Theme Tracker

Each cell shows the rerank_score of the single best-matching chunk from that company's 10-K or 10-Q in that period, against the theme's fixed search query.

rerank_score = cosine_similarity + 0.08 × lexical_score
Cosine similarity is the dot product of the chunk's embedding vector and the query's embedding vector — it measures semantic proximity in the model's latent space. The 0.08 × lexical term adds a small boost when query keywords appear literally in the chunk text.

What a high score means: a chunk exists in this period whose embedding is close to the theme query. The theme's vocabulary is present and retrievable.

What a high score does not mean:

• It is not a count — one highly-matching chunk scores the same as a filing saturated with that topic.
• It does not capture sentiment — a company disclosing a serious breach and one mentioning routine password policies can score similarly on "cybersecurity."
• It does not measure prominence — a passing footnote and a dedicated Risk Factors section score on the same scale.
• Score differences within a row are small. 0.46 → 0.64 looks dramatic as a colour gradient but the absolute embedding distance is modest. Do not read this as a precise ranking between companies.

Use this as an exploration signal: "does this topic show up in this period's filings?" — not as a quantitative measure of exposure or emphasis.