Skip to main content

​
RAG Example: E-commerce Product Assistant

ScaleDown Team • February 2025 • 10 min read Online shoppers ask specific questions: “Is this jacket waterproof?”, “Does this laptop have a USB-C port?”, “Which running shoes work for flat feet?” A product assistant powered by RAG can answer these from your catalog data, and you don’t need a vector database to build one. With ScaleDown, you feed your product catalog as context and the shopper’s question as prompt. ScaleDown compresses the catalog down to only the relevant product details before the LLM generates an answer.
This example builds a product Q&A assistant for an electronics store. It answers shopper questions from a product catalog using only ScaleDown and an LLM.

​
The Problem

Product catalogs have hundreds of items, each with specs, descriptions, pricing, and availability. When a shopper asks about one product or feature, the LLM doesn’t need to see every item in the catalog. Just the ones that match.

​
The Product Catalog

Here’s a sample electronics catalog with 6 products: 
product_catalog = """
## MacBook Pro 16" M3 Max
- Price: $3,499
- Processor: Apple M3 Max, 16-core CPU, 40-core GPU
- RAM: 48GB unified memory
- Storage: 1TB SSD
- Display: 16.2" Liquid Retina XDR, 3456x2234, 120Hz ProMotion
- Battery: Up to 22 hours
- Ports: 3x Thunderbolt 4 (USB-C), HDMI, SDXC, MagSafe 3, headphone jack
- Weight: 4.8 lbs
- Best for: Professional video editing, 3D rendering, software development

## Dell XPS 13 Plus
- Price: $1,299
- Processor: Intel Core Ultra 7 155H
- RAM: 16GB LPDDR5x
- Storage: 512GB SSD
- Display: 13.4" FHD+, 1920x1200, 60Hz, touch
- Battery: Up to 13 hours
- Ports: 2x Thunderbolt 4 (USB-C), no USB-A, no headphone jack
- Weight: 2.73 lbs
- Best for: Ultraportable daily driver, business travel

## Sony WH-1000XM5 Headphones
- Price: $349
- Type: Over-ear, wireless, noise-canceling
- Driver: 30mm carbon fiber composite
- Battery: 30 hours (ANC on), 40 hours (ANC off)
- Connectivity: Bluetooth 5.3, 3.5mm audio cable, USB-C charging
- ANC: Adaptive noise canceling with Auto NC Optimizer
- Weight: 8.8 oz
- Best for: Commuting, office work, frequent flyers

## Samsung Galaxy S24 Ultra
- Price: $1,299
- Processor: Snapdragon 8 Gen 3
- RAM: 12GB
- Storage: 256GB (expandable: no)
- Display: 6.8" Dynamic AMOLED 2X, 3120x1440, 120Hz
- Camera: 200MP main + 12MP ultrawide + 50MP 5x telephoto + 10MP 3x telephoto
- Battery: 5000mAh, 45W wired charging, 15W wireless
- S Pen: Included, built-in
- Best for: Photography, productivity, note-taking

## Apple AirPods Pro 2
- Price: $249
- Type: In-ear, wireless, noise-canceling
- Driver: Custom Apple H2 chip
- Battery: 6 hours (ANC on), 30 hours with case
- Connectivity: Bluetooth 5.3, USB-C charging case
- ANC: Active noise cancellation with Adaptive Transparency
- Features: Personalized Spatial Audio, conversation awareness, IP54 dust/water resistance
- Best for: iPhone users, workouts, daily commute

## LG C4 65" OLED TV
- Price: $1,799
- Display: 65" OLED evo, 3840x2160, 120Hz
- Processor: Alpha 9 Gen 7
- HDR: Dolby Vision, HDR10, HLG
- Audio: Dolby Atmos, 40W 2.2ch speakers
- Smart TV: webOS 24, built-in Alexa and Google Assistant
- Gaming: 4x HDMI 2.1, VRR, ALLM, G-Sync, FreeSync Premium
- Best for: Movie watching, console gaming, home theater
"""

​
Build the Product Assistant

1

Set up your clients

import requests
from openai import OpenAI

SCALEDOWN_URL = "https://api.scaledown.xyz/compress/raw/"
SCALEDOWN_HEADERS = {
    "x-api-key": "YOUR_SCALEDOWN_API_KEY",
    "Content-Type": "application/json"
}
openai_client = OpenAI()
2

Compress the catalog against the question

A shopper asks about USB-C ports. The catalog has 6 products. ScaleDown keeps only the ones with relevant port information.
shopper_question = "Which laptops have USB-C ports and how many?"

response = requests.post(
    SCALEDOWN_URL,
    headers=SCALEDOWN_HEADERS,
    json={
        "context": product_catalog,
        "prompt": shopper_question,
        "model": "gpt-4o",
        "scaledown": {"rate": "auto"}
    }
)

result = response.json()
compressed = result["compressed_prompt"]

print(f"Original: {result['original_prompt_tokens']} tokens")
print(f"Compressed: {result['compressed_prompt_tokens']} tokens")
The catalog has 6 products. The shopper asked about laptops with USB-C. ScaleDown kept the MacBook Pro (3x Thunderbolt 4 USB-C) and Dell XPS 13 (2x Thunderbolt 4 USB-C) details while compressing away the headphones, phone, earbuds, and TV since none of those are laptops.
3

Generate the answer

response = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {
            "role": "system",
            "content": (
                "You are a helpful electronics store assistant. "
                "Answer the customer's question using ONLY the provided "
                "product information. Be specific about specs and prices."
            )
        },
        {
            "role": "user",
            "content": f"Products:\n{compressed}\n\nQuestion: {shopper_question}"
        }
    ],
    temperature=0.2
)

print(response.choices[0].message.content)
Example output:
We have two laptops with USB-C ports:
  1. MacBook Pro 16” M3 Max ($3,499) - 3x Thunderbolt 4 (USB-C) ports, plus HDMI, SDXC, and MagSafe
  2. Dell XPS 13 Plus ($1,299) - 2x Thunderbolt 4 (USB-C) ports (note: no USB-A or headphone jack)
Specific port counts, accurate prices, and a useful note about the Dell’s limited port selection. All grounded in the catalog data.

​
Reusable Product Q&A Function

def ask_product(catalog: str, question: str) -> str:
    """Answer a shopper's question from the product catalog."""

    # Compress: keeps relevant products, drops the rest
    compress_resp = requests.post(
        SCALEDOWN_URL,
        headers=SCALEDOWN_HEADERS,
        json={
            "context": catalog,
            "prompt": question,
            "model": "gpt-4o",
            "scaledown": {"rate": "auto"}
        }
    )
    compressed = compress_resp.json()["compressed_prompt"]

    # Generate
    ai_resp = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "system",
                "content": (
                    "You are a helpful electronics store assistant. "
                    "Answer using ONLY the provided product information. "
                    "Include prices and key specs."
                )
            },
            {
                "role": "user",
                "content": f"Products:\n{compressed}\n\nQuestion: {question}"
            }
        ],
        temperature=0.2
    )
    return ai_resp.choices[0].message.content


# Example queries
print(ask_product(product_catalog, "What's the best option for noise canceling under $300?"))
print(ask_product(product_catalog, "Compare the two laptops for software development"))
print(ask_product(product_catalog, "Which products support Dolby Atmos?"))
print(ask_product(product_catalog, "I need a phone with a good camera, what do you have?"))

​
Why This Works for E-commerce

Catalog-wide search

Shoppers ask cross-product questions like “What’s your cheapest noise-canceling option?” ScaleDown surfaces the right products from the entire catalog.

Spec accuracy

Product specs need to be exact. Wrong prices or features break trust. Compression preserves numbers, specs, and prices while removing unrelated products.