Search
Three kinds: a model searching the web, you searching everything you have, and library search that understands meaning.
Web search for models
Web search is a tool a model calls when a question needs it; the sources stay in the transcript. Configure it under Admin → Web search:
| Provider | Needs |
|---|---|
| DuckDuckGo | Nothing — no account, no key. (From source, install the search extra.) |
| SearXNG | Your own instance with its JSON format enabled: add - json under search.formats in its settings.yml. |
| Firecrawl | An API key, or your own self-hosted Firecrawl. |
It is only offered to models marked for tools under Admin → Models, and to people with the web-search permission. A companion fetch tool reads a page. Addresses on the server’s own machine and network are refused unless an administrator allows them — an address can come from a model, and the server can reach things your browser cannot.
Searching your own things
The magnifier at the top of the sidebar searches, in one box, your chats and agent sessions (titles and what was said, archived ones included), knowledge, notes, skills, reports and memories. Results are grouped by kind, five of each with Show all, or narrowed to one kind. Only what you may already read is searched.
Semantic search in the library
Out of the box, library search matches keywords and ranks them well — nothing to run, nothing sent anywhere, nothing that stops working offline. It will not match a paraphrase, so a line of description on a document is worth writing.
Choose an embedding model and search fuses keywords with meaning: “how do I get paid” finds a document that says “invoicing”. Choosing none is not a degraded mode; it is exactly the keyword search, with nothing written and no requests made.
- Records are indexed in the background as they are saved; a record that could not be indexed is still saved and found by keywords.
- Changing the embedding model needs a rebuild of the index. Until then the old vectors are ignored, never mixed with the new model’s.
- A data group can have an embedding model of its own.
Over the API
POST /api/v1/search and POST /api/v1/fetch give a program — such as the LLeMbas CLI — the server’s web search and page fetching, so a device needs no search provider of its own. See API.