My embedding server died and I didn't notice for two weeks

Short summary

The author built an MCP server called Codicil that indexes repo docs for AI coding assistants, using Ollama for embeddings. When the embedding host died, the system silently fell back to keyword grep for two weeks without the user noticing. The real lesson is about deliberate fallback design — graceful degradation, file locking, and per-model Chroma collections — that kept the tool useful even when its primary service was down.

• •MCP server Codicil fell back from semantic search to keyword grep for 2 weeks unnoticed
• •Fallback design: grep on disk, fcntl locking, per-model Chroma collections
• •Keyword search is worse but the system stayed functional instead of breaking

Generated with AI, which can make mistakes.