From 80-page PDF to a one-glance decision

The meaning layer

Keyword search only finds the words you typed. We go a layer deeper: every chunk of every opportunity — and every one of your capability facets — is turned into a 1,024-number vector that encodes its meaning. Close meanings sit close together in that space, even with zero shared words.

Matching becomes geometry — cosine distance between your facet and every chunk. “UST removal” lands right next to “underground storage tank decommissioning” and “petroleum spill response” — all three a plain keyword search would miss.

Vectors alone over-smooth, so a keyword (full-text) pass runs in parallel for the exact codes and proper nouns, and the two are fused by reciprocal-rank fusion — then gated by NAICS, geography and value, and reranked by a cross-encoder. Meaning for recall, keywords for precision.