semlayer reads your data warehouse and works out what an AI needs to know to answer business questions correctly. It shipped as a command-line tool, and people asked to see it work first. Now you can watch it map a warehouse, explore what it found, and compare right answers with wrong ones, in a browser tab.
semlayer is a tool that reads a data warehouse and writes down what the data means, so that an AI asked a business question gets the number right. We shipped it as a command-line tool. The first thing people asked was some version of the same question: can I see it in action before I install anything?
What it costs to try a data tool
Before a tool like ours shows you anything, it asks for three things.
An install. A Python environment on your machine, and at many companies, permission to put software on it.
A warehouse login. A read-only account and its credentials. That is a request to somebody else, and it waits in their queue.
An AI key. An account with a model provider, a budget behind it, and approval to send anything to it.
None of these is unreasonable. Together they turn “I’ll take a look” into a small project, and a small project waits for a free afternoon that does not come. If you are an analyst or a manager and not an engineer, it may not be something you can do at all.
There is a fourth cost, and it is the one people were really pointing at. When a command-line tool finishes, what you have is a text file. Everything it worked out about your warehouse is in there, but you find out by reading it. You cannot watch it think, click on a table, or ask it why.
People did not ask to read the output. They asked to see it in action.
So we removed the first three costs, and built the fourth thing: a page where you can see it.
First, why semlayer exists
The short version, for anyone arriving here first.
Companies are pointing AI at their data warehouses and asking it business questions. The AI writes a query, the query runs without an error, and the number that comes back is wrong. Not because the AI is weak, but because the rules that make a number right are in people’s heads. Cancelled orders never count. One of the two order tables stopped being updated years ago. Nothing in the database says so.
The fix is to write those rules down in one place the AI can read. That place is called a semantic layer, and until now a person had to write it by hand, which takes months. semlayer works the rules out from the data itself, in minutes.
The longer version, with the measurements, is here: https://medium.com/@rajdevnath/we-measured-what-a-semantic-layer-is-actually-worth-to-an-ai-agent-b5d7e62d9376
What you can see now
Open the page. There is nothing to sign up for.
To give you something to look at straight away, the page comes with a sample warehouse we call Messy Mart: a made-up retailer with 36 tables, built to be as untidy as a real warehouse that has been running for years. Each step below works the same way on your own data. The Messy Mart figures are there to show you what to expect.
Watch the map of your warehouse draw itself. A warehouse is a set of tables that relate to each other, and those relationships are almost never written down. semlayer finds them by reading the data. On the page you watch this happen: tables appear as they are read, and lines appear between them as each connection is found. On Messy Mart that is 36 tables and 37 connections, none of them declared anywhere, in about a minute.
Click any table to see what it connects to. Knowing a table exists is not the same as knowing how to use it. Click one and you see what it holds, which tables it joins to, and through which columns. On Messy Mart, clicking the orders table shows it leading to the order lines, the customers, the shops and the sales reps, among others.
The orders table, with the nine tables it joins to and the rule it requires.
See the metrics it proposed. A metric is a number your business reports, such as revenue or average order value. Teams usually build these up painstakingly, one definition at a time. semlayer gives you a starting list by surfacing the metrics that are already visible in your data: the amounts and counts in your detail tables, and the totals somebody has already built into your summary tables. They are proposals, each marked as inferred until a person confirms it. On Messy Mart it proposed 19 metrics from the 36 tables, among them total revenue, revenue from completed orders only, returns, and average order value.
Slice and dice any metric. A total on its own rarely answers the question. You want it by quarter, by shop, by channel. For each metric, the page shows the time periods and the dimensions you can break it down by, and you pick. Each combination you run gives you the answer and the query that produced it, so you can try one after another. On Messy Mart, revenue can be broken down by 65 different columns, and revenue by quarter comes back as twelve quarters.
Picking a metric from the list, grouping it by sales channel, and getting revenue for each of the four channels.
See what it used, and what it turned down. An answer you cannot inspect is an answer you cannot trust. So every answer shows its working: which tables it used, which rules it applied, and which tables it refused and why. On Messy Mart there are two orders tables. It refuses the one with the friendlier column names, because that table stopped being updated at the end of 2022.
An answer, with the table it used, the table it refused and the rule it applied on the right.
See the wrong answer next to the right one. The claim behind all of this is that a query can run perfectly and still be wrong. The page lets you test that. It shows queries an AI would plausibly write, beside the query semlayer would have it write, and you run both. On Messy Mart the plain revenue query returns $16.3M. The right figure is $14.6M. The gap is 818 cancelled orders, and the page shows the evidence it used to work out that they do not count.
Both queries run without an error. One is $1.7M too high.
Then turn to your own data. Drop in CSV, Parquet or JSON files, or a DuckDB file. You get the same map, the same proposed metrics and the same working, for your tables. When semlayer is unsure about something it asks you. When you are finished, download the result as one file.
Which AI it uses, including a free one
This was the third cost, so it is worth being specific. You have three choices.
None. Everything on Messy Mart works with no AI connected, including the metrics and the right-against-wrong queries.
A free one. To type questions in your own words, sign in with OpenRouter. It offers free models alongside paid ones, and the page lets you pick. The free ones cost nothing. OpenRouter limits how many requests they take, and one is sometimes busy; if so, pick another.
Your company’s. If your company runs its own approved AI service, the page can use that. It also takes your own Anthropic, OpenAI or Gemini key.
You can use one model to build and a different one to answer questions.
What stays in the tab
Your files are opened inside the browser tab and are not uploaded to us. There is no server of ours for them to go to.
If you connect an AI, the page talks to that provider directly from your browser. By default the AI sees column names and a few example values so it can describe them. One switch in the settings stops the example values.
When you are ready for the real warehouse
The tab works for small and medium warehouses, brought in as files.
For large ones we recommend running semlayer on your own machine, where it can connect to Snowflake, BigQuery and the rest directly and use all the power the machine has. We built a VS Code extension for exactly that, and it also works in Cursor. The file you downloaded from the tab opens there, so nothing you did is thrown away.
We will follow this with a longer article on the extension.
Until then: open the page, press the button, and see whether what it finds looks like your warehouse.
Try it: https://try.semlayer.dev
The VS Code extension: https://marketplace.visualstudio.com/items?itemName=semlayer.semlayer
What semlayer is: https://semlayer.dev





Top comments (0)