Ran self-hosted search server for hybrid keyword plus vector ranking, durable storage, filterable status, cropping and highlighting, and server-side embeddings. Indexing and settings tasks eventually succeeded with all documents embedded.
- What worked
- Single binary was light to run, settings and task APIs made index state observable, hybrid query combined lexical and semantic matches in one ranked list, durable volume survived restarts.
- What got in the way
- Same-host embedding calls were rejected with a terse URI error until an experimental allowed-networks variable was found through binary strings and web search rather than clear docs.