt} {score:.2f}", fill="lime")
    return img

# ======================== API Routes ========================
@app.route("/models", methods=["GET"])
def list_models():
    """List available models and their status."""
    result = []
    for name, loader in MODEL_REGISTRY.items():
        cached = name in MODELS
        info = dict(MODELS[name]["info"]) if cached else {"name": name, "loaded": False}
        info["loaded"] = cached
        result.append(info)
    return jsonify({"models": result})

# Default fusion weights: GDINO 0.6, PaliGemma 0.4
FUSION_WEIGHTS = {"grounding-dino": 0.6, "paligemma": 0.4}

@app.route("/detect", methods=["POST"])
def detect():
    data = request.json or {}
    uuid = data.get("uuid", "aeed71342a899fe4b4c57b7d41bcb692")
    t_sec = data.get("time", 0)
    prompt = data.get("prompt", "gun")
    model_name = data.get("model", "grounding-dino")
    threshold = data.get("threshold", 0.1)
    weights = data.get("weights", None)  # e.g. {"grounding-dino":0.7,"paligemma":0.3}
    fusion_weights = weights if weights else \
        ({model_name: 1.0} if model_name != "fusion" else FUSION_WEIGHTS)

    # Determine which models to run
    if model_name == "fusion":
        models_to_run = list(INFERENCE.keys())
    elif model_name in INFERENCE:
        models_to_run = [model_name]
    else:
        return jsonify({"error": f"Unknown model: {model_name}"}), 400

    video = find_video(uuid)
    if not video: return jsonify({"error": "Video not found"}), 404

    cap = cv2.VideoCapture(video)
    cap.set(cv2.CAP_PROP_POS_FRAMES, int(t_sec * (cap.get(cv2.CAP_PROP_FPS) or 25.0)))
    ret, frame = cap.read()
    cap.release()
    if not ret: return jsonify({"error": f"Cannot read frame at {t_sec}s"}), 400

    img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))

    all_detections = {}
    fusion_results = []
    t0 = time.time()

    for mn in models_to_run:
        if mn not in INFERENCE: continue
        detections = INFERENCE[mn](img, prompt, threshold)
        all_detections[mn] = detections
        w = fusion_weights.get(mn, 0.5)

        for d in detections:
            gdino_score = d.get("score", 1.0)
            # PaliGemma has no score, treat detected=1.0
            model_score = gdino_score if mn == "grounding-dino" else 1.0
            fused = round(model_score * w, 3)

            fusion_results.append({
                "bbox": d["bbox"],
                "label": d["label"],
                "score": model_score,
                "fused_score": fused,
                "source_model": mn,
            })

    infer_ms = (time.time() - t0) * 1000

    # Deduplicate by bbox IOU for fusion mode
    if model_name == "fusion" and len(fusion_results) > 1:
        deduped = []
        fusion_results.sort(key=lambda x: -x["fused_score"])
        for r in fusion_results:
            overlap = False
            for d in deduped:
                b1, b2 = r["bbox"], d["bbox"]
                iou = calc_iou(b1, b2)
                if iou > 0.5:
                    overlap = True
                    break
            if not overlap:
                deduped.append(r)
        fusion_results = deduped

    # Annotate with best result
    display_dets = [{"bbox": r["bbox"], "score": r["fused_score"], "label": prompt} for r in fusion_results]
    if model_name != "fusion":
        display_dets = all_detections.get(model_name, [])

    img_ann = annotate_image(img.copy(), display_dets, prompt)
    shot_name = f"{uuid[:8]}_{int(t_sec)}s_{prompt}_{model_name}.jpg"
    img_ann.save(os.path.join(SHOTS_DIR, shot_name))

    return jsonify({
        "model": model_name,
        "fusion_weights": fusion_weights,
        "models_used": models_to_run,
        "per_model": {mn: {"detections": all_detections.get(mn, []),
                          "n_detections": len(all_detections.get(mn, []))}
                     for mn in models_to_run},
        "fusion": fusion_results if model_name == "fusion" else None,
        "detections": display_dets,
        "time_ms": round(infer_ms, 1),
        "n_detections": len(display_dets),
        "shot_url": f"/shots/{shot_name}",
    })

def calc_iou(b1, b2):
    xi1 = max(b1[0], b2[0]); yi1 = max(b1[1], b2[1])
    xi2 = min(b1[2], b2[2]); yi2 = min(b1[3], b2[3])
    inter = max(0, xi2 - xi1) * max(0, yi2 - yi1)
    a1 = (b1[2]-b1[0])*(b1[3]-b1[1])
    a2 = (b2[2]-b2[0])*(b2[3]-b2[1])
    return inter / (a1 + a2 - inter + 1e-10)

@app.route("/search", methods=["POST"])
def search():
    data = request.json or {}
    uuid = data.get("uuid", "aeed71342a899fe4b4c57b7d41bcb692")
    target_str = data.get("target", "")
    query = data.get("query", "find the gun")
    range_str = data.get("range", "0-6780")
    interval = data.get("interval", 30)
    threshold = data.get("threshold", 0.15)
    model_name = data.get("model", "grounding-dino")

    if model_name not in INFERENCE:
        return jsonify({"error": f"Unknown model: {model_name}. Available: {list(INFERENCE.keys())}"}), 400

    # Parse query → object name
    prompt = parse_query(query)
    if not prompt:
        return jsonify({"error": f"Cannot parse query: {query}"}), 400

    # Resolve target → time range
    resolved_label = ""
    if target_str:
        resolved = resolve_target(target_str)
        if not resolved:
            return jsonify({"error": f"Cannot resolve target: {target_str}"}), 404
        uuid, range_start, range_end = resolved
    else:
        parts = range_str.split("-") if "-" in range_str else ["0", "6780"]
        range_start = float(parts[0])
        range_end = float(parts[1]) if len(parts) > 1 else 6780

    video = find_video(uuid)
    if not video: return jsonify({"error": "Video not found"}), 404

    cap = cv2.VideoCapture(video)
    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))

    hits = []
    t_start = time.time()
    infer_fn = INFERENCE[model_name]
    frame_step = int(interval * fps)

    for frame_num in range(int(range_start * fps), min(int(range_end * fps), total_frames), frame_step):
        cap.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
        ret, frame = cap.read()
        if not ret: continue
        img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
        detections = infer_fn(img, prompt, threshold)
        if detections:
            ts = frame_num / fps
            best = max(d.get("score", 1.0) for d in detections)
            hits.append({
                "time": round(ts, 1),
                "time_str": f"{int(ts//60)}:{int(ts%60):02d}.{int((ts%1)*fps):02d}",
                "frame": frame_num,
                "n_detections": len(detections),
                "best_score": best,
                "detections": detections[:3],
            })
        if len(hits) >= 100: break

    cap.release()
    elapsed = time.time() - t_start

    return jsonify({
        "model": model_name,
        "query": query, "object": prompt,
        "target": target_str or None,
        "range": f"{range_start:.0f}-{range_end:.0f}",
        "interval_secs": interval,
        "hits": hits,
        "n_hits": len(hits),
        "elapsed_secs": round(elapsed, 1),
    })

@app.route("/multimodal", methods=["POST"])
def multimodal_search():
    """Multi-modal search across all chunk types.
    For sentence chunks: ASR text + visual confirmation.
    For trace/story/cut chunks: visual detection only (no ASR text).
    Input:
      {"keyword":"gun"}        — find chunks mentioning "gun" in ASR + visually confirm
      {"keyword":"gun","chunk_type":"trace"}  — search trace chunks visually (no ASR)
      {"target":"file_uuid:chunk_id"} — search a specific chunk visually
    """
    data = request.json or {}
    uuid = data.get("uuid", "aeed71342a899fe4b4c57b7d41bcb692")
    keyword = data.get("keyword", "")
    prompt = data.get("prompt", keyword or "")
    target_str = data.get("target", "")
    chunk_type = data.get("chunk_type", "sentence")  # sentence, trace, story, cut
    threshold = data.get("threshold", 0.15)
    model_name = "grounding-dino"

    conn = psycopg2.connect(DB_URL)
    cur = conn.cursor()

    # Resolve target first if provided
    if target_str:
        resolved = resolve_target(target_str)
        if not resolved:
            return jsonify({"error": f"Cannot resolve target: {target_str}"}), 404
        uuid, st, et = resolved
        cur.execute("SELECT chunk_id, chunk_index, chunk_type, text_content FROM dev.chunks WHERE file_uuid=%s AND start_time=%s AND end_time=%s LIMIT 1",
                    (uuid, st, et))
        chunks = [(r[0], r[1], r[2], st, et, r[3] or "") for r in cur.fetchall()]
    elif keyword and chunk_type == "sentence":
        # Search sentence chunks by ASR text keyword
        cur.execute("""
            SELECT chunk_id, chunk_index, chunk_type, start_time, end_time, text_content
            FROM dev.chunks
            WHERE file_uuid=%s AND chunk_type='sentence'
              AND text_content ILIKE CONCAT('%%', %s, '%%')
            ORDER BY start_time
        """, (uuid, keyword))
        chunks = cur.fetchall()
    else:
        # Search any chunk type by time range (visual only, no ASR)
        range_str = data.get("range", "0-6780")
        parts = range_str.split("-") if "-" in range_str else ["0", "6780"]
        rs, re = float(parts[0]), float(parts[1]) if len(parts) > 1 else 6780
        cur.execute("""
            SELECT chunk_id, chunk_index, chunk_type, start_time, end_time, COALESCE(text_content, '')
            FROM dev.chunks
            WHERE file_uuid=%s AND chunk_type=%s
              AND start_time BETWEEN %s AND %s
            ORDER BY start_time
        """, (uuid, chunk_type, rs, re))
        chunks = cur.fetchall()

    conn.close()

    if not chunks:
        return jsonify({"error": f"No matching chunks found"}), 404

    # Visual confirmation
    video = find_video(uuid)
    if not video:
        return jsonify({"error": "Video not found"}), 404

    cap = cv2.VideoCapture(video)
    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
    infer_fn = INFERENCE.get(model_name)

    results = []
    t_start = time.time()

    for chunk_id, chunk_idx, ctype, st, et, text in chunks:
        center = (st + et) / 2
        frame_num = int(center * fps)
        cap.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
        ret, frame = cap.read()
        if not ret: continue

        img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
        detections = infer_fn(img, prompt or keyword, threshold)

        entry = {
            "chunk_id": chunk_id,
            "chunk_index": chunk_idx,
            "chunk_type": ctype,
            "time_range": f"{st:.1f}-{et:.1f}",
            "time_str": f"{int(st//60)}:{int(st%60):02d}-{int(et//60)}:{int(et%60):02d}",
            "visual_confirmed": len(detections) > 0,
            "best_score": round(max(d.get("score", 1.0) for d in detections), 3) if detections else 0,
            "n_visual_dets": len(detections),
        }
        if keyword and ctype == "sentence":
            entry["asr_text"] = text[:150]
            entry["asr_matched"] = keyword.lower() in text.lower()

        results.append(entry)

    cap.release()
    elapsed = time.time() - t_start

    return jsonify({
        "keyword": keyword or prompt,
        "chunk_type": chunk_type,
        "target": target_str or None,
        "total_chunks": len(chunks),
        "visual_confirmed": sum(1 for r in results if r["visual_confirmed"]),
        "asr_matched": sum(1 for r in results if r.get("asr_matched")),
        "elapsed_secs": round(elapsed, 1),
        "results": results,
    })

@app.route("/shots/<filename>")
def serve_shot(filename):
    path = os.path.join(SHOTS_DIR, filename)
    if not os.path.exists(path): return jsonify({"error": "Not found"}), 404
    return send_file(path, mimetype="image/jpeg")

@app.route("/health")
def health():
    loaded = list(MODELS.keys())
    available = list(MODEL_REGISTRY.keys())
    return jsonify({
        "status": "ok",
        "models_loaded": loaded,
        "models_available": available,
        "device": DEVICE,
        "port": PORT,
    })

if __name__ == "__main__":
    # Register both as resources
    gdino_info = {"params_m": 232, "resolution": 384, "has_confidence": True, "license": "Apache 2.0"}
    pg_info = {"params_m": 2923, "resolution": 224, "has_confidence": False, "license": "Gemma license"}
    register_resource("eye-gdino", "grounding-dino", gdino_info)
    register_resource("eye-paligemma", "paligemma", pg_info)

    # Start heartbeat
    t = threading.Thread(target=heartbeat_loop, args=(["eye-gdino", "eye-paligemma"],), daemon=True)
    t.start()

    # Pre-load grounding-dino by default
    print(f"\n{'='*60}")
    print(f"  👁️  Momentry Eye — port {PORT}")
    print(f"{'='*60}")
    print(f"  Models: {', '.join(MODEL_REGISTRY.keys())}")
    print(f"  Device: {DEVICE}")
    print(f"  Resources: eye-gdino, eye-paligemma")
    print(f"  Loading default model...")
    get_model("grounding-dino")
    print(f"  👁️  Ready: http://localhost:{PORT}")
    app.run(host="0.0.0.0", port=PORT, threaded=True)