I Built a Pipeline That Turns Messy Supplier Data Into On-Brand Product Listings — Automatically

The Problem

tb-0092, STAINLESS STEEL WATER BOTTLE 750ML, kitchen, Silver,
stainless steel, 750ml, $24.95, , BPA free

TB0094, Linen Throw Cushion, Home, Charcoal, 100% linen,
45cm x 45cm, 34.5, 12, Cover only no insert

TB-0098, Wool Throw Blanket, HOME, Grey Marle, 80% wool 20% acrylic,
130cm x 170cm, 89.0, 22, "Made in Australia, supplier states ""hypoallergenic"""

What It Does

Drop a supplier CSV into an S3 bucket. Within three minutes, clean draft listings appear in Shopify — grounded only in verified product data, written in a consistent brand voice, with unverified claims filtered out before anything goes live.

• Inconsistent formatting and casing normalisation
• Missing prices, missing stock quantities, duplicate SKUs
• Supplier claims that can't be published ("hypoallergenic", "100% natural fragrance")
• Products that aren't ready to list — flagged and held for review
• On-brand copy that follows your voice guide and learns from gold standard examples

Everything lands as a draft. A human approves before publishing. That's intentional.

Architecture

How the Prompts Work

Extraction — structured output with XML tags and JSON fencing

The extraction prompt uses a pattern that consistently produces valid, parseable JSON. XML tags separate the instructions from the data, and a fenced example makes the expected output format unambiguous.

EXTRACTION_PROMPT = """
You are a product data normalisation engine.

<instructions>
- sku: uppercase with hyphen (e.g. TB-0091)
- product_name: Title Case, preserve scent/variant names
- dimensions: if product name contains volume (350ml, 750ml),
  extract it here when dimensions is otherwise empty
- flags: array of out_of_stock, price_missing, qty_missing,
         possible_duplicate, unverified_claim
</instructions>

<example>
Input: SKU: tb-0092, Item Name: STAINLESS STEEL WATER BOTTLE 750ML...
Output:
```json
{ "sku": "TB-0092", "product_name": "Stainless Steel Water Bottle",
  "dimensions": "750 ml", "price": 24.95, "flags": ["qty_missing"] }
```
</example>

<raw_row>{row_text}</raw_row>

Respond with ```json ... ``` only.
"""

Generation — RAG-grounded retrieval

Before generating each listing, the pipeline queries the Knowledge Base. The retrieved chunks — brand voice rules and example listings — become the brand context in the generation prompt, so the model has both the product data and the style guide before it writes a single word.

def retrieve_kb_context(query):
    resp = agent_runtime.retrieve(
        knowledgeBaseId=KB_ID,
        retrievalQuery={'text': query},
        retrievalConfiguration={
            'managedSearchConfiguration': {'numberOfResults': 3}
        }
    )
    return '\n\n---\n\n'.join(
        r['content']['text'] for r in resp['retrievalResults']
    )

The Guardrail — and Why It Runs Output-Only

def generate_listing_guarded(product):
    # Step 1: generate without guardrail
    listing_text = generate(product)

    # Step 2: validate only the output
    guard = bedrock.apply_guardrail(
        guardrailIdentifier = GUARDRAIL_ID,
        guardrailVersion    = GUARDRAIL_VERSION,
        source              = 'OUTPUT',          # <- key
        content             = [{'text': {'text': listing_text}}]
    )

    if guard['action'] == 'GUARDRAIL_INTERVENED':
        return {'sku': product['sku'], 'blocked': True}

    return json.loads(listing_text)

Results

Test run: 20 products from a messy supplier CSV

Evaluation: 10-product scored set

Scoring axes: factual grounding (no invented claims) · brand tone (warm, direct, no superlatives) · completeness (all fields present and informative).

What Improved v1 → v2

The first eval run averaged 7.7/9. Four issues, four fixes, taking it to 8.5.

1. 01

   Capacity missing from titles

   "Ceramic Mug 350ml" was extracted as "Ceramic Mug" with dimensions: null.

   Fix: Updated the extraction prompt to parse volume and capacity from product names.

2. 02

   Spec-sheet bullet format

   Sparse products defaulted to "Material: Bamboo" / "Category: Kitchen" labels.

   Fix: Added explicit WRONG/RIGHT examples to the generation prompt and gold standard docs.

3. 03

   Scent names dropped

   "Soy Wax Candle - Sandalwood" became "Soy Wax Candle" during extraction.

   Fix: Updated the extraction prompt to preserve scent and variant names.

4. 04

   Minor hallucination

   "Gets softer with every wash" appeared in a listing — not in the supplier data.

   Fix: Added a "Do not add general category knowledge" rule to the generation instructions.

Open Questions

What's Left for Production

This is a personal reference implementation — not a production system. The gap is about 3.5 days of work:

• Lambda + S3 trigger — replace the notebook with an event-driven function
• Idempotency — check if a SKU exists in Shopify before creating
• AWS Secrets Manager — move credentials out of .env
• IAM role — replace long-lived access keys with an execution role
• Guardrail versioning — promote from DRAFT to a numbered version
• Run-level audit log — write a JSON summary per run to S3
• Error alerting — CloudWatch + SNS on failure or block-rate spike