Building a Product Discovery Agent: AI-Powered Shopping Assistance and Recommendations

Why Product Discovery Matters in E-Commerce

The average online store carries thousands of SKUs. Customers who cannot find what they want within a few interactions leave — and most never return. Traditional keyword search misses intent. A customer typing "something warm for a mountain trip" gets zero results on a site that indexes by material and garment type.

An AI-powered product discovery agent bridges this gap by understanding natural language queries, mapping them to product attributes, and personalizing results based on browsing history, past purchases, and stated preferences.

Architecture of a Product Discovery Agent

The agent sits between the customer and your product catalog. It receives free-text queries, extracts structured intent, queries your catalog database, applies personalization, and returns ranked results. The core loop involves three tools: semantic search, attribute filtering, and comparison.

flowchart LR
    CALLER(["Shopper"])
    subgraph TEL["Telephony"]
        SIP["Twilio SIP and PSTN"]
    end
    subgraph BRAIN["E-commerce AI Agent"]
        STT["Streaming STT<br/>Deepgram or Whisper"]
        NLU{"Intent and<br/>Entity Extraction"}
        TOOLS["Tool Calls"]
        TTS["Streaming TTS<br/>ElevenLabs or Rime"]
    end
    subgraph DATA["Live Data Plane"]
        CRM[("CRM and Notes")]
        CAL[("Calendar and<br/>Schedule")]
        KB[("Knowledge Base<br/>and Policies")]
    end
    subgraph OUT["Outcomes"]
        O1(["Order status answered"])
        O2(["Return RMA created"])
        O3(["Specialist handoff"])
    end
    CALLER --> SIP --> STT --> NLU
    NLU -->|Lookup| TOOLS
    TOOLS <--> CRM
    TOOLS <--> CAL
    TOOLS <--> KB
    NLU --> TTS --> SIP --> CALLER
    NLU -->|Resolved| O1
    NLU -->|Schedule| O2
    NLU -->|Escalate| O3
    style CALLER fill:#f1f5f9,stroke:#64748b,color:#0f172a
    style NLU fill:#4f46e5,stroke:#4338ca,color:#fff
    style O1 fill:#059669,stroke:#047857,color:#fff
    style O2 fill:#0ea5e9,stroke:#0369a1,color:#fff
    style O3 fill:#f59e0b,stroke:#d97706,color:#1f2937

from agents import Agent, Runner, function_tool
from dataclasses import dataclass
from typing import Optional

@dataclass
class ProductResult:
    id: str
    name: str
    price: float
    category: str
    attributes: dict
    score: float

# Simulated catalog with embeddings
PRODUCT_CATALOG = [
    {"id": "SKU-001", "name": "Merino Wool Hiking Jacket",
     "price": 189.99, "category": "Outerwear",
     "attributes": {"warmth": "high", "activity": "hiking", "material": "merino"}},
    {"id": "SKU-002", "name": "Down Insulated Parka",
     "price": 249.99, "category": "Outerwear",
     "attributes": {"warmth": "extreme", "activity": "general", "material": "down"}},
    {"id": "SKU-003", "name": "Fleece Midlayer Pullover",
     "price": 79.99, "category": "Midlayers",
     "attributes": {"warmth": "medium", "activity": "hiking", "material": "fleece"}},
]

@function_tool
def search_products(query: str, max_results: int = 5) -> str:
    """Search the product catalog using natural language."""
    query_lower = query.lower()
    scored = []
    for product in PRODUCT_CATALOG:
        score = 0
        searchable = f"{product['name']} {product['category']} " \
                     f"{' '.join(product['attributes'].values())}".lower()
        for word in query_lower.split():
            if word in searchable:
                score += 1
        if score > 0:
            scored.append({**product, "relevance_score": score})
    scored.sort(key=lambda x: x["relevance_score"], reverse=True)
    return str(scored[:max_results])

@function_tool
def filter_by_price(min_price: float, max_price: float) -> str:
    """Filter products within a price range."""
    results = [p for p in PRODUCT_CATALOG
               if min_price <= p["price"] <= max_price]
    return str(results)

@function_tool
def compare_products(product_ids: str) -> str:
    """Compare products side by side. Accepts comma-separated IDs."""
    ids = [pid.strip() for pid in product_ids.split(",")]
    matches = [p for p in PRODUCT_CATALOG if p["id"] in ids]
    if len(matches) < 2:
        return "Need at least two valid product IDs to compare."
    comparison = []
    for p in matches:
        comparison.append({
            "name": p["name"],
            "price": p["price"],
            "attributes": p["attributes"],
        })
    return str(comparison)

Building the Discovery Agent

With the tools defined, wire them into an agent with instructions that guide the conversational flow.

discovery_agent = Agent(
    name="Product Discovery Assistant",
    instructions="""You are a helpful shopping assistant for an outdoor
    gear retailer. Your job is to help customers find the right products.

    Guidelines:
    - Ask clarifying questions when a query is vague
    - Always search the catalog before making recommendations
    - When showing results, highlight why each product matches the request
    - Offer to compare products when showing multiple options
    - Mention price range and key differentiators
    - Never invent products that are not in the catalog""",
    tools=[search_products, filter_by_price, compare_products],
)

result = Runner.run_sync(
    discovery_agent,
    "I need something warm for a hiking trip under 200 dollars",
)
print(result.final_output)

Adding Personalization

Personalization transforms generic search into a tailored experience. Store user preferences and purchase history, then inject them into the agent context.

def build_personalization_context(user_id: str) -> str:
    """Fetch user history and build context string."""
    # In production, query your user profile database
    user_profile = {
        "preferred_brands": ["Patagonia", "Arc'teryx"],
        "size": "M",
        "past_categories": ["Outerwear", "Footwear"],
        "price_sensitivity": "moderate",
        "climate": "cold",
    }
    return (
        f"Customer preferences: prefers {', '.join(user_profile['preferred_brands'])}. "
        f"Size {user_profile['size']}. Usually buys {', '.join(user_profile['past_categories'])}. "
        f"Price sensitivity: {user_profile['price_sensitivity']}. "
        f"Lives in a {user_profile['climate']} climate."
    )

# Inject personalization into the agent at runtime
personalized_agent = Agent(
    name="Personalized Discovery Assistant",
    instructions=f"""You are a shopping assistant with knowledge of
    this customer. {build_personalization_context("user-123")}
    Use these preferences to rank and explain recommendations.""",
    tools=[search_products, filter_by_price, compare_products],
)

Handling Edge Cases

Production discovery agents must handle empty results gracefully. When no products match, the agent should suggest broadening the search, offer related categories, or ask the customer to adjust their criteria rather than returning a blank response.

@function_tool
def suggest_alternatives(category: str) -> str:
    """Suggest alternative categories when no exact match is found."""
    alternatives_map = {
        "Outerwear": ["Midlayers", "Vests", "Rain Shells"],
        "Footwear": ["Socks", "Insoles", "Gaiters"],
        "Midlayers": ["Base Layers", "Outerwear", "Vests"],
    }
    suggestions = alternatives_map.get(category, ["Popular Items"])
    return f"No exact matches. Try browsing: {', '.join(suggestions)}"

FAQ

How do I integrate vector search for semantic product discovery?

Use an embedding model like OpenAI text-embedding-3-small to encode product descriptions into vectors. Store these in a vector database such as Pinecone, Weaviate, or pgvector. At query time, embed the customer query and perform a cosine similarity search against product vectors. This allows the agent to find "warm jacket for mountains" even if no product contains those exact words.

How should I handle product availability in real time?

Add an inventory check tool that queries your inventory management system before presenting results. Filter out-of-stock items by default but offer a "notify me when back in stock" option. Cache inventory status with a short TTL (30-60 seconds) to avoid hammering your inventory API on every search.

What metrics should I track for a product discovery agent?

Track search-to-cart conversion rate, zero-result rate, average number of agent turns before a product is selected, and click-through rate on recommended items. Compare these against your baseline keyword search. A well-built discovery agent should reduce zero-result rates by 40-60 percent and increase add-to-cart rates measurably.

---

#ProductDiscovery #ECommerceAI #RecommendationEngine #RetailAI #ShoppingAssistant #AgenticAI #LearnAI #AIEngineering