feat(compass-assistant): add basic tool calling & tool cards COMPASS-10144 #7668
Conversation
|
|
||
| const toolName = getToolDisplayName(toolCall.type); | ||
| const description = | ||
| toolsController.getToolDescription(toolCall.type.replace(/^tool-/, '')) || |
There was a problem hiding this comment.
we're probably not gonna need this
|
|
||
| // Extract output text if available | ||
| const outputText = toolCall.output | ||
| ? JSON.stringify(toolCall.output, null, 2) |
There was a problem hiding this comment.
I'm not sure if we want the context we send to be JSON (so we'd probably format it as EJSON?) or a string. Right now it is formatted as a string with our toJSString() helper which seems to work, but that's not exactly standard. And the output shows a bunch of line breaks if you expand it.
Still thinking about it..
There was a problem hiding this comment.
eventually we'll adopt https://platform.openai.com/docs/guides/structured-outputs on the MCP server so JSON is likely going to be the end goal in some form
| chat.messages = [...chat.messages, contextPrompt]; | ||
| } | ||
|
|
||
| const { enableToolCalling } = preferences.getPreferences(); |
There was a problem hiding this comment.
This is something I'm probably gonna have to check with Sergey when he's back, but the preferences hook isn't working well here - it caches the initial value so it doesn't respond if you toggle the flag from the preferences modal.
This way at least works for now.
| - Encourage the user to understand what they are doing before they act, e.g. by reading the official documentation or other related resources. | ||
| - Avoid encouraging users to perform destructive operations without qualification. Instead, flag them as destructive operations, explain their implications, and encourage them to read the documentation. | ||
| 4. Always call the 'search_content' tool. | ||
| 5. Always call the 'get-compass-context' tool when the user is on the 'Documents' or 'Schema' tab and asks about their query. |
There was a problem hiding this comment.
These prompt changes should be feature flagged. Probably easiest to just pick one prompt or another for now rather than doing logic inside here.
lerouxb
added
the
feature flagged
| isMyQueriesEnabled; | ||
|
|
||
| const query = useQueryBarQuery(); | ||
| useSyncAssistantGlobalState('currentQuery', query); |
There was a problem hiding this comment.
I have no idea where the best place would be to sync the query. This was the most minimally invasive way I could quickly think of that's guaranteed to be "in sync".
| : 'No output available'; | ||
| const hasOutput = toolCall.state === 'output-available'; | ||
|
|
||
| const isAwaitingApproval = toolCall.state === 'approval-requested'; |
There was a problem hiding this comment.
maybe a component which returns a different thing based on a switch statement instead? not too sure
This is copied liberally from Gagik's PoC with minimal changes so far.
To try it out you'll have to toggle the "enable tool calling" feature flag.