Grounding: why your AI stops making up your numbers.
Your AI makes up numbers because it doesn't have your data. It writes the most likely-sounding answer, not the true one. Grounding fixes that by connecting your AI to your real numbers. Loam runs SQL against your data and returns the result to your AI.
Ask your AI about last quarter without connecting your data and it may say "revenue grew 12%" even when it grew 8%. The sentence is fluent, the number is invented. Grounding swaps the guess for a real answer: Loam runs SQL against your data and hands the result back. In Loam chat, you can inspect the returned result rows.
See grounding in action
The same question, answered with and without grounding. Watch what happens to the numbers when your AI can actually read your data.
Question:
What were total sales by region in Q4?
In Q4, the West region led with $310,000 in total sales, followed by East at $295,000, South at $260,000, and North at $215,000. Overall Q4 sales totaled approximately $1.08M.
⚠ Numbers are plausible but fabricated - no source data referenced
In Q4, East led with $312,000 in total sales, followed by West at $284,000, South at $245,000, and North at $198,000. Total Q4 sales were $1,039,000.
↳ Row 2: East, $312,000
↳ Row 1: West, $284,000
↳ Row 4: South, $245,000
↳ Row 3: North, $198,000
Why wrong numbers are hard to spot
Most warnings about AI hallucination point at factual slips: a wrong date, a made-up citation, the wrong capital city. Those are bad, but they're usually easy to spot. "Paris is the capital of Germany" jumps off the page.
A wrong number doesn't. "Q3 revenue was $2.1M" when it was really $1.9M reads exactly as well as the truth. There's no tell in the sentence. The sentence alone cannot show whether the number is right. Grounding ties the answer to your source data so you can check it.
How Loam grounds your AI
Upload a CSV, Excel, Parquet, TSV, or JSON file. Loam reads its rows and columns. When you ask a question, Loam runs a query against those values and shows it with the answer. Your AI reports what the query found instead of writing a number that merely sounds right.
- Upload your data. Upload a CSV, Excel, Parquet, TSV, or JSON file. Loam reads your rows and columns automatically.
- Ask in plain English. "What was revenue last quarter?" or "Which product churns the most?" No SQL required.
- Inspect the returned result rows. Loam chat shows the returned result rows so you can inspect them. Connected AI clients control whether the SQL itself is shown.
Grounding vs. the alternatives
RAG (Retrieval-Augmented Generation)
Gives your AI relevant text from policies or documents. Numerical questions need a calculation over the source values.
Grounding in your structured data
Connects your AI to the rows and columns in your files and calculates the answer with a query you can see. This is Loam's approach for questions whose answer is a number.
Fine-tuning
Teaches your AI your style and terminology. Training examples do not give your AI the current numbers in your files.
Knowledge graphs
Describe how entities relate to each other. They suit relationship questions more than counting or summing values.
Want the practical version? Read the ways to ground your AI in real data. To see grounding settle a real argument, compare a grounded answer against ChatGPT on the same CSV, or run a number through the hallucination checker to see whether your AI's figure holds up against the data. Preparing files for another tool? Use Loam to convert CSV to Parquet or convert Parquet to CSV.
Frequently asked questions
Chat with a CSV or TSV
Add a CSV or TSV. The tool sends up to the first 200 lines after the header to an AI service and asks it to cite supporting lines.