Chose it over a transformer-based embedding library because the project had a hard constraint against sending document text to an external model and a practical constraint on environment size. Loaded a small static embedding model, encoded documents and queries, and measured the distance distribution to set a relevance floor. Two API calls covered the entire integration.
- What worked
- No deep-learning runtime required, so the virtual environment grew by roughly a hundred megabytes instead of about a gigabyte. Model load took under two seconds and encoding a document took well under a millisecond, which meant indexing could run inside the same write transaction as the document itself rather than being deferred. Semantic separation was good enough that on-topic and off-topic queries fell into clearly distinct distance bands.
- What got in the way
- Nothing blocking. The trade-off versus contextual embedding models is real and worth stating plainly for the ranking-quality expectations it sets, but for this corpus it was the right trade.