feat: deep dive alerts #851
Conversation
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
…s/integrations Signed-off-by: Matthis Holleville <matthish29@gmail.com>
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
cmd/analyze/analyze.go
Outdated
| @@ -88,6 +122,8 @@ func init() { | |||
| AnalyzeCmd.Flags().BoolVarP(&nocache, "no-cache", "c", false, "Do not use cached data") | |||
| // anonymize flag | |||
| AnalyzeCmd.Flags().BoolVarP(&anonymize, "anonymize", "a", false, "Anonymize data before sending it to the AI backend. This flag masks sensitive data, such as Kubernetes object names and labels, by replacing it with a key. However, please note that this flag does not currently apply to events.") | |||
| // alert flag | |||
| AnalyzeCmd.Flags().StringVarP(&alert, "alert", "", "", "Alert to be analyzed") | |||
There was a problem hiding this comment.
Let's offer a shorthand flag also, it doesn't need to be "-a"
There was a problem hiding this comment.
What do you think about "l" ?
There was a problem hiding this comment.
that's fine, might be easier to make it a less ambiguous char like g or y
|
What makes this OpenAI specific? I expected to see us use OpenAI functions, but it looks like we're just doing some smart prompt manipulation instead. |
|
We had the choice between using OpenAI's assistant mode and being specific to OpenAI, or using a multi-agent system that we maintain ourselves. This is the choice I made so that this mode is available for all AI backends. |
Okay that's even better. I am looking for an opportunity to re-use some of your work in the interactive loop I am making |
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
|
If you need help or suggestions for integrating multi-agent functionality, reach out to me. I can finalize the few remaining jobs quickly to have a 'ready for review' pull request. |
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
Signed-off-by: Matthis Holleville <matthish29@gmail.com>
… in the prompt and providing it with more error context. Signed-off-by: Matthis Holleville <matthish29@gmail.com>
|
Sorry I lost this thread @matthisholleville - I think my primary reservation was was originally around it being bound to OpenAI functionality; I think you subsequently disproved that. Shall we get this through another review? The schema has just been approved. |
|
@AlexsJones I'm diving back into it this week and I'll see about getting it ready for review as quickly as possible. |
Closes #795
📑 Description
Disclaimer : This PR is a first implementation and should be consolidated
It's deeply reworks our interaction with the AI. Until now, we had an AI agent (which we could call "analyzer"). Its goal is to provide an explanation for an error sample. This agent is still used for the "classic" explain mode.
For "deep dive alert" analysis, it was necessary to transform this model to switch to a multi-agent autonomous AI system.
This system consists of 3 AI agents:
router: its goal is to identify the appropriate analyzers to activate based on the provided alert (k8sgpt analyze --alert ...).evaluator: once the analysis has been performed, this agent's objective is to filter errors related to the provided alert. This agent helps refine errors that may be related to the alert.analyzer: it's the historical analyzer responsible for finding a solution to a problem.A new
pkgcalledagenthas been created to host the code for these agents.full-demo-deep-dive.mov
✅ Checks
ℹ Additional Information
Other improvements:
jsonformat provided )A description has been added to each of the analyzers to provide context to the "router" agent about the purpose of the analyzer.removed due to frequent maxToken hitThis PR is a first implementation and should be consolidated with the following elements:
jsonformat is provided.With this mode, we make 2 additional AI API calls per command (router and evaluator agent).
Analysis flow with error