Known issues

These are the known issues in this release.

Platform

When querying multiple documents in Converse:
- With the multistep model, the information source is accurately returned only at a page level.
- Only the multistep model can return visualizations such as graphs and charts.
- Single specific tables can’t be extracted.
Table extraction is in public preview, and Converse can’t yet reliably do the following:
- Correctly extract a single table that’s continued on a different page, unless the table header is present on the subsequent page. If there is no consistent table header across all pages, the tables on each page are detected separately.
- Select a table from an Excel sheet that contains multiple tables on one sheet. If a single Excel sheet contains more than one table, Converse can extract only the top left table.
- Select a table from a document with many (approximately 30 or more) tables in a single document.
- Use non-English language queries for table extraction from non-English documents.
- Respond to prompts about tables that aren’t found by Converse’s object detection.
- List the exact source of information in a table. Instead, Converse cites the entire table as the information source.
When you publish a new chatbot version, including updating any of the chatbot’s details or settings, all users of the chatbot lose their chat history and are redirected to the chatbot’s latest version.

With an extraction prompt, table extraction results can be somewhat unreliable, particularly if you enabled object detection in your project settings. If you’re not seeing expected results, switch to a reasoning prompt or disable object detection.