Release Notes v0.4

We’ve been working behind the scenes to bring smarter, intent-driven knowledge retrieval across DataQI assistants. This update introduces SQL connections, intent detection, knowledge mode, and hallucination guardrails, along with several usability improvements.

Features

1. SQL Connections (Beta) - Add databases as a data source

You can now connect directly to Microsoft SQL, PostgreSQL, and MySQL databases, and use them as data sources for your assistants.

What you’ll see:

Option to add a SQL connection from the Connections page
Ability to browse and select tables when adding SQL data in the Data page
SQL-informed assistants generate graphs and tables directly in chat, which display seamlessly and persist in history
Initial support for tables, bar charts, line graphs and pie/donut charts

⚠️ Note: This feature is in beta and we welcome your feedback.

Known Issues (Beta):

Performance may be slower when working with very large datasets
Responses may occasionally be duplicated or include repetitive information
In some cases, the assistant may display a message such as “I was unable to create a response” after already providing an answer
Results will be inconsistent when assigning multiple SQL data sources to an assistant, we’d suggest limiting this to one per assistant during the beta feedback phase

2. Intent Detection - We detect the tools you require on-demand

Assistants can now decide when to use on-demand tools to answer questions, starting with SQL.

What you’ll see:

Assistants with a SQL database connection can automatically detect when a user’s question requires querying live data
Context-aware decision-making on when to run SQL queries

💡 Tip: The more descriptive and precise you are with instructions and data source descriptions, the better the assistant can detect intent. These details directly influence how accurately it decides when to query data.

3. Knowledge Mode

You now have control over whether an assistant should use only your private data or also include global knowledge.

What you’ll see:

New “Knowledge Mode” option during assistant creation or ‘Manage’ pages
Private Data Only (default) or Global Knowledge selection
A banner in chat, making it clear which knowledge mode is active

4. Hallucination Guardrail

We’ve introduced a new safety feature that flags when a response may be inaccurate and highlights parts of the answer that fall outside your provided data sources.

What you’ll see:

Toggle in Advanced assistant settings to enable/disable guardrail
Warnings in chat or document writer when a response may contain hallucinations, with areas of content flagged

Other Improvements

Assistant Creation UX – Clearer staging area and improved multi-select options for assigning data sources when creating an assistant
File Training Reliability – Better handling of training errors, to prevent files getting “stuck” in training
Search Assistant Refresh – Updated design for clearer results and improved information display in the Search Assistant
…and other minor improvements for performance and usability