The AI That Turns a Raw Product Feed Into a Revenue-Ready Catalogue, in Seconds.

The taxonomy model uses a combination of fine-tuned category classifiers and general embedding-based similarity matching. For well-established categories, apparel, electronics, furniture, beauty, the model performs at high confidence from day one. For niche or proprietary categories, the system falls back to embedding similarity against the target taxonomy structure, surfaces low-confidence matches for human review, and learns from every correction made in the Audit Dashboard. In practice, most client catalogues have 80–90% of their SKUs in standard categories where the model performs with full confidence; the remaining 10–20% are flagged and reviewed once, after which the model generalizes those corrections to future similar products. No new category requires a full retraining cycle.

Thin input is the most common scenario, and the pipeline is designed for it. When text input is minimal, the extraction pipeline attempts to use any available image data to infer attributes: product type from visual classification, color from image analysis, and material or construction from texture recognition. For products where both text and image data are insufficient to extract attributes with confidence, the system generates the best available classification with a low confidence score and routes the record to the Audit Dashboard for admin review rather than publishing an incomplete listing. The platform tracks "data completeness" per SKU so operations teams can see exactly where their supplier feeds are weakest and request improvements at source, which over time raises the quality of input data across the entire pipeline.

The storytelling engine is grounded-generation only: it generates content exclusively from the structured attributes that have already been extracted and validated by the upstream pipeline. It does not invent claims. If the composition of a fabric hasn't been extracted, the generated description won't mention composition. If the power rating of an appliance is unknown, the copy won't cite a wattage. The generation model is given only the confirmed attribute set as its factual basis, every sentence it writes must be traceable to an input attribute. This is a deliberate architectural choice: the pipeline treats extraction and generation as separate stages so that content quality is bounded by data quality, not by model confidence. The Confidence Score layer adds a final gate: descriptions that score below threshold are held back from publishing until reviewed.

Catalogue migration is one of the highest-value use cases for Naratix. Legacy PIMs accumulate years of inconsistent data, products added by different teams, under different category conventions, with varying attribute completeness. Moving that data to a new system without cleaning it first just transfers the problem. The Naratix pipeline can ingest the full legacy catalogue export, run taxonomy audit and attribute extraction across the entire dataset, surface inconsistencies and gaps, and produce a clean, normalized catalogue ready for import into the target PIM. What would otherwise be a six-to-twelve month manual data remediation project can be completed in days. The Admin Dashboard provides a complete audit trail of every change made to the data, giving migration teams the governance documentation their compliance processes require.

Most clients are processing live product data within one to two weeks of onboarding. The client-side requirement is straightforward: a product data export in any standard format (CSV, JSON, XML, or direct API connection to their existing ERP or PIM) and access to their target taxonomy structure, either a standard schema like Google Product Taxonomy or their own internal category hierarchy. Agix handles the pipeline configuration, model calibration, and marketplace integration setup. There is no code deployment required on the client side for standard integrations. For custom ERP connections or proprietary data formats, the onboarding typically extends to three to four weeks depending on how complex the source data structure is. The rate-limiting factor is almost always data access, clients who can export their full catalogue on day one move fastest.