Chatwize allows you to upload training data in the form of simple CSV or Excel tables. Simple tables refer to tabulated data that do not contain merged cells, have a single primary key column with unique elements, unique and informative column / row titles, and either column-wise or row-column structural configurations.

Example column-only simple table:

Data across each row corresponds to a single product key in the first column

Example row / column simple table:

Each row-column combination points to a corresponding cell value.

While there is no maximum number of columns, there is a limit to the total number of tokens that can be included within each row of the table. If you exceed that, the training will result in an error. As of March 2024, the maximum number of tokens per row is ~8000 tokens (this number includes non-cell-value JSON code used to represent table structure, so actual maximum reserved for cell-values is lower and depends on column / row name lengths).

If your data matches the format above, you can proceed to upload your table onto Chatwize as a training source. To do so, simply go to Sources -> Add Sources -> Tables:

For row-column tables, after you upload it, select it and go to the three-dot-menu on the rightmost edge and click “Edit Table Data”:

Then, in the table edit dialog, select “Row-Column Header” and click “Save Changes”. This ensures that your data is pre-processed correctly for LLM understanding.

You can then set the tables as referenced data when assigning knowledge base to your Agents.

Please note that based on our experience, GPT-4 series significantly outperform GPT-3.5 series in retrieval accuracy and consistency when working with tables.

Example LLM retrieval using the sample tables above.

​
My Agent or chatbot doesn’t seem to understand or interpret my table correctly!

Large language models (LLMs) like GPT-4 are extremely adept at working with unstructured text data. With a multimodal training method, LLMs can even interpret visual or image data and associate it with natural language (like GPT-Vision). However, tabulated data is very different. There is no universal rule set or standard language patterns when it comes to representing structured information. Given the probabilistic nature of LLMs, they are not innately proficient at handling this type of data directly.

An article from Microsoft Research recently evaluated GPT-4’s performance when processing structured data. Within it, model performance against numerous programmatic representations of structured data are discussed.

Comparing GPT-3.5 vs. GPT-4 performance across a variety of table operations and structured data representations. Source: https://www.microsoft.com/en-us/research/blog/improving-llm-understanding-of-structured-data-and-exploring-advanced-prompting-methods/

Chatwize currently uses a version of JSON to handle tables. While not perfect, it does support a limited range of use cases when bot makers want to train chatbots with structured data.

We understand that many use cases involve much larger and more complex datasets that may even be updated in real-time. The static table feature within Chatwize is not optimized for them, and we instead recommend that you look into function-calling.

The most robust and “legitimate” way to do retrieval augmented generation (RAG) on structured data is via function calling, where you design and host custom functions with templated SQL queries to retrieve specific data snippets and feed it to your Chatwize chatbot as additional RAG context on-demand. However, this requires a bit of programming and server hosting setup on your end. We are working on some concrete examples of this setup, so we appreciate your patience as we improve our documentation.

Bring Your Own Key (BYOK) and PricingBest practices for preparing training data