AACSearch Assistant builds its answer from the documents it found and returns the supporting sources. Your visitor can check where a statement came from.Try the AssistantSee the API
Point it at one or more collections — no separate vector store to stand up.
Every turn retrieves, grounds the answer in real documents, and streams the response with citations attached.
Pass a conversation id on the next turn — the assistant remembers what was already asked.
Instead of this model:
AACSearch uses:
That makes retrieval part of a controlled search infrastructure.
The AI Assistant does not have to replace classic search. Use them together:
an exact term
regular results
a question
a grounded answer
a complex exploration
answer + sources + related results
Reliable AI search has to be able to:
The Search UI, the Assistant and MCP use one retrieval infrastructure. That means fewer parallel indexes and fewer diverging access models.
What matters for AI search:
Retrieval quality is measurable today with golden query sets in the Quality Lab; connect the Assistant to it and to analytics.
Every answer links back to the specific documents it was grounded in.
Entities and relationships extracted from your content, deduplicated and clustered into communities.
A primer answer plus targeted follow-up searches for questions that need breadth and depth.
Server-Sent Events for the UI, billed per assistant turn from a credit wallet.
The assistant, search and recommendations share one session id, so a conversation, a click and a purchase all attribute to the same visitor — no separate identity to wire up.
{ "messages": [{ "role": "user", "content": "Does it come in wide fit?" }] }
// -> Server-Sent Events: delta chunks, then
// { "sources": [{ "doc_id": "42", "title": "Trail Runner GT" }] }Explore more
The best way to evaluate a search platform isn't reading one more comparison table.
Only then make the decision.