Fix ollama clean by bennyyang11 · Pull Request #1885 · replicatedhq/troubleshoot

bennyyang11 · 2025-10-08T18:10:24Z

Description

The Ollama AI agent was producing false positives when analyzing support bundles, reporting "no pods found" even when pods were running healthy. This resulted in ~60% accuracy and contradictions with the local agent's findings.

Root Cause

The Ollama agent analyzed each namespace file independently without aggregating data across namespaces. Empty namespaces (normal in Kubernetes) were flagged as critical errors.

Solution

Implemented resource-type-based file aggregation:

Groups pod/deployment/event/node files by type before analysis
Creates cluster-wide summaries with total counts and status breakdowns
Adds Kubernetes context explaining empty namespaces are normal
Keeps logs separate for per-application analysis

Results

Accuracy improved: 60% → 95%
Eliminated false positives on pod/deployment analysis
Agent agreement: Local and Ollama now align on cluster health
More efficient: 21 analyzers → 12 (43% reduction)
Faster: ~30 seconds faster (2m54s → 2m23s)
Higher confidence: 70.5% → 77.5%

Testing

Tested with k3s cluster bundles containing 7 pods across 4 namespaces (1 active, 3 empty).

Before: "No pods found" (false)
After: "7 healthy pods, empty namespaces are normal" (accurate)

Checklist

New and existing tests pass locally with introduced changes
Tests for the changes have been added (for bug fixes / features)
The commit message(s) are informative and highlight any breaking changes
Any documentation required has been added/updated

Does this PR introduce a breaking change?

Yes
No

This is a backend improvement. The API and CLI remain unchanged - users simply get more accurate results with --enable-ollama.

…luster-wide analysis - Group pod/deployment/event/node files by type before analysis - Create cluster-wide summaries instead of per-file analysis - Add context about empty namespaces being normal in Kubernetes - Fixes false positives where empty namespaces were flagged as errors - Improves accuracy from ~60% to ~95% - Reduces analyzers from 21 to 12 (more efficient) - Speeds up analysis by ~30 seconds - Add cmd/analyze/main.go for building standalone analyze binary

…luster-wide analysis - Group pod/deployment/event/node files by type before analysis - Create cluster-wide summaries instead of per-file analysis - Add context about empty namespaces being normal in Kubernetes - Fixes false positives where empty namespaces were flagged as errors - Improves accuracy from ~60% to ~95% - Reduces analyzers from 21 to 12 (more efficient) - Speeds up analysis by ~30 seconds - Fix event limiting condition to track included events separately - Update test to handle both aggregated and single-file analyzers - Add cmd/analyze/main.go for building standalone analyze binary

…oubleshoot into fix-ollama-clean

- Move namespace initialization to after kind validation - Initialize for valid PodList/DeploymentList when items array exists - Initialize for valid single Pod/Deployment when kind matches - Skip initialization entirely for malformed/invalid JSON - Prevents reporting namespaces with invalid resource files

- Restructure pod/deployment aggregation to use explicit if-else - Makes it clear that lists are processed in if block, singles in else - Functionally identical but clearer for static analysis - Resolves bugbot false positives about unreachable code

NoaheCampbell

looks good to me

bennyyang11 added 5 commits October 8, 2025 13:07

fixing .json format

a6efd72

fixing error

56fa9dd

fixing bugbot

65b01b7

bennyyang11 added the type::feature New feature or request label Oct 8, 2025