recommender

Similarity search recommendation engine for Fastly Compute

This is a short Fastly Compute program written in Rust 🦀. It provides ML-powered content recommendations based on a given set of content IDs.

Running locally

To test this Compute program locally, open this folder and run:

fastly compute serve

Then, use curl in a separate shell (or use your browser) to make HTTP requests to the local server. Watch the log output in the first shell.

curl -s http://127.0.0.1:7676/\?ids\=84948,97843,85035,753076,569378
# ids - comma-separated ids of objects in the Met Collection

The local development version of this program generates HNSW graphs on the fly (src/recommender_otf.rs). Inference latency increases proportionally with the number of embeddings in a target cluster.

The production version (src/recommender_kv.rs) uses precompiled search graphs stored in a Fastly KV Store, and consistently achieves 🚀 sub-100ms 🚀 response times for searches on the Met Museum's entire dataset (480K objects).

To build and publish your recommendation system, follow the instructions in ../README.md.