Indexing & sync

Pipeline

1. Fetch — the connector retrieves changed items (by change notification or by comparing modification times).
2. Extract — text and structure are extracted; scanned pages are OCR'd; spreadsheets are read sheet by sheet.
3. Chunk — documents are split into overlapping chunks that respect headings, tables and lists. Chunk size is tuned per content type.
4. Enrich — each chunk gets metadata: document, location, author, modified date, language, permissions.
5. Embed & index — chunks are embedded for semantic search and indexed for keyword search. Embedding runs in the workspace's region.
6. Publish — new versions replace old ones atomically; a document is never half-updated.

Freshness

With change notifications, a change in the source is typically searchable within 2–5 minutes. Scheduled sources refresh at the interval you set. Knowledge → (knowledge base) → Freshness shows the distribution of document ages.

Deletions

Deleted or unshared items are removed on the next sync. Permission changes are treated the same way as content changes.

Languages

Retrieval is multilingual: a question in Finnish finds relevant passages in English documents and vice versa. Answers are given in the language of the question unless the user sets a preference.

Limits

Re-indexing

Changing chunking settings or upgrading the embedding model triggers a background re-index. The old index serves queries until the new one is complete.