Embeddings shortlist the records closest in meaning to what you typed. An optional rerank model then reads your text together with each candidate and re-scores them, so the most relevant result rises to the top before anything reaches the AI step.
Every record and your query are turned into vectors by an embedding model. Records whose vectors point the same way are close in meaning — even when they share no words.
Embeddings are fast but approximate. A rerank model is slower and sharper: it reads your text together with each shortlisted record and scores how well they truly fit, then reorders the list.
Each step narrows the field. Every step also degrades gracefully, so a missing model never breaks the screen.
Your text and every candidate record are embedded and compared. Records above the threshold, best first, make the shortlist.
Local or hostedAn optional rerank model reads your text with each shortlisted record and reorders them. Only the strongest few move on.
Hosted · optionalThe AI step receives only the records you keep selected — a focused, relevant context instead of the whole ledger.
Your chosen AI engineSemantic features should never be the reason a screen stops working. Each layer falls back to the one below it.
Catalogs are read live from your providers, so new models appear without an app update. The lists below show examples available today.
| Where it runs | Example models | Good to know |
|---|---|---|
| Local (Ollama) | nomic-embed-text, mxbai-embed-large | Free, private — the text never leaves your machine. Needs Ollama running. |
| Hosted (Modelion) | bge-m3 | No local setup. Text goes through the TotalApp server to the provider. |
| Hosted (OpenRouter) | voyage-4 and other embedding models | Large catalog, free models are marked. |
| Provider | Example models | Good to know |
|---|---|---|
| OpenRouter | NVIDIA Llama Nemotron Rerank VL 1B V2 (free) | Free to try; 1B parameters, fast. |
| OpenRouter | Cohere Rerank 4 Pro / 4 Fast / v3.5, VoyageAI rerank-3 / rerank-2.5, Qwen3 Reranker 8B | Paid options, higher capacity. |
| Modelion | qwen3-reranker-0.6b | Gateway model, no separate key to manage. |
Model names and availability change. Open Settings → Agentic to see the live list for your account.
Local and hosted options trade convenience for control. You decide per feature.
| Local embedding (Ollama) | Hosted embedding & rerank | |
|---|---|---|
| Where the text goes | Stays between your browser and your own machine | Sent through the TotalApp server to the model provider |
| Setup | Install Ollama and pull a model | Pick a model from the list — nothing to install |
| Cost | Free | Free models available; paid models billed by the provider |
| Works offline | Yes | No |
Rerank is hosted-only. If your text must not leave your environment, use a local embedding model and leave Rerank unselected.
Embedding powers similar-record search in several places. Rerank is available today on one screen and will grow from there.
Better retrieval means a smaller, sharper context for the AI step — fewer wrong turns, lower cost.
Rerank reads the question and each candidate together, so near-misses stop outranking the real answer.
Send the AI step the few records that matter instead of the whole list.
Free embedding and rerank models are available, with a one-click filter in Settings.
Each layer falls back automatically, so a missing model never blocks your work.
Choose local or hosted, pick the model, and turn rerank off with a single click.
Exact records come from deterministic queries; similarity search covers free-text discovery.