Algorithm 2 Query Encoding for Retrieval

Require: query q, embedding meta meta, model name M, flags NormalizeSQL, MaskPII Ensure: query vector z ∈ Rd 1:  q' ← NormalizeGeneric(q, NormalizeSQL, MaskPII) 2:  if meta.type = st then 3:      z ← SentenceTransformerEncode(M, q') 4:      z ← L2Normalize(z) 5:  else if meta.type = tfidf then 6:      Xc ← meta.vec_char.transform(q') 7:      Xw ← meta.vec_word.transform(q') 8:      X ← [Xc; Xw]    ▷ concatenate sparse features 9:      X ← L2Normalize(X) 10:    z ← ToDense(X) 11: end if 12: return z