Skip to main content

AI agent integration

Overview​

Once your Sales Catalog is configured with products, categories, and properties, you can connect it to one or more AI agents. When an agent is connected, your customers can ask about products in natural language — "do you have waterproof jackets?", "anything under $500", "color blue size M" — and the agent searches the catalog, filters by the mentioned criteria, and responds with concrete results: name, from-price, brand, available variants. The connection is per-agent, not at the account level: you choose which agent consults the catalog and you can restrict it to a branch of the category tree.

How to Access​

  • Navigate to Tools > Aurora AI > Knowledge Base in the agent you want to configure.
  • Select the Sales Catalog tab.
  • Required role: Chatbot Admin.

Interface Overview​

The Sales Catalog tab inside Knowledge Base shows:

  • An activation toggle to connect/disconnect the catalog to the agent.
  • When active, an optional scope category selector — to restrict what the agent can see to a branch of the tree.
  • An informational panel that explains the requirements (Advanced model) and the scope (the agent can query every product in the catalog whose status is Active).

The Advanced model requirement​

The Sales Catalog uses hybrid search — combining semantic understanding with structured filters — and requires an AI model with sufficient capabilities. For this reason, when you activate the catalog, if the agent is not using the Advanced model, Aurora will ask for your confirmation to upgrade it automatically to the Advanced model.

While the catalog is active on the agent, you cannot downgrade the model from the agent's settings tab: the option appears locked with a clear hint. If you need to change the model, first deactivate the catalog on that agent.

Per-category scope (optional)​

After activating the catalog, you can optionally restrict the agent to a category and all its descendants. Useful when the same catalog serves multiple agents with different responsibilities:

  • A "Home care" agent only sees that branch of the tree.
  • An "Electronics" agent sees a different branch.
  • A general agent sees the entire catalog (no restriction).

The restriction is enforced server-side: the agent never receives nor mentions products outside its branch, regardless of how the customer phrases the question.

Features & Actions​

Activate the catalog on an agent​

What it does: Connects the Sales Catalog to the agent.

Steps:

  1. Open the agent and navigate to Knowledge Base > Sales Catalog tab.
  2. Turn on the activation toggle.
  3. If the agent isn't using the Advanced model, a confirmation dialog appears: "Activate and upgrade model". Confirm to continue.
  4. (Optional) Select a scope category to restrict the agent's reach.
  5. Changes save automatically.

Important notes:

  • Only products with Active status are visible to the agent. Products in Draft, Discontinued, or Archived are ignored.
  • Custom properties are exposed to the agent automatically — you don't need to configure them per agent.

Change the scope category​

What it does: Narrows (or expands) the agent's reach over the catalog.

Steps:

  1. With the catalog active, open the Scope category selector.
  2. Pick a category — the agent will see that category and all its sub-categories.
  3. To remove the restriction, set the category to "No restriction" (entire catalog).

Deactivate the catalog on an agent​

What it does: Disconnects the catalog. The agent stops having access to the products.

Steps:

  1. Turn off the activation toggle.
  2. Confirm. Once deactivated, you can change the agent's model again if you wish.

What it does: Verifies how the agent responds to real questions before promoting it to production.

Steps:

  1. Once the catalog is activated, open the agent's test chat.
  2. Ask questions in natural language, for example:
    • "What do you have in home care?"
    • "Anything in color blue size M?"
    • "Products under $500"
    • "Do you have X brand?"
  3. The agent should respond with concrete products: name, from-price, brand, and — where applicable — variant options.

How the agent uses the catalog​

When a customer asks a product question, the AI agent:

  1. Passes the query exactly as the customer said it to the hybrid search engine. It doesn't add filters the customer didn't mention (for example, if the customer says "a sweatshirt", the agent searches "sweatshirt" — it doesn't add "for men" or "for women").
  2. When the customer does mention a concrete attribute (color, brand, max price, a custom property), the agent translates it into a structured filter and applies it.
  3. The engine combines semantic understanding (it knows synonyms: "jacket" = "coat" = "windbreaker") with keywords and structured filters, and returns the most relevant results.
  4. The agent presents those results naturally in the conversation — not as a raw table, but integrated into the response.

If a search returns no results, the agent may offer to relax one of the filters ("I didn't find anything in color blue size M; would you like to see other sizes in blue?") instead of just saying "no".

Best practices​

Make sure products are in Active status​

The most common cause of "the agent doesn't mention my product" is that the product is in Draft or Archived. Change the status to Active before testing.

Fill in brand and category​

Brand and category are frequent filters ("do you have brand X?", "show me something from category Y"). Filling in both fields significantly improves precision.

Define clear custom properties​

If your business distinguishes products by color, size, weight, voltage, or any other attribute, define custom properties with clear names. The agent uses them to filter precisely.

Use the per-category scope​

If you have several agents for different business lines, restrict each one to its branch. It improves the relevance of responses and prevents an agent from recommending products outside its context.

Test with real questions​

Before exposing the agent to customers, test with the most common questions your support team receives. If the agent doesn't handle them well, it's usually a data issue — not a configuration one. Review the brand, category, and properties of the product.

Important Notes​

  • Per-agent activation is independent of the Aurora AI Catalogs (/aurora-ai/catalogs). You can use both at the same time on different agents — or on the same one — as complementary data sources.
  • When you add new products to the catalog, the agent may take a few minutes before recommending them via semantic search. Meanwhile, keyword search and structured filters already find them.
  • AI-agent catalog usage counts toward your plan's AI message usage; there's no extra cost for activating the catalog.

FAQ​

Q: Why does Aurora ask me to change the agent's model when I activate the catalog? A: Hybrid search and accurate structured filtering require a model with sufficient capacity. The Advanced model ensures the agent understands complex questions and filters precisely.

Q: Can I activate the catalog on several agents at once? A: Yes. Each agent is configured separately, with its own per-category scope where applicable.

Q: How do I prevent the agent from recommending products outside the segment it serves? A: Use the per-category scope. The agent respects the branch of the tree and never mentions products outside the assigned sub-tree.

Q: What do I do if the agent "invents" attributes when searching? A: Aurora is configured not to add context the customer didn't provide. If you notice the agent adding unsolicited filters, report it — we treat that as a real issue because it degrades result quality.

Q: Can the agent add products to a cart or close the sale? A: For now, the integration is for consultation and recommendation. The sale itself still happens through the usual commercial channels (your human team, your online store, etc.).