add tool to run queries against data frames #56
Conversation
| btw_tool_env_query_data_frame <- function(query, data_frame) { | ||
| d <- get(data_frame) | ||
| conn <- btw_connection() | ||
|
|
||
| if (!DBI::dbExistsTable(conn, data_frame)) { | ||
| duckdb::duckdb_register(conn, data_frame, d, experimental = FALSE) | ||
| } |
There was a problem hiding this comment.
What do you think about adding a registration system for these tables? The usage would be something like
btw_register_table(mtcars)
btw_register_table(mtcars, "my_mtcars")
client <- btw_client()If done before the client is created, the tool description could include the table names. Or we could also have a "list tables" tool to introduce the available tables to the LLM.
The other advantage is that the LLM wouldn't need to provide data_frame, it would only need to write the SQL code.
There was a problem hiding this comment.
There'd be friction in needing to register tables for use with btw, but I also like the explicit consent nature.
There was a problem hiding this comment.
Oh and the registration function would be a generic, so we could also take existing connections, remote tables, .csv files etc. (eventually, but I'd stick with starting with R data frames via duckdb).
There was a problem hiding this comment.
Do you imagine that the registration is optional or required? I do like the "if it's in my global environment, the model can find it" paradigm that's set with the existing environment and data frame description tools, which makes me lean towards making registration optional. There's also already some existing models of explicit consent happening depending on the UI; i.e. if btw is hooked up to Claude Desktop or Claude Code via acquaint, the first usages of the tools will ask for the user's permission anyway.
I do see the value in the registration system in that get() doesn't have a lot of uses beyond the global environment and models are otherwise unable to "fetch" data up to this point. The registration system sounds really helpful for those other types of data sources.
2 participants
Closes #52.
Opening as draft just to communicate that I'd want to test more thoroughly and make sure I understand the security implications here, but I was impressed with how easy this was to make happen, how powerful it feels already, and how effectively sandboxed it feels from the R session. Do feel free to leave comments before I mark it as ready for review. :)