How retrieval works

You don't need to understand the machinery to use collections well — the defaults are good and the agent does the work. But a little insight into how the agent finds answers helps you build better collections and trust the results. So here's the short version of what happens when a customer asks a question.

Two ways to search, at once

When the agent looks something up, it searches your collection two ways simultaneously, because each catches what the other misses:

• By meaning — it understands the intent behind the words. A search for "how do I get my money back?" finds your "Refund policy" article even though they share almost no words. This is what makes the agent feel like it understands.
• By keyword — it also matches exact words and terms. When a customer types a specific SKU, a product name, or an unusual word, exact matching makes sure it isn't lost.

Vivollo blends both into one ranked list — this is called hybrid search — so you get the understanding of meaning-based search with the precision of keyword search. No setup required; it's on by default.

Filtering happens during the search

If the customer's request implies filters — "under 500 TL", "in stock", "size 38" — those aren't applied as an afterthought. They're part of the search itself, so only qualifying items are ever considered. The result is fast and exact: the agent never shows you a perfect match that's out of stock or over budget. (More on filters in Fields & filters.)

Reranking: putting the best answer first

Hybrid search produces a strong list of candidates. Then a final step, reranking, takes a closer look and reorders them by how well each one actually answers the specific question — floating the single best result to the top.

Think of it as a two-pass approach: a fast first pass gathers good candidates, and a careful second pass picks the winner. It's especially good at nuance, and it handles Turkish and other languages well, so the most relevant answer leads — not just the one that happened to share the most words.

Why prose gets broken into passages

For written knowledge, long articles are split into smaller passages before they're stored (you saw this in Templates). That's so retrieval can be surgical: instead of handing the agent a whole troubleshooting guide and hoping, it retrieves the exact passage about your problem. The agent then answers from that passage and can point back to the source article.

What this means for you

You don't tune any of this — but knowing how it works pays off in how you build:

• Write knowledge in focused pieces. One clear article per topic retrieves better than one giant page covering everything.
• Use real customer language. Because search understands meaning, content written the way customers actually talk gets found more reliably.
• Get your fields right. Good filters (price, stock, size) are what let precise requests resolve to precise answers.
• Keep it fresh. Retrieval is only as good as the content behind it — so let your sources refresh on a sensible schedule.

That's the whole knowledge story — templates to shape it, sources to fill it, fields to structure it, and retrieval to find it. Next, let's give the agent a personality to deliver all this knowledge with.

→ Assistants