AI Vector Search (Beta)

This feature is optional and switched off by default. A standard install has no Qdrant service and no QDRANT_URL, the Search page shows a "not enabled" notice, and no model weights are ever downloaded.

Once enabled, Saved Posts Tracker indexes every saved post's image (and detected faces) into Qdrant for semantic search — no third-party AI APIs, self-hosted.

Should you enable it?

It is worth it if you want to find posts by describing them, by uploading a similar image, or by a person's face. It costs:

If you only want the archive, the scraper and the dashboard, skip it. Everything else works exactly the same without it.

Enabling it

Search turns on when the app can see a QDRANT_URL. That is the only switch — there is no separate feature flag.

Docker Compose

A second compose file layers the Qdrant service and the env var onto the base stack, so nothing needs editing:

docker compose -f docker-compose.yml -f docker-compose.search.yml up -d

To turn it back off, drop the second file and recreate:

docker compose -f docker-compose.yml up -d --remove-orphans

Your posts are untouched either way — vectors are derived data and are rebuilt by reindexing.

Dokploy / Coolify

Those templates are pasted as a single file, so add the pieces by hand. In the app service environment: block:

      - QDRANT_URL=http://qdrant:6333

Give the app a volume for the downloaded model weights, so they survive restarts and image upgrades:

    volumes:
      - model_cache:/app/node_modules/@huggingface/transformers/.cache

Add qdrant to the app's depends_on::

    depends_on:
      mongo:
        condition: service_healthy
      qdrant:
        condition: service_healthy

Then add the service and its volume:

  qdrant:
    image: qdrant/qdrant:v1.13.4
    restart: unless-stopped
    ports:
      # Loopback only — see the security note below.
      - "127.0.0.1:6335:6333"
    volumes:
      - qdrant_data:/qdrant/storage
    ulimits:
      nofile:
        soft: 65535
        hard: 65535
    healthcheck:
      test: ["CMD-SHELL", "bash -c ': >/dev/tcp/127.0.0.1/6333' || exit 1"]
      interval: 5s
      timeout: 5s
      retries: 10
      start_period: 2s

volumes:
  qdrant_data:
  model_cache:

After enabling

Open Search and press Index Archive Now (or POST /api/search/reindex). The first run downloads the model weights, so it is slower than later ones.

Security: Qdrant ships with no authentication. Publishing its port on 0.0.0.0 exposes every vector and payload — and lets anyone delete your collections. Keep the 127.0.0.1: prefix, or set QDRANT__SERVICE__API_KEY and put it behind your reverse proxy. The app refuses to send QDRANT_API_KEY over plaintext HTTP to a non-local host.

1. Multimodal CLIP Embeddings

• Model: HuggingFace CLIP (Xenova/clip-vit-base-patch16), run server-side via @huggingface/transformers at full precision.
• Dimensionality: 512-d vectors, cosine distance.
• Query mechanism: post images and text queries are both projected into the same latent space, so a plain-language query retrieves visually matching posts (/search, "text prompt" tab). Captions and creator usernames are also matched lexically and merged in via Reciprocal Rank Fusion.
• Storage: post_images Qdrant collection, one point per post thumbnail / carousel slide.

Example natural language queries:

• "Moody neon cyberpunk street"
• "Warm wooden Scandinavian interior"
• "Minimalist typography posters with Swiss grid"

2. Face Detection & Facial Descriptors

• Model: @vladmandic/face-api on @tensorflow/tfjs, run server-side during indexing.
• Dimensionality: 128-d facial descriptor vectors, Euclidean distance.
• Storage: post_faces Qdrant collection — lets you filter all posts containing a matching face.

Reindexing Vectors

After changing the embedding model, or to index newly saved posts, run:

npm run reindex:vectors -- --all

or trigger it per-profile from the UI, or via POST /api/search/reindex.