How to work with tables

Master working with tables in AI Chatbot Hub: learn tips and techniques for efficient table handling.

AI Chatbot Hub enables you to upload training data in the form of basic CSV or Excel tables. Basic tables are defined as those that have a single primary key column with unique entries, no merged cells, descriptive titles for rows and columns, and a consistent structure with either a column-based or row-and-column setup.

Here’s an example of a column-based table:

Each row represents a single product key found in the first column.

And for a row-and-column setup:

Every cell value corresponds to a unique row-column pairing.

While there’s no fixed limit on the number of columns you can include, each row has a maximum token limit, beyond which an error may occur. As of March 2024, this limit is approximately 8,000 tokens per row. Note that this count includes non-data JSON used for table structure, so the space available for actual data is somewhat less and can vary based on the length of row and column names.

Once your data meets these criteria, you can upload your table as a training source on AI Chatbot Hub by navigating to Sources -> Add Sources -> Tables:

For row-and-column tables, upload the file, if it's the first time you will see this message below with a link highlighting Microsoft's guidelines.

See here - Transpose (rotate) data from rows to columns or vice versa

Proceed with the upload. Once the table has been uploaded and trained, you need to make sure the formatting is correct. Click on the table in the "Sources" menu, then on "View content". In the table edit menu, select "Row-Column Header" and hit "Save Changes" to ensure proper data formatting for large language models (LLM). View below:

Your tables are now ready to be referenced within knowledge bases for your agents.

If you want to further optimize data retrieval and display, you can explore the following, all within the Sources tab:

• Adding tags to each source

• Customizing the citation reference

• Customizing the citation link

In our experience, the GPT-4 series provides enhanced retrieval accuracy and consistency compared to the GPT-3.5 series when working with table data.