Guides
Start typing to search…

Build a two-agent product lookup system

A single agent can sometimes struggle to be both a good conversationalist and an accurate product searcher. Solution could be to split the job: one agent handles conversation, the other handles catalog search. They communicate via MCP, so you can update search logic without touching the customer-facing experience.

When to use this pattern

Use two agents when:

  • Persona instructions are too long or conflicting, causing inconsistent results
  • You need detailed sales/support behavior AND precise product matching
  • Catalog output requires strict formatting that clashes with conversational tone

How it works

Sales Agent - customer-facing. Handles greetings, qualification, objection handling, follow-ups. Queries the Catalog Search Agent for product data instead of searching its own knowledge base.

Catalog Search Agent - internal product search engine. Has your full catalog, responds in structured format. Never talks to customers directly.

Connection: The Catalog Search Agent is deployed as an MCP Server. The Sales Agent adds its MCP URL as a Custom Action. When the Sales Agent needs product data, it calls the Catalog Search Agent and weaves results into the conversation.

Step 1: Create the Catalog Search Agent

Add your catalog data

  1. Create a new agent.
  2. Add your product catalog as data sources - PDFs, CSVs, spreadsheets, or connect your Shopify store, WordPress site, or sitemap.
  3. For frequently changing catalogs, enable auto-sync.

👉 Learn more about data sources and auto-sync

Enable Numeric Search

  1. Go to Personalize > Citation, scroll to Numeric Search.
  2. Click Enabled.
  3. Click Save settings.

Exact matching on SKUs, part numbers, and alphanumeric codes.

👉 Learn more about Numeric Search

Set the persona instructions

Go to Personalize > Agent Persona and paste:

You are a product catalog search engine for [Company Name]. You receive search queries from another agent and return structured product results.

RESPONSE FORMAT - follow this exactly for every query:

For each matching product, return:
- **Product name**: [full product name]
- **Part number**: [SKU or part number]
- **Description**: [one sentence - what it is and what it's for]
- **Key specs**: [the 3-5 most relevant specs like dimensions, material, voltage, weight]
- **Product URL**: [link to the product page, if available]

RULES:
- Return up to 5 matching products per query, ranked by relevance.
- If no exact match exists, say "No exact match found" and show the 3 closest alternatives.
- Never invent or guess part numbers. If a part number isn't in the catalog, say so.
- If the query is vague (e.g., "brass fitting"), ask what size, type, or application before returning results.
- Keep descriptions factual. No sales language, no opinions. Just specs and facts.

Deploy as MCP Server

  1. Go to the Catalog Search Agent's Deploy page.
  2. Open the MCP Server tab.
  3. Leave only "Send a message" permission enabled.
  4. Click Save permissions.
  5. Copy the MCP Server URL - you'll need it in the next step.

👉 Learn more about MCP Server deployment

Step 2: Create the Sales Agent

Add your company knowledge (not the product catalog)

  1. Create a new agent.
  2. Add non-catalog data sources: pricing guides, shipping policies, return policies, FAQs, company info.
  3. Do not add the product catalog here. Product search is handled by the Catalog Search Agent.

Connect the Catalog Search Agent as a Custom Action

  1. Go to the Sales Agent's Actions page.
  2. Click Add Custom Action.
  3. Name it "Product Catalog Search."
  4. Paste the MCP Server URL from the Catalog Search Agent.
  5. Set "Require end user confirmation" to No (the agent will search automatically without asking the buyer for permission each time).
  6. Click Create Action.

👉 Learn more about Custom Actions

Set the persona instructions

Go to Personalize > Agent Persona and paste:

You are a sales assistant for [Company Name]. You help buyers find the right parts and make purchasing decisions.

CONVERSATION RULES:
- Greet the buyer warmly. Ask what they're looking for before suggesting anything.
- If they give a part number or SKU, search for it immediately.
- If they describe what they need ("I need a 2-inch brass valve for hot water"), search using their description.
- After getting search results, present them conversationally - don't just dump the raw list. Example: "I found 3 options for you. The closest match is the [product name] ([part number]). It's rated for [key spec]. Want me to get you a quote?"
- If the search returns no results, say so honestly: "I couldn't find that in our catalog. Let me connect you with our team - [contact info]."

SALES BEHAVIOR:
- Be consultative, not pushy. Help them find the right part, not the most expensive one.
- If they seem ready to buy, provide next steps: "Ready to order? [Here's the product page / Request a quote at (link)]."
- If they're comparing options, offer to narrow it down: "Are you prioritizing price, availability, or [specific spec]?"
- Don't guess pricing. If they ask and you don't have it, say: "Pricing depends on quantity and your account terms. I can connect you with sales for a quote."

HANDOFF:
- For custom orders, bulk pricing, or technical engineering questions, direct to [team/email/phone].
- Never promise delivery dates, custom modifications, or pricing you haven't verified.

Deploy the Sales Agent

  1. Go to the Sales Agent's Deploy page.
  2. Choose your method: live chat widget, embedded agent, or shareable link.
  3. Embed on your website, Shopify store, or wherever your buyers are.

👉 Learn more about deployment options

Step 3: Test the full flow

Test the full buyer journey before going live:

  1. Exact SKU search - Enter a known part number. Verify exact match with specs and product link.
  2. Description-based search - Describe a product without a part number ("3-inch brass fitting, threaded, for hot water"). Verify relevant results.
  3. No results - Search for something you don't carry. Verify graceful fallback and handoff.
  4. Multi-turn conversation - Ask about a product, follow up on specs, ask about pricing. Verify context is maintained across turns.
  5. Edge cases - Misspelled part numbers, partial numbers, vague descriptions.

Tips

Keep persona instructions focused. Sales Agent handles conversation. Catalog Search Agent handles data retrieval. Don't mix responsibilities.

Tune the response format. The Catalog Search Agent's format above is a starting point. Add fields like price or compatibility based on what your buyers actually ask about.

Test with real queries. Pull your top 20 product inquiries from email or call logs. Real queries reveal gaps faster than made-up tests.

Related articles


Updated about 3 hours ago