Stored Datasets & Embeddings

Everything you want an assistant to use (product descriptions, uploaded documents, entire knowledge bases) goes through an embedding process first. Embeddings turn your text into something the model can actually search through, so the assistant can read across all of it and surface what's relevant, instead of being limited to whatever fits in a single prompt.

Two kinds of search work together. Semantic search lets customers ask in their own words and still get the right result from unstructured text. Structured, parameterized search narrows by exact attributes (category, brand, or vehicle fitment, for example), so the assistant can combine what the customer meant with what exactly matches.

For more complex catalogs we build a parser tuned to that business's data, pulling structured fields out of messy product information so a query like "parts that fit a 2018 model" resolves to exactly the right items. Depending on the data, that extraction is rule-based or AI-driven, whichever gives the most reliable result.

Larger customers often need search where a wrong answer simply isn't acceptable. For them, we mirror the catalog into our own database and run an AI extraction step through our ETL pipeline as the data syncs, cleaning and enriching it so the assistant can query it quickly and precisely. The mirror keeps in step with the source, so what the assistant sees stays current.