A quant asked us: have you built a semantic layer mapping a phrase like “mortgage rates” to the sectors it belongs to? Yes — grounded in 88 million SEC filing sentences, 2019–2025. No keyword lists, no hand-mapping. The meaning of the phrase finds the companies.
Each phrase is embedded and matched against every filing sentence; the nearest matches are rolled up by the issuer’s SIC industry. Bars show each industry’s share of the on-theme sentences; chips are real tickers in that bucket. Nothing here is curated — this is what the meaning of the phrase returned.
For a phrase and its home sectors, we track how much of that sector’s filing language sits on the theme — on-theme sentences per 10,000, by filing year. This is the sector-level trending signal: it rises before the narrative is obvious.
The query runs through E5-base-v2 (768-dim) — the same model that embedded the corpus, so phrase and filings live in one vector space.
Cosine search over 88M filing sentences (2019–2025) held in Qdrant. Synonyms, paraphrases and jargon all land near the concept — “rate-lock pipeline” matches “mortgage rates” without either word.
Each matched sentence carries its issuer. We aggregate by SIC for the sector map, and by year for the trend — both computed on the fly, not stored.
GICS is licensed; SIC is public, so the demo maps to SIC and stays freely shareable. The mapping isn’t to a taxonomy — the layer works the same against any sector scheme you hand it.
Look at data-center power: the top bucket is filed under “Finance Services” because it’s crypto miners repurposing rigs for AI compute — the static SIC code hasn’t caught up, but the language has. A semantic layer sees the theme years before the classification does.