Semantic connections now work with your provider
Arivu can now create embeddings with OpenAI, OpenRouter, Mistral, Ollama, LM Studio, and other providers, chosen separately from the model that writes text.
If you run Arivu with a provider other than Gemini, some of its quieter features were not working for you. Starting with Arivu 1.8.4, they are. You can now choose an embedding model separately from your text model, and most popular providers, hosted or local, can supply it.
What embeddings do in Arivu
An embedding turns a piece of text into a list of numbers. Two items with nearby numbers are likely about similar things, even when they use different words. Arivu uses embeddings for four things:
- Semantic ranking in Graph search, so a search can find items that match the idea, not only the exact words.
- Related items, the “you saved something like this before” suggestions next to a source.
- Similarity links in the Graph, the quieter connections between items that are about the same thing.
- Similar-content duplicates, so a second copy of an article under a different URL is easier to spot.
None of these is required. Capture, reading, notes, full-text Search, explicit links, and the rest of the Graph work without any provider at all.
What was wrong
Since Arivu let you pick your model provider, text generation could come from almost anywhere. Embeddings could not. They only worked when the text provider was Gemini. On any other provider, every embedding request quietly failed, and the four features above stopped adding anything. There was no clear message saying why.
What changed
Admin settings (and Settings, under Provider settings) now have two labelled groups: Text generation model and Embedding model.
The embedding provider can be:
- Same as text generation. This is the default. If your text provider also offers embeddings, Arivu uses that provider’s default embedding model and the key you already entered. Most people need to change nothing.
- A different provider. Pick any embedding-capable preset and give it its own model, key, and endpoint. The key is stored encrypted and only ever shown masked.
- Off. Arivu keeps working, and the features above fall back to keywords and extracted topics.
Embedding-capable presets include Gemini, OpenAI, OpenRouter, Mistral, Fireworks AI, Perplexity, Ollama or another local server, LM Studio, and Custom for any compatible endpoint. Some providers only offer text generation. Those are marked as text-only, so you can pair them with a separate embedding provider instead of guessing.
If embeddings are not set up, the Embedding model section tells you why and lists which features are affected. Admin > API Usage now reports “The embedding model is not configured” as its own message, instead of blaming your text provider, and shows how many items are embedded, out of date, or missing.
Switching models safely
Embeddings from two different models cannot be compared with each other, any more than two maps drawn at different scales. So Arivu now records which model created each embedding and never mixes them.
When you change the embedding provider, model, or size, older embeddings stop being used straight away. A low-priority background job then rebuilds them in small batches. It picks up where it left off after a restart, runs again after settings changes, and can be started by hand with Re-embed now. If the model name or endpoint looks wrong, the job pauses and shows a visible failure instead of working through your whole library for nothing.
While it runs, Graph search falls back to keyword ranking, and the Graph keeps its concepts and entities. Nothing disappears.
Before you upgrade
Your existing embeddings are labelled as Gemini embeddings on first start, because that is the only place they could have come from.
- On Gemini, nothing changes and nothing is rebuilt.
- On another provider that offers embeddings, such as OpenRouter or OpenAI, “Same as text generation” now starts working. On first start, Arivu embeds your existing library in the background, and that uses your provider’s quota. If you do not want that, choose a different embedding model, or Off, in Admin settings before or after upgrading.
Read Search and Ask and Environment Variables for the details, or Arivu 1.8.4 for the rest of the release.
If you have been running Arivu with a local model or a non-Gemini provider, your Graph and related items should now start filling in.