A semantic layer turns vague column names into business meaning, so AI answers stay accurate. What a business dictionary is and how to build one.
Ask an AI tool "what was our revenue last month?" and it has to guess which column holds revenue. Real spreadsheets and databases are full of names like `amt`, `net_val`, `rev2`, or `status_3` — names that mean something to the person who built the sheet and nothing to anyone else.
When the meaning is ambiguous, the AI guesses. Sometimes it guesses right. Sometimes it sums the gross amount instead of the net, or counts cancelled orders as sales. The answer looks confident either way — which is exactly the problem. The fix is not a smarter model; it is giving the model your business context.
A business dictionary — also called a semantic layer — is a small set of definitions that map your raw columns to business meaning. For each important field you record:
- The plain-English name — that `net_val` means "Net Revenue".
- The formula or rule — that revenue counts only rows where status is "completed".
- Synonyms — that "sales", "turnover", and "revenue" all mean the same metric.
Once these definitions exist, every question routes through the same shared meaning. "Revenue", "sales", and "how much did we make" all resolve to the identical calculation, instead of the AI re-deciding each time.
You do not need to define everything. Start with the handful of things people actually ask about and the columns most likely to be misread:
1. Your core metrics — revenue, active customers, orders, churn. Define exactly how each is calculated.
2. Status and category columns — what each code or label means ("DC" is Day Care, status 2 is refunded).
3. Date columns stored as text — note the format so the AI parses them correctly instead of returning empty results.
4. ID-like columns to ignore — flag the columns that look numeric but should never be summed or averaged.
A dozen good definitions usually covers most of the questions a team asks.
VedaX has a built-in Business Dictionary for exactly this. You can add definitions by hand, or use Auto-Suggest from Schema to have VedaX scan your connected tables and propose definitions for the meaningful columns — revenue metrics, key counts, status-dependent aggregations — which you then accept or edit.
Because the dictionary is injected into the AI prompt as trusted business context, every query the AI writes uses your definitions. One person defines "active customer" once, and the whole team gets the same answer to "how many active customers do we have?" forever after.
The biggest win is not speed — it is consistency. Without a semantic layer, two people asking the same question can get two different numbers because the AI made two different assumptions. With one, "revenue" means revenue, every time, for everyone.
That consistency is what lets a team actually trust AI analytics enough to make decisions on it. The model handles the SQL; your business dictionary makes sure it is answering the right question.
It is a set of business definitions that sit between your raw data and your analytics, mapping technical column names to plain-English meaning and formulas. It is how "net_val where status=completed" becomes simply "Revenue" for everyone who asks.
No. Start with your core metrics, your status/category codes, and any date columns stored as text. A dozen well-chosen definitions usually covers the majority of real questions; you can add more as new questions come up.
It removes guesswork. Instead of the AI inferring what "revenue" means from a column name on every query, it uses your stored definition and formula. That makes answers both correct and consistent across users and over time.
See it in VedaX: AI business intelligence for small businesses