What changed
Each entry summarises merged work and links to the pull request that made it. GitHub Releases remains the authoritative record — this page is a reading list, not the source of truth.
Open GitHub ReleasesA built-in local embedding backend
Embedding now runs in the application by default, with the model downloaded on demand and remote loading disabled.
The release this site documents. Import and retrieval no longer require an embedding endpoint:
multilingual-e5-small runs in-process through ONNX, quantised, with its revision pinned, and the
model files are fetched once on demand and cached locally. Remote model loading is switched off, so
a retrieval cannot reach out.
The same release carries a renderer error boundary and two editor/UI fixes, listed above.
One vector table per notebook
Vector storage moved to a table per notebook, each carrying its own embedding width, so vectors from different embedding spaces can no longer be compared.
Retrieval became honest about embedding spaces. Each notebook now owns a vector table whose width follows the embedding configuration that produced it, which removes the possibility of ranking across two incomparable spaces rather than relying on a check to prevent it.
Model configuration was also reworked into protocol-based connections, so an OpenAI-, Anthropic- or Google-compatible endpoint is described by its protocol rather than by a registry entry.
A design system, and a script that enforces it
The interface moved onto a semantic surface ladder, gained a written design system, and acquired a token guard that fails a build which violates it.
Panel hierarchy stopped being expressed with shadows and started being expressed with a surface ladder and hairlines, described in a written design system. A token guard runs over the renderer sources and fails on a radius, shadow or raw palette value that is not in the system.
This site inherits that system, which is why its palette and type are the application’s rather than a separate set of decisions.
Before this page existed
Earlier releases are tagged in the repository and published on GitHub Releases with generated notes. This page starts at the point where the changes were worth writing up individually rather than only listing.