From scattered documents to a structured product catalogue.
One orchestrator. Seven specialist AI agents. Product data prepared with manual approval at every stage.
The business context.
Supplier information arrives in separate catalogues, price lists, brochures and supporting documents. Model codes, descriptions and technical specifications can be spread across different pages. Turning that material into consistent product records requires finding the right sources and understanding how the information fits together.
Our role.
We built a product-data workflow in which one orchestrator coordinates seven specialist agents. Each agent passes its findings and notes to the next. The workflow accepts an existing website's attribute matrix and category matrix as optional inputs. Every stage pauses for manual confirmation so the reviewer can inspect the results and resolve uncertain findings before work continues.
How we shaped the solution.
1. Analyse and classify the inputs
The first agent analyses incoming documents, classifies them and assigns ratings to guide the subsequent work.
2. Locate the product information
The second agent searches across all documents and records where products, descriptions and possible images appear, creating source notes for the next agents.
3. Extract models and valid variants
The third agent returns to the identified sources and extracts the models. When requested, it combines manufacturer model-code options into all valid product variants according to the catalogue's rules, rather than treating every possible combination as a real product.
4. Prepare short and full descriptions
The fourth agent revisits the same source pages and gathers the material needed to compose short descriptions and full product descriptions.
5. Extract attributes and assign categories
The fifth agent extracts technical attributes. If an existing website attribute matrix is supplied, it maps equivalent attributes only when fully confident and flags ambiguous matches for manual review. Without a matrix, it collects the attributes stated in the source documents. An optional website category matrix guides product categorisation.
6. Create SEO metadata
The sixth agent uses the assembled product information to prepare meta titles, meta descriptions, SEO page headings and other required SEO tags.
7. Check and assemble the final catalogue
The final agent checks the assembled product data and prepares the final output for review. Product records are assembled in Excel tables, with other output formats selectable. Manual confirmation remains required at each stage, including review of uncertain extractions and mappings.
What was delivered.
The workflow brings models and valid variants, short and full descriptions, attributes, categories and SEO metadata together in a reviewable product catalogue. Reviewers control progress between stages and inspect uncertainties before approving the final export.
A project of your own?
Bring your business context. We’ll help define the website and the next steps.