Catalogs: Not Tables, but a Knowledge Base for AI
You have an expense table. Transactions, amounts, categories — something like that. It lives in Google Sheets or a Notion database, and it works fine for you. Then you want AI to answer a question about that data — "how much did I spend on food this month?" — and the hassle begins.
The problem: AI can't see your table
A familiar scenario. You open Claude, and to get any answer about your expense table you have to:
- Export the data to CSV or copy a chunk of it;
- Paste it into the chat;
- Explain what each column means — because
sumandcatsay nothing on their own; - Ask your question;
- Repeat everything tomorrow because there are new entries.
AI parses the grid of cells from scratch every time. Sheets and CSV carry no structure: to the machine it's text separated by commas. Which column is a number, which is a date, which is a link to another table — it has to guess. Usually from the contents, and sometimes wrong.
The root of the problem isn't AI. The root is that a table is a format for human eyes, not an interface for a program.
What a catalog is
A catalog in Korfix looks like a table — rows, columns, familiar layout. But under the hood it's different. A catalog has four things from the start that a Sheets tab doesn't:
- Field types — each field knows what it is: text, number, date, select, relation;
- Relations — a field can reference a record in another catalog;
- API — the catalog is available via REST right after creation;
- MCP — AI reads it directly, without exports.
A catalog isn't a "table for human eyes". It's a structured store designed so that both people and programs — including AI — understand it the same way.
Creating a catalog — 30 seconds
Click + Catalog, give it a name — say, "Transactions" — and add fields, choosing a type for each. That's it, the catalog exists and is working: available via API, visible to AI, ready for records.
No server setup, no database schema, no migrations. The field type you pick with your mouse is the schema.
In practice, catalogs are rarely created by hand. A marketplace app deploys the needed catalogs itself on install, and an app built for you through AI brings its own structure. Manual creation is just the clearest way to see what a catalog is made of.
Field types — why they matter
A field type isn't cosmetic alignment. It's the meaning you assign to data once, and everyone uses it: the interface, the API, and AI.
| Type | What it gives | In our expense tracker |
|---|---|---|
| text | arbitrary string | transaction comment |
| number | can be summed, averaged, compared | transaction amount |
| date | range filtering, "last 30 days" | transaction date |
| select | fixed set of values | type: expense / income / transfer |
| relation | link to a record in another catalog | wallet and expense category |
Compare two worlds. In CSV, the amount column is the text "1200", and AI has to guess that's money, not an index. In a catalog, the "amount" field is declared as a number — it can be summed without caveats. The "category" column in CSV is a string where today it's "food", tomorrow "Food", the day after "groceries". A relation field pointing to the Expense Categories catalog has exactly the set of values you created, and "food" can only be spelled one way.
When data is typed, AI doesn't guess the structure — it reads it.
Relations: data stops being flat
Real data doesn't live in one table. An expense isn't just a row: each transaction has a wallet the money left from and a category it went to. In Sheets this is separate sheets or duplicate columns that you mentally join by matching names.
In a catalog a relation is a field. The Transactions catalog has a "wallet" field that points to a record in the Wallets catalog, and a "category" field pointing to Expense Categories. The platform knows about these relations, and AI knows too. "How much did I spend from my card?" doesn't require you to manually join two sheets — the relation is already in the data.
AI reads the catalog as its own context
This is what it's all for. You connect the catalog to your AI client via MCP — one token from settings — and never export anything again. No CSV, no copy-paste, no explaining column names.
AI sees the catalog itself: its fields, their types, value sets, relations to other catalogs. That's the context. You ask in plain language — AI answers from facts.
how much did I spend on food this month
AI knows "date" is a date, "category" is a relation, "amount" is a number. It filters transactions by category and month, sums the amounts, replies. You didn't explain where what lives. The field types did it for you.
what was over a thousand
AI filters by the numeric field — no new setup, same catalog.
which wallet do I spend most from
AI follows the relation from Transactions to Wallets, groups, counts. Multiple catalogs, one question, zero manual work.
Add transactions tomorrow — AI sees them immediately. The catalog is connected not to a snapshot frozen in time, but to live data.
Practical case: expense tracking
Let's build in five minutes what usually lives across three browser tabs.
Three catalogs:
- Wallets — cash, card, savings.
- Expense Categories — food, transport, subscriptions, misc.
- Transactions — fields below.
| Transactions field | Type |
|---|---|
amount | number — transaction amount |
wallet | relation → Wallets |
category | relation → Expense Categories |
date | date |
Add transactions. Connect MCP. Then work with questions, not formulas:
- "sum of expenses by category for the month" — aggregate on a number, grouped by relation;
- "what was over $100" — filter on a numeric field;
- "balance per wallet" — relation plus aggregate.
In Sheets each of those questions is a separate formula or pivot table. In a catalog — one sentence, because AI relies on types and relations, not your memory of what's in which column.
Importing data: CSV and Sheets
You don't have to start from a blank catalog. Korfix imports CSV — and everything exports CSV: Google Sheets, Notion, Airtable, Excel.
Simple steps:
- Export your expense table to CSV from your current tool;
- Create a catalog in Korfix and upload the file;
- Map columns to fields and assign types — this is the key step that turns a flat file into a catalog;
- Done: same data, now with structure and AI access.
Step three is what matters most. Importing CSV without assigning types would just give you another table. Assigning types is the moment the table becomes a knowledge base for AI.
The bottom line
A table answers "what does this look like for me?". A catalog also answers "what does a program understand about this?"
- Field types give data meaning that doesn't need to be re-explained.
- Relations make data three-dimensional instead of flat.
- API and MCP open it to AI without exports or parsing.
While your expenses are in Sheets, AI sees a grid of text. Once they're in a catalog — AI sees structure. The difference between "parse this table" and "answer from my data".
Move your data and see. Registration at vibe.korfix.info is free and requires no credit card. Export any working table to CSV, load it into a catalog, assign types — and ask AI the first question. It'll notice the difference.
Back to blog
