Early on, one of our testers called PowerDataChat "a chart generator". It was a fair description. It could turn a question into a correct chart, but it did not know anything about the business behind the data.
What changed it was context, not a bigger model.
Data descriptions. A short description of the file and each column (which the AI can draft and a person confirms) tells the system that amt_out is the outstanding loan amount in GEL, not a random number.
Domain skills. In the Enterprise edition, skills carry industry vocabulary, KPI definitions and calculation rules. The banking and microfinance skill knows PAR30, NPL and IFRS 9 stages, and rules like "never sum across currencies" and "weight averages by outstanding amount".
Company vocabulary. What "branch", "client" or "portfolio" mean in this company, which test accounts to exclude, the reporting currency, which personal fields are never displayed.
With that context, the same question produces analysis instead of just a chart: the right filters, the right weighting, the right caveats.
If you are building with LLMs on structured data, our takeaway is simple: invest in describing the data before tuning prompts.
Top comments (0)