Prompt samples for Product Recommender
Learn how to interact with the Product Recommender agent using effective prompts.
What is it?
The Product Recommender agent analyzes a single customer in Reltio and recommends three products based on profile attributes, household and other relationships, recent interactions, and data source confidence. It uses a scoring model to weigh these signals and explains its reasoning with evidence pulled from the Reltio Intelligent Data Graph.
In a typical interaction, the agent retrieves the customer’s unified profile, relationships, and interactions, then ranks candidate products by probability and lists the key signals behind each recommendation. You can also ask follow-up questions about households or about how many data points were used to generate the recommendations.
For more information, see Product Recommender.
Recommend products for a specific customer
✅ Prompt: Recommend three products for customer ID abc123.
Why it works: This prompt gives the agent a clear, unambiguous target. It can immediately fetch the unified profile for abc123, analyze attributes, relationships, interactions, and contributing sources, and then return three ranked product recommendations with explanations.
✅ Prompt: Recommend products for customer ID xyz789, but only consider interactions from the last 60 days.
Why it works: You still anchor the request on a single customer ID but add a clear interaction window. The agent can adjust its interaction retrieval parameters (for example, overriding a default 90-day window) and focus on the most recent behavior when scoring recommendations.
✅ Prompt: Recommend financial products for customer ID def456.
Why it works: Adding a product category narrows the candidate set. The agent can filter products by category attributes before scoring them, reducing noise from unrelated product types.
Search for a customer, then recommend products
✅ Prompt: Find an active Individual customer in California and recommend products.
Why it works: This prompt asks the agent to search for a single customer using profile attributes it can query directly (entity type, status, and location). Once it finds a suitable Individual in California, it can retrieve that customer’s profile, relationships, and interactions and then run the standard recommendation workflow. If you want to limit recommendations by interaction recency, first identify the customer and then control the interaction window in a follow-up prompt.
✅ Prompt: Pick a premium-tier customer from the Individual entity type and recommend products based on their household relationships and recent browsing behavior.
Why it works: You specify entity type (Individual), segment (premium tier), and which signals to emphasize (household and browsing behavior). The agent can select a representative customer in that tier and weight graph and interaction signals more heavily in its scoring model.
Control which signals and evidence to emphasize
✅ Prompt: Recommend products for customer ID abc123, emphasizing profile fit over interaction history.
Why it works: Instead of disabling interactions entirely, you tell the agent how to weight signals. It can still use the full mix of attributes, relationships, and interactions, but tilt the scoring model toward static profile fit (for example, demographics, preferences, or status).
✅ Prompt: Recommend three products for customer abc123, then tell me how many data points you used — attributes, relationships, sources, and interactions.
Why it works: This prompt asks for both recommendations and transparency. After ranking products, the agent can summarize how many attributes, relationships, sources, and interactions contributed to the result, helping you judge whether the profile is well-populated or sparse.
✅ Prompt: Recommend three products for this customer based on their household relationships, and then suggest one additional product for the household as a whole.
Why it works: The prompt makes household context explicit and separates the individual recommendation from the household-level one. The agent can inspect relationships, ages, locations, and preferences across the household and then propose an additional product tailored to the group, as demonstrated in the Sarah Brennick scenario.
Avoid ambiguous or unsupported prompts
⚠️ Prompt: Recommend products for all customers in Texas.
Issue: The agent is designed to work with one customer at a time. Asking for recommendations for “all customers” implies looping over many records and aggregating results, which isn’t supported in a single session.
Better prompt: Find one active customer in Texas and recommend products.
⚠️ Prompt: Show me some good recommendations.
Issue: No customer is specified, and “good” is subjective. The agent would have to guess which customer to pick and which filters to apply.
Better prompt: Pick a customer who has interacted in the last 7 days and recommend products.
⚠️ Prompt: Recommend products for customer abc123 but ignore their interactions and relationships, just use their age.
Issue: This contradicts the agent’s multi-signal scoring model. Turning off core signals (relationships and interactions) undermines the design and may still result in the agent using them behind the scenes.
Better prompt: Recommend products for customer abc123, emphasizing profile fit over interaction history.
⚠️ Prompt: Recommend products based on the customer’s social media sentiment.
Issue: Many tenants don’t store social media sentiment attributes in Reltio. The agent must first check the data model, discover that sentiment isn’t available, and explain the limitation, which wastes a turn.
Better prompt: What customer attributes are available in this tenant? Then recommend products for customer abc123 using those.
⚠️ Prompt: Generate a CSV of product recommendations for 500 customers.
Issue: The agent focuses on one customer at a time and doesn’t generate export files. Bulk recommendation jobs require external orchestration or separate tooling.
Better prompt: Recommend products for customer abc123. After that, I’ll give you the next customer ID.
Advanced patterns
✅ Metadata discovery first
Prompt: What entity types, attributes, and relationship types exist in this tenant? Then recommend products for a sample customer.
Why it works: Starting with metadata discovery helps you understand which customer and product attributes are actually available. The agent can inspect the data model first and then pick a representative customer for an initial recommendation.
✅ Data point transparency
Prompt: Recommend products for customer abc123, then tell me how many data points you used — attributes, relationships, sources, and interactions.
Why it works: Asking for counts gives you insight into how rich the underlying data is. In demos, the agent has explained recommendations using dozens of attributes, relationships, and interactions across many sources, which helps build trust in the outcome.
✅ Iterative refinement
Prompt sequence:
- Step 1: Recommend three products for customer
abc123. - Step 2: Now re-run but exclude products they already own and increase the weight on household influence.
Why it works: You build on the previous result instead of starting over. The agent can reuse context from the first recommendation, adjust scoring weights, and refine the product list based on ownership and household relationships.
Best practices
- Treat the Product Recommender agent as a single-customer assistant: always anchor on one customer per interaction.
- Use precise identifiers, such as a customer entity ID, whenever possible to avoid ambiguity.
- When you don’t have an ID, describe clear search criteria (status, location, recency of interactions) so the agent can pick a representative customer.
- Control scope with time windows and product categories to keep recommendations relevant and manageable.
- Ask for evidence and data point counts when you need to understand or validate how the recommendations were generated.
- Avoid prompts that assume non-existent data, bulk processing, or export capabilities; instead, iterate one customer at a time.
- Use iterative refinement: start with a baseline recommendation, review the rationale, then adjust weights or filters in follow-up prompts.
- When you need to factor in recent interactions (for example, “last 60 days”), first select a specific customer using search on profile attributes, then ask the agent to apply an interaction window for that customer. Avoid prompts that require filtering the entire population by interaction recency until that capability is explicitly supported.