Skip to content

llm: avoid exposing the model to "variables" concept - #9993

Merged
vito merged 3 commits into
dagger:mainfrom
vito:tidy-llm
Mar 28, 2025
Merged

llm: avoid exposing the model to "variables" concept#9993
vito merged 3 commits into
dagger:mainfrom
vito:tidy-llm

Conversation

@vito

@vito vito commented Mar 27, 2025

Copy link
Copy Markdown
Contributor

This was a bit of a halfway measure and is currently detrimental to consistent model behavior.

For example, I've observed:

  • needless calls to readVariable
  • attemps to pass variables as function args

With this change it will instead be 100% focused on objects and their functions. String variables are always just prompt vars, never exposed to the model.

vito added 3 commits March 27, 2025 18:43
This was a halfway measure to allowing the LLM to discover vars that
were set for it. As-is, it confuses the model into unnecessary calls and
into trying to pass variables as arguments to tools, which is a bit too
much ambiguity.

Signed-off-by: Alex Suraci <suraci.alex@gmail.com>
Let's just make everything 100% focused and strictly around IDs, so the
model never (lol) tries anything funny.

Signed-off-by: Alex Suraci <suraci.alex@gmail.com>
these are fine because the LLM never actually sees them

Signed-off-by: Alex Suraci <suraci.alex@gmail.com>
@vito
vito merged commit fe21113 into dagger:main Mar 28, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants