Silent Fallback: How a Missing Config Route Sent PDF Parsing to a CPU-Only Ollama Model

Short summary

A RAG app's PDF parsing job silently fell back from GPT-4.1 to a local Ollama model on a CPU-only container after a reliability fix moved indexing to a background Worker — without copying the routing config. The result: 390% CPU for an hour, zero errors, zero OpenAI traffic, and a frozen UI. The post distills three lessons: fallbacks succeed with the wrong tool, nobody alerts on silence, and config lives per process even when your mental model doesn't.

• •A reliability fix moved PDF parsing to a Worker process but didn't copy the routing config, causing silent fallback from GPT-4.1 to local Ollama on CPU
• •No errors, no failed requests — the system was 'working' by its own rules, just on the wrong hardware at 42 seconds per page
• •Key lessons: alert on absence of expected signals, treat config as per-process, and add a checklist item for config migration when moving work between processes

Generated with AI, which can make mistakes.