Spreadsheets and CSV in a Chatbot Knowledge Base: How to Make Tabular Data Answerable

Businesses keep a great deal of knowledge in spreadsheets: price lists, product catalogues, class timetables, member lists, service areas, staff rotas. It is natural to upload the file to a chatbot and expect it to answer questions about it. Often it does not, or it answers wrongly. The reason is that retrieval systems are built for prose. A table of rows and columns, flattened into text, loses the link between a row's label and its values, and a question like 'how many members joined in March?' is a calculation, not a lookup. This guide explains why tabular data is awkward for retrieval, how to prepare a spreadsheet so that an assistant can answer questions about it, when to convert rows to sentences, how summaries help with counts and totals, what sizes to expect, what you should not upload and how to test. Crescelo reads CSV and similar files and adds a dataset overview to help with broad questions.

Why tables are hard for retrieval

Retrieval finds passages that resemble the question. A row of numbers resembles nothing a visitor would type.

Technique 1: turn each row into a sentence

Convert every row into a short sentence that includes the column names. The sentence is searchable, understandable and self-contained.

Crescelo's file parsing for CSV keeps column names with each row, so rows are searchable and readable. Where you can, tidy the header names first: 'Price (USD)' is better than 'P'.

Technique 2: add a summary block

Questions about counts, totals and breakdowns cannot be answered by retrieving a few rows. Add summary text that answers the common ones: how many rows, the main categories and their counts, ranges of prices or dates, and the most common values. Crescelo adds a dataset overview for uploaded tables, describing the columns and summarising categories, and prioritises it for questions about counts and overviews. You can improve on it by writing your own summary in a text document.

Technique 3: split big tables by topic

A single sheet with a thousand rows is unwieldy. Split it into meaningful parts: by category, by location or by date. Give each part a clear title. A file called 'Colour prices' retrieves better than one giant 'Prices' file with everything in it.

Technique 4: keep the source of truth elsewhere for calculations

An assistant is not a spreadsheet. For calculations that must be exact, such as totals, live stock or balances, use a system built for it, and let the assistant answer descriptive questions from the data. Do not rely on a language model to add columns of figures. If the answer is a simple aggregate, put the aggregate in the text.

What not to upload

• Personal data such as member lists with names, emails and phone numbers, unless you need the assistant to answer about individuals and have the legal basis.
• Financial records and anything confidential.
• Live data that changes hourly, such as stock levels, unless you can keep it updated.
• Huge exports of everything, when only a small part is useful.
• Files with merged cells, multiple header rows or notes in cells, without cleaning.

If a dataset contains personal information, tell visitors what the assistant uses and follow your privacy rules. See the GDPR checklist for website chatbots.

Prepare the file

1. Use one header row with clear names.
2. Remove blank rows and columns, merged cells and decorative formatting.
3. Put units in the header: 'Duration (minutes)', 'Price (USD)'.
4. Keep one table per sheet, or one topic per file.
5. Save as CSV (UTF-8) or a simple spreadsheet format.
6. Add a short text note describing what the file contains and its date.

Pruébalo

Fix problems by improving the file, not by adding rules. See how to stop a chatbot from making things up.

Worked example: a class timetable

A studio uploaded its timetable spreadsheet, and the assistant answered 'what time is Pilates on Tuesday?' wrongly. The sheet had merged cells for days, a blank column and abbreviated headers. The owner cleaned the sheet into one row per class with the columns 'Class', 'Day', 'Start time', 'Room' and 'Places', and added a short note: 'Timetable valid from 1 September. Classes marked Beginner are suitable for first-timers.' After re-uploading, the assistant answered correctly, including 'which classes are good for beginners?', because the note gave it the meaning. Ten minutes of cleaning fixed the problem.

Common mistakes

• Uploading a raw export with merged cells and abbreviations.
• Expecting the assistant to add up columns.
• Leaving stale versions of the same sheet in the knowledge base.
• Uploading personal data that is not needed.
• Skipping tests with real questions.

Frequently asked questions

See the knowledge base feature, try uploading a table in the playground and read preparing your service menu and pricing documents.

Further reading

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.): The original research paper behind retrieval-augmented generation.
• Google Search Central: structured data introduction: How structured data improves search results.