TUSDW AI CARE · TROUBLESHOOTING
AI problems. Clear first steps.
Practical troubleshooting for Ollama, RAG, ChromaDB, MCP, Claude, Gemini, Kimi K3 and local LLMs — plain-language, safety-first and verification-oriented.
KNOWLEDGE BASE
The first 10 problem pages
Diagnose → Explain → Recommend → Verify
Ollama Connection Refused on Port 11434
If an app, RAG pipeline or agent cannot reach Ollama on port 11434, the cause is usually that Ollama is not running, the endpoint is wrong, or the client is trying to reach the wrong machine.
Open troubleshooting ↗02Ollama Is Installed but Not Running
A local AI app can find the Ollama installation and still fail because the Ollama process or service is not active.
Open troubleshooting ↗03Ollama Model Not Found
A model-not-found error often means the application expects a model name or tag that is not installed locally or differs from the configured name.
Open troubleshooting ↗04RAG Is Not Finding Your Documents
When a RAG system answers as if your documents do not exist, the failure can be in ingestion, chunking, embeddings, storage, retrieval or filtering.
Open troubleshooting ↗05ChromaDB Connection or Collection Error
ChromaDB problems can come from a wrong persistence location, a missing collection, client/server mismatch or a project pointing at a different database than expected.
Open troubleshooting ↗06MCP Server Not Connecting
An MCP connection can fail because the server process is not starting, the client configuration points to the wrong command or endpoint, or required environment variables are missing.
Open troubleshooting ↗07Claude API Authentication Error
Authentication errors usually mean the application is not sending a valid credential in the way the current provider API expects.
Open troubleshooting ↗08Gemini API Quota or Rate Limit Error
Quota and rate-limit errors are different from authentication failures. They often mean the account, project or model has reached a current usage limit.
Open troubleshooting ↗09Kimi K3 API Troubleshooting
When working with Kimi K3, avoid assuming that authentication, endpoints, limits or compatibility are identical to another AI provider. Start from the exact current provider configuration.
Open troubleshooting ↗010Local LLM Is Running Very Slowly
Slow local inference can be caused by model size, available RAM/VRAM, CPU/GPU use, context length, concurrent workloads or runtime configuration.
Open troubleshooting ↗TUSDW AI CARE
Still not fixed?
TUSDW AI Care is being built to diagnose AI problems, explain them in plain language and verify the result afterwards.
Request AI Care Early Access ↗