Handling Sensitive Data with Positron Assistant

[Seviks]

First, thank you for developing Positron and the Positron Assistant — it’s been a very useful tool in my workflow.
In practice we often work with sensitive or personal data (in RStudio until today). While the syntax, names and structure of our data objects are not considered sensitive, the content of data.frames (e.g. values exposed via glimpse()) may contain sensitive information.

Therefore my questions:

• Does Positron (Assistant) access or transmit the actual values contained in data.frames (for instance glimpse() or similar functions)?
• If yes, is there a way to prevent this behaviour (so that only metadata/object structures are used, not the contained data)?
• Regarding console output: I am aware that it can be toggled off, but is there a way to set this default behaviour to “off”?

I haven't found information regarding this topic (https://positron.posit.co/assistant.html). Feel free to direct me to resources that already explain this,.

[juliasilge]

Thank you for these questions!

• Positron Assistant does not have infrastructure to send an entire dataset to the model provider, but it can use "tool calls" to send the schema of your datasets. These schemas can include information about your dataset that may be considered sensitive or personal, like the mode of a string column or factor levels in R. You'll probably need to think carefully about whether metadata for data objects is truly safe/appropriate for you, given your own domain and constraints. Also, you can pretty easily prompt Positron Assistant in such a way that it will send row-level individual data points to the model provider (for example, "tell me about the fifth row of my dataframe").
• We are tracking support for explicit opt-out of various tools in Allow users to configure which tool calls to opt out of (especially the tool to install packages) #9271. This would let you opt out of, say, the getTableSummary, inspectVariables, and similar tools that can send data schema to the model provider. If you want to add any info on your use case there, that would be super helpful! There is some support right now for opting in/out of tool calls when in Agent mode (click the tool icon in the bottom right of your chat window) but we need to do better on this.
• You can control how Positron Assistant has access to your console context with chat.implicitSessionContext.enabled. Change the value for "panel" to "never" to keep the console context from ever being available to the Assistant, and read more at Add user control of assistant session context #8495 (comment).

[Seviks]

Thank you for the response, this is very close to what I was hoping for. It would be really useful if we could create a profile where the available tool calls could be defined and configured. That way, the setup could be tailored more easily to different use cases.

[Seviks]

Aside from opting in or out for different tools, it would also be nice to have a general option for how PositronAssistant sends data to the LLM. For example, similar to this option in gander:

# Only describe relevant data frame columns and their types, but don't
# provide any rows:
options(.gander_dims = c(0, Inf))