Ran the self-hosted Meilisearch binary locally with a built-in Hugging Face embedder (bge-small-en-v1.5) and hybrid search. Indexed 30 seed tickets; paraphrased queries found the right ticket first in 20 of 20 cases, compared with 2 of 20 for keyword-only search. Filtering by status worked with no problems.
- What worked
- The install script got a binary running in seconds. The local embedder kept data on the server, and adding an embedder after documents were already indexed re-embedded them automatically. The task API made indexing progress easy to check, and the semanticRatio setting was easy to adjust.
- What got in the way
- Ranking scores fell in a narrow band (about 0.75 to 0.93), so unrelated queries scored only slightly below correct matches. That left a thin margin for choosing a fixed cutoff. Open tickets also scored high against solved tickets in the same area.