feat(W-19126747): use RAG to list tools #129
Conversation
| const filteredResults = searchResults.labels | ||
| .map((id) => assets.tools.find((c) => c.id === id)!) | ||
| .filter((tool) => !enabledToolNames.has(tool.name)) | ||
| .slice(0, 5); |
There was a problem hiding this comment.
Thought: What if we had the "return count" be a parameter that defaults to 5? We could instruct the LLM to increase the tools or commands returned if the user does not initially find the tool/command they are looking for.
There was a problem hiding this comment.
This could also help with this type of question. If someone wants to know what all the tools are, this new approach does not work.
|
In most cases, this is finding the correct tool to enable 🎉 |
| // Filter search results to exclude enabled tools and get top 5 | ||
| const filteredResults = searchResults.labels | ||
| .map((id) => assets.tools.find((c) => c.id === id)!) | ||
| .filter((tool) => !enabledToolNames.has(tool.name)) |
There was a problem hiding this comment.
This !enabledToolName might cause some trouble... Check out this video, it all looked like it was going well until the model chose to run sf-list-all-tools a second time. Since the result filters out enabled tools, the sf-query-org tool was not returned the second time and it chose to build an sf command instead.
In the second call the query was:
{
"query": "query SOQL records from Salesforce org"
}
Maybe this can be fixed by tweaking the description? Something like: "Don't run this tool twice for the same basic query." or "Always check enabled tools before running sf-list-all-tools?
Screenshare.-.2025-08-12.10_54_45.AM.mp4
What does this PR do?
Implements a RAG for listing tools so that we can limit the amount of tokens used when calling
sf-list-toolsWhat issues does this PR fix or reference?
@W-19126747@