When you read about good infrastructure and what's needed for fast and accurate data prompting it can quickly become a daunting task especially for small companies with limited in house data/tech capabilities. A well structured DWH, a good semantics layer and potentially an external tool with a considerable price tag. I will talk about how we took a slightly different route at De Koffiejongens to get the whole business prompting data with Claude and Google BigQuery.

In short we found a way that we could setup for the basics within a week and worked for us by:
- limiting the used data to already well defined views on frequently used properties (think an exhaustive orders table)
- used LLMs to extract company knowledge from communication
- used this as a base for a makeshift semantics layer in Claude with layers of skills within a plugin
- manage data access at the user level in order to keep privacy standards
- implemented scheduled accuracy testing to maintain standards

During the presentation I will talk about the setup and different considerations we made along the way that shaped our decision.