Knowledge and RAG
Upload your content, and have the agent answer from it — with citations.
An agent with no knowledge base answers from whatever the model happens to know, which for your business is nothing. A knowledge base is how you fix that.
Creating one
Knowledge → New knowledge base. Name it after what is in it — teams end up with several ("Product docs", "Policies", "FAQ") and an agent can search more than one.
Adding documents
Three ways in:
- Upload a file — PDF, Word, spreadsheet, CSV, Markdown or plain text.
- Add a URL — Vicero fetches and extracts the page.
- Paste text — for the things that live in someone's head rather than a file.
Each document then moves through a pipeline: extract → chunk → embed → ready. The status column tells you where it is.
| Status | Meaning |
|---|---|
queued | Waiting for a worker |
processing | Being extracted and embedded |
ready | Searchable |
failed | Something went wrong — the row explains what |
Chunks
Documents are split into chunks — passages a few hundred words long — because a whole PDF will not fit in a prompt and, even if it did, burying the relevant paragraph in fifty irrelevant pages makes answers worse, not better.
You can inspect the chunks of any document. It is worth doing once, on your most important document: if the chunks are cut in the wrong places, retrieval will be fighting the formatting.
Chunk size and overlap are configurable per knowledge base. The defaults are sensible; change them only with a reason.
Turning on retrieval
On the agent → Knowledge tab: enable RAG, select the knowledge bases, and set top-k — how many chunks to retrieve per question. Five is a good starting point. More context is not automatically better: it dilutes the relevant passage and costs tokens.
How retrieval works
Two searches run and their results are merged:
- Semantic — the question and the chunks are compared as vectors, so a chunk can match without sharing any words with the question.
- Keyword — full-text search.
Semantic search alone is bad at exact strings: product codes, SKUs, error numbers and version numbers all embed to roughly the same place. The keyword half is what makes "error BF-4021" find the page about error BF-4021.
Citations
Answers come back with the documents they used. This is the part that makes a grounded agent trustworthy — a reader can check, and you can tell at a glance whether an answer came from your content or from the model's imagination.
An agent that answers confidently while citing nothing is answering from the model, not from you. That is the thing to watch for when testing.
Keeping it current
- Reingest a document after you replace the source file.
- Delete removes the document and its chunks from retrieval immediately.
- Stale content is worse than missing content: an agent will quote last year's prices with total confidence, because as far as it knows they are the prices.
Help Center articles
Articles you publish to your Help Center can be synced into a knowledge base, so the same text serves both your public help pages and your agent. Write it once.