Product data in the wild is a mess. The same running shoe appears on five retail sites with five different names, three different colour conventions, and specifications scattered across marketing copy, bullet lists and image alt text. If you're building price comparison, recommendations or a marketplace, turning that chaos into clean, structured records is the whole game.
We've been using Amazon Bedrock to do exactly that, and the approach has held up well at scale.
Why an LLM, and why Bedrock
Traditional scrapers break the moment a template changes, and rule-based extractors need a rule for every retailer. Large language models flipped that economics: instead of writing parsers per source, you describe the output you want — a schema — and let the model map messy input onto it.
Bedrock earns its place in the stack for three reasons:
- The data stays in our AWS account. No third-party API means no additional data processing agreements and no product catalogues leaving the VPC.
- Model choice without integration churn. We can benchmark models against the same extraction task and switch when the price-accuracy trade-off moves.
- It composes with the rest of AWS. S3 events in, Lambda for orchestration, SQS for backpressure, structured results out — boring, reliable plumbing.
The pattern that works
- Define a strict schema first. Field names, types, allowed values. The schema is the contract everything else depends on.
- Extract with the schema in the prompt and validate the response against it mechanically. Anything that fails validation gets retried or flagged — never silently accepted.
- Score confidence per field, not per document. A record with a confident price and an uncertain colour is still useful; treating documents as all-or-nothing throws away good data.
- Keep humans in the loop where it counts. Low-confidence fields route to a lightweight review queue instead of blocking the pipeline.
Where this went next
The pattern in this post — schema-first extraction, field-level confidence, human review queues — turned out to generalise far beyond product data. Invoices, contracts, ID documents, claims: same problem, same architecture. We eventually productised the whole pipeline as LangParse, our document intelligence platform, and it now powers the intelligent document processing work we do for clients.
If you're staring at a pile of unstructured data and wondering whether this approach fits, get in touch — we can usually tell you within a conversation.