Our great sponsors
-
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
hey hn, supabase ceo here
As the post points out, this comes in 2 parts:
1. Embeddings models for RAG workloads (specifically pgvector). Available today.
2. Large Language Models for GenAI workloads. This will be progressively rolled out as we get our hands on more GPUs.
We've always had a focus on architectures that can run anywhere (especially important for local dev and self-hosting). In that light, we've found that the Ollama[0] tooling is really unbeatable. I heard one of our engineers explain it like "docker for models" which I think is apt.
To support models that work best with GPUs, we're running them with Fly GPUs - pretty much this: https://fly.io/blog/scaling-llm-ollama (and then we stitch a native API around it). The plan is that you will be able to "BYO" model server and point the Edge Runtime towards it using simple env vars / config.
We've also made improvements for CPU models. We built a native extension in Edge Runtime that enables using ONNX runtime via the Rust interface. This was made possible thanks to an excellent Rust wrapper, Ort[1]. We have the models stored on disk, so there is no downloading, cold-boot, etc.
The thing I most like about this set up is that you can now use Edge Functions like background workers for your Postgres database, offloading heavy compute for generating embeddings. For example, you can trigger the worker when a user inserts some text, and then the worker will asynchronously create the embedding and store it back into your database.
I'll be around if there are any questions.
[0] ollama.com
[1] Ort: https://github.com/pykeio/ort