Creating a Knowledge Base
Embedding models, search modes, and advanced index configuration.
Go to Studio → Knowledge → Create Knowledge Base. A name is the only thing you can't change later without consequence — everything else can be edited afterward, though some changes trigger a re-index (see Changing settings later).
Embedding model
Every Knowledge Base embeds its content through an LLM Configuration you've already connected (Studio → Settings → LLMs) — there's no separate, Elizon-managed embedding provider to configure. Pick a connected LLM Configuration, then one of its embedding-capable models. If nothing you need is listed, connect it as an LLM Configuration first.
One embedding model per Knowledge Base
The embedding model is set at creation and used for every document and every search query against this KB. Changing it later re-embeds everything (see below) — mixing models within one KB isn't supported, since vectors from different models aren't comparable.
Search mode
- Vector — pure semantic similarity search (approximate nearest-neighbor over embeddings). Best for conceptual/paraphrased queries.
- Keyword — pure BM25 full-text search. Best for exact terms, part numbers, error codes.
- Hybrid (default) — blends both, weighted by Hybrid Alpha (
0= pure keyword,1= pure vector; default0.5). Start here unless you have a specific reason not to.
Advanced configuration
Three collapsible sections, collapsed by default — sensible defaults are used if you skip them entirely:
- Vector Index (HNSW) —
ef_construction,ef,max_connections,distance_metric, and related tuning parameters for the approximate nearest-neighbor index. Only worth touching if you understand the recall/latency tradeoff you're making. - Inverted Index (BM25) —
bm25_b,bm25_k1, and related keyword-search tuning parameters. - Metadata Schema — built-in fields (document name, uploader, upload date, source) plus your own typed fields (
text,text[],int,int[],bool,number,number[],date,phone) for filtering search results later.
Knowledge Base status
| Status | Meaning |
|---|---|
PENDING | Being created. |
ACTIVE | Ready — documents can be added and search works. |
INDEXING | A settings change is triggering a full re-index (see below). |
FAILED | Creation or re-index failed. |
A brand-new KB reaches ACTIVE almost immediately — creating the underlying index is synchronous and fast. It's documents that take time to index (see Documents); the KB itself is ready to accept them right away.
Changing settings later
Renaming a KB is free — no re-index. Changing anything else (embedding model, search mode configuration, HNSW/BM25 tuning, the metadata schema) triggers a full re-index: every document is re-embedded and re-indexed into a new, versioned copy of the index, and only once that finishes does Elizon atomically switch search over to it. Search keeps working on the old configuration throughout — there's no downtime window. The KB shows INDEXING status with a banner while this runs; a second settings change is rejected with a conflict until the current re-index finishes.
Duplicate names
KB names must be unique within a project. Creating a second KB with a name already in use in that project is rejected — pick a different name.