app.py

# app.py
import os
import uuid
import io
from PIL import Image
import gradio as gr
import numpy as np

# CLIP via Sentence-Transformers (text+image to same 512-dim space)
from sentence_transformers import SentenceTransformer

# Gemini (Google) client
from google import genai

# Qdrant client & helpers
from qdrant_client import QdrantClient
from qdrant_client.http.models import VectorParams, Distance, PointStruct

# -------------------------
#  CONFIG (reads env vars)
# -------------------------
GEMINI_API_KEY = os.environ.get("GEMINI_API_KEY")  # set in Hugging Face Space secrets
QDRANT_URL = os.environ.get("QDRANT_URL")          # set in Hugging Face Space secrets
QDRANT_API_KEY = os.environ.get("QDRANT_API_KEY")  # set in Hugging Face Space secrets

# Local fallbacks (for local testing) - set them before running locally if needed:
# os.environ["GEMINI_API_KEY"]="..." ; os.environ["QDRANT_URL"]="..." ; os.environ["QDRANT_API_KEY"]="..."

# -------------------------
#  Initialize clients/models
# -------------------------
print("Loading CLIP model (this may take 20-60s the first time)...")
MODEL_ID = "sentence-transformers/clip-ViT-B-32-multilingual-v1"
clip_model = SentenceTransformer(MODEL_ID)  # model maps text & images to same vector space

# Gemini client (for tags/captions)
if GEMINI_API_KEY:
    genai_client = genai.Client(api_key=GEMINI_API_KEY)
else:
    genai_client = None

# Qdrant client
if not QDRANT_URL:
    # If you prefer local Qdrant for dev: client = QdrantClient(":memory:") or local url
    raise RuntimeError("Please set QDRANT_URL environment variable")
qclient = QdrantClient(url=QDRANT_URL, api_key=QDRANT_API_KEY)

COLLECTION = "lost_found_items"
VECTOR_SIZE = 512

# Create collection if missing
if not qclient.collection_exists(COLLECTION):
    qclient.create_collection(
        collection_name=COLLECTION,
        vectors_config=VectorParams(size=VECTOR_SIZE, distance=Distance.COSINE),
    )

# -------------------------
#  Helpers
# -------------------------
def embed_text(text: str):
    vec = clip_model.encode(text, convert_to_numpy=True)
    return vec

def embed_image_pil(pil_img: Image.Image):
    # sentence-transformers supports directly encoding a PIL image for CLIP models
    vec = clip_model.encode(pil_img, convert_to_numpy=True)
    return vec

def gen_tags_from_image_file(local_path: str) -> str:
    """Upload image file to Gemini and ask for 4 short tags.
       Returns the raw text response (expected comma-separated tags)."""
    if genai_client is None:
        return ""
    # Upload file (Gemini Developer API supports client.files.upload)
    file_obj = genai_client.files.upload(file=local_path)
    # Ask Gemini: produce short tags only
    prompt_text = (
        "Give 4 short tags (comma-separated) describing this item in the image. "
        "Tags should be short single words or two-word phrases (e.g. 'black backpack', 'water bottle'). "
        "Respond only with tags, no extra explanation."
    )
    response = genai_client.models.generate_content(
        model="gemini-2.5-flash",
        contents=[prompt_text, file_obj],
    )
    return response.text.strip()

# -------------------------
#  App logic: add item
# -------------------------
def add_item(mode: str, uploaded_image, text_description: str):
    """
    mode: 'lost' or 'found'
    uploaded_image: PIL image or None
    text_description: str
    """
    item_id = str(uuid.uuid4())
    payload = {"mode": mode, "text": text_description}

    if uploaded_image is not None:
        # Save image to temp file (so we can upload to Gemini)
        tmp_path = f"/tmp/{item_id}.png"
        uploaded_image.save(tmp_path)
        # embed image
        vec = embed_image_pil(uploaded_image).tolist()
        payload["has_image"] = True
        # optional: get tags from Gemini (if available)
        try:
            tags = gen_tags_from_image_file(tmp_path)
        except Exception as e:
            tags = ""
        payload["tags"] = tags
        # store image bytes (tiny) so we can show result in the UI (base64)
        with open(tmp_path, "rb") as f:
            b64 = f.read()
        payload["image_b64"] = True  # flag (we will return/show image via Gradio from file bytes)
    else:
        # only text provided
        vec = embed_text(text_description).tolist()
        payload["has_image"] = False
        # ask Gemini to suggest tags from text
        if genai_client:
            try:
                resp = genai_client.models.generate_content(
                    model="gemini-2.5-flash",
                    contents=f"Give 4 short, comma-separated tags for this item described as: {text_description}. Reply only with tags."
                )
                payload["tags"] = resp.text.strip()
            except Exception:
                payload["tags"] = ""
        else:
            payload["tags"] = ""

    # Upsert into Qdrant
    point = PointStruct(id=item_id, vector=vec, payload=payload)
    qclient.upsert(collection_name=COLLECTION, points=[point], wait=True)

    return f"Saved item id: {item_id}\nTags: {payload.get('tags','')}"


# -------------------------
#  App logic: search
# -------------------------
def search_items(query_image, query_text, limit: int = 5):
    # produce query embedding
    if query_image is not None:
        qvec = embed_image_pil(query_image).tolist()
        q_type = "image"
    else:
        if (not query_text) or (len(query_text.strip()) == 0):
            return "Please provide a query image or some query text."
        qvec = embed_text(query_text).tolist()
        q_type = "text"

    hits = qclient.search(collection_name=COLLECTION, query_vector=qvec, limit=limit)

    # Format output (list)
    results = []
    for h in hits:
        payload = h.payload or {}
        score = getattr(h, "score", None)
        results.append(
            {
                "id": h.id,
                "score": float(score) if score is not None else None,
                "mode": payload.get("mode", ""),
                "text": payload.get("text", ""),
                "tags": payload.get("tags", ""),
                "has_image": payload.get("has_image", False),
            }
        )
    # Return a simple list for Gradio to show
    if not results:
        return "No results."
    # Convert to text for display
    out_lines = []
    for r in results:
        out_lines.append(f"id:{r['id']} score:{r['score']:.4f} mode:{r['mode']} tags:{r['tags']} text:{r['text']}")
    return "\n\n".join(out_lines)

# -------------------------
#  Gradio UI
# -------------------------
with gr.Blocks(title="Lost & Found — Simple Helper") as demo:
    gr.Markdown("## Lost & Found Helper (image/text search) — upload items, then search by image or text.")
    with gr.Row():
        with gr.Column():
            mode = gr.Radio(choices=["lost", "found"], value="lost", label="Add as")
            upload_img = gr.Image(type="pil", label="Item photo (optional)")
            text_desc = gr.Textbox(lines=2, placeholder="Short description (e.g. 'black backpack with blue zipper')", label="Description (optional)")
            add_btn = gr.Button("Add item")
            add_out = gr.Textbox(label="Add result", interactive=False)
        with gr.Column():
            gr.Markdown("### Search")
            query_img = gr.Image(type="pil", label="Search by image (optional)")
            query_text = gr.Textbox(lines=2, label="Search by text (optional)")
            search_btn = gr.Button("Search")
            search_out = gr.Textbox(label="Search results", interactive=False)

    add_btn.click(add_item, inputs=[mode, upload_img, text_desc], outputs=[add_out])
    search_btn.click(search_items, inputs=[query_img, query_text], outputs=[search_out])

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)