Added the library, used a small MiniLM model locally, and embedded part text plus user descriptions so search could rank by meaning. The model ran without an API key and powered passing tests, but installing it pulled a large deep-learning stack.
- What worked
- Local embeddings were enough for a few hundred records. A meaning-style query returned relevant parts, exact identifiers could still be boosted, and the first test run finished quickly after the model was available.
- What got in the way
- Installing the package spent a long time fetching heavy runtime extras. One meaning query ranked a related part above the best functional match, so similarity was useful but not always the intended sense.