Firn has a Python package now! Just `pip install firn` Vector and full-text search on object storage, straight from Python. The engine embeds in-process, so there's no server: point it at a local folder or an S3 bucket. Imagine how happy your SysOps team will be with no cluster to manage. Or your CFO, with no managed solution to approve

A bit more on what WASM is doing in metabare.com. WebAssembly lets compiled code run inside the browser. In this case, that’s ONNX Runtime doing the model inference, with Transformers.js handling loading and preprocessing. I’m running two models: CLIP ViT-B/32 turns images and search phrases into comparable vectors, so you can find an image by describing it. MiniLM-L6-v2 does the same for notes and text queries. The browser compares those vectors locally to find matches. Both use quantised weights to reduce the download and memory footprint. The inference runs in a Web Worker, off the main UI thread, using the device’s CPU. The same approach could support speech transcription, translation, text classification and smaller text-generation models. What’s practical depends on the model size, available memory and how long you’re willing to wait

I took metabare.com offline recently, keeping AWS infra running models was too expensive for a side project. Now it’s back online, Image and note search running entirely in your browser using WebAssembly (WASM). WASM lets the browser run the model inference on your own CPU, so it can generate embeddings on the device. Your images, notes and searches stay on your device too. That means I don’t need to run a GPU instance, run Triton Inference Server or manage a Kubernetes cluster to serve the models. There’s no server-side inference ticking away while nobody’s using the site.

Huh, that’s really clever. An average open source project might have a readme. Better still, a Skill for an LLM, or a pre-trained agent. Imagine trying a new database with its own agent and it figures out importing your data format so you don’t have to

Very cool. Could have used this a few years ago for Mariadb schema changes 😅

Great to see Turbopuffer testing Jev too. I recently ran a similar experiment with Firn: nDCG@10 improved from 0.727 → 0.800, with relevant results reaching rank #1 70% of the time vs 61% before. I also tested a few open-source alternatives: gordonmurray.ie/data/2026/09…

Jev is all over my timeline this week, so I added it to Firn's search as a reranker. On SciFact, nDCG@10 went from 0.727 to 0.800: the right abstract first for 70% of claims instead of 61%. About $0.0016 a search. A write-up and the free models I tried: gordonmurray.ie/data/2026/09…

I tested classifier.dev against direct Jev for reranking Firn search results. It uses the same Jev model, matched Jev’s ranking quality, gave me worse score resolution, allows about 800 searches/day/IP at my 25-candidate rerank depth, has no published terms page I could find, and doesn’t appear to support model pinning.