hailey.at
bloblens#
Tracks similar avatars and banners across Bluesky by hashing profile images and comparing them.
What it does#
Listens to the ATProto firehose for profile updates, hashes any avatar/banner images using PDQ perceptual hashing, and stores them in a vector database. When it finds multiple profiles using similar images, it logs them.
Useful for finding coordinated behavior, tracking ban evasion, or whatever else
Setup#
You'll need three things running:
- bloblens (this)
- Retina - image hashing service (handles PDQ hashing)
- Qdrant - vector database (stores and searches hashes)
With Docker Compose#
docker-compose up -d
That's it. It'll create the Qdrant collection automatically on first run.
Manual#
Build:
go build -o bloblens ./cmd/bloblens
Run:
./bloblens \
--retina-host=http://localhost:8080 \
--websocket-host=wss://jetstream.atproto.tools/subscribe \
--qdrant-host=localhost \
--qdrant-port=6334 \
--qdrant-collection=blobs \
--max-hamming-distance=31 \
--seen-threshold=5
Configuration#
| Flag | Description | Default |
|---|---|---|
--retina-host |
URL of your Retina instance | required |
--websocket-host |
ATProto firehose URL | required |
--qdrant-host |
Qdrant server hostname | required |
--qdrant-port |
Qdrant gRPC port | required |
--qdrant-collection |
Collection name for storing hashes | required |
--max-hamming-distance |
Max distance to consider images "similar" | 31 |
--seen-threshold |
How many matches before logging | required |
--max-limit |
Max results to fetch from Qdrant per search | required |
--max-search-time |
Timeout for vector searches | required |
How it works#
- Watches the firehose for
app.bsky.actor.profilerecords - Extracts avatar and banner blob refs
- Sends them to Retina for PDQ hashing
- Stores the 256-dimensional vector in Qdrant
- Searches Qdrant for similar existing hashes
- Logs when threshold is exceeded