Dust agents should answer spreadsheet questions with Table Query SQL, not semantic search

TL;DR

Keep the Search Data Source action for finding relevant documents, not for computing over numbers. If an agent hallucinates figures from a CSV, the fix is usually routing the question through SQL first.

Steps

1. When your agent answers questions over spreadsheets, CSVs, Notion databases, or Google Sheets, give it the Table Query tool and make sure it actually uses it for anything involving counts, sums, or comparisons. Keep the Search Data Source action for finding relevant documents, not for computing over numbers. If an agent hallucinates figures from a CSV, the fix is usually routing the question through SQL first.

When to use

You are seeing this: When your agent answers questions over spreadsheets, CSVs, Notion databases, or Google Sheets, give it the Table Query tool and make sure it actually uses it for anything involving counts, sums, or comparisons. Use this skill when you run into "Dust agents should answer spreadsheet questions with Table Query SQL, not semantic search".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

The original report does not dig into a root cause. It documents the symptom and the fix that resolved it.

Source: Official Dust docs on LLM limitations document when search-based retrieval gives wrong answers on structured data. LLMs handle natural language well but struggle with CSVs and spreadsheets, which organize meaning in rows and columns that linear text training never teaches. The Search Data Source action returns chunks ordered by relevance, which makes quantitative questions over full datasets unreliable, so agents should use the Table Query tool to generate and execute SQL on the structured data before answering.

Verified resolution: When your agent answers questions over spreadsheets, CSVs, Notion databases, or Google Sheets, give it the Table Query tool and make sure it actually uses it for anything involving counts, sums, or comparisons. Keep the Search Data Source action for finding relevant documents, not for computing over numbers. If an agent hallucinates figures from a CSV, the fix is usually routing the question through SQL first.

Source URL: https://docs.dust.tt/docs/understanding-llm-limitations-counting-and-parsing-structured-data