Improved vector search

[kuraisle]

Vector search seems an effective step in semantic search for OMOP concepts. We've just used an off-the-shelf model so far, and it got the correct answer in the top 5 for 214/400 medications in an initial test. There are several routes for improving the results

• Add vector search metrics
  • Top k accuracy: Is the correct mapping in the top k results?
  • Mean reciprocal rank: Where does the correct answer appear? We would have to limit the list to e.g. 20 https://en.wikipedia.org/wiki/Mean_reciprocal_rank
  • Similarity: by some vector similarity metric (cosine similarity etc.), how similar is the input vector to the correct answer? Comparing this between models may not be valid.
• Test different models: There are biomedical specialist BERTs available. Also, larger and smaller generalist models. We should try some of these and compare how they do.
• Model fine-tunes: This will be a stretch, but doable