I spent 3 years training ML and RL models on market data and kept hitting the same wall: raw price series and the indicators derived from them have terrible signal-to-noise ratio, and a model with frozen weights goes stale as regimes shift. So I tried feeding models news instead, and hit a different problem. Getting the relevant news. Every news API I tried matches keywords. Some ship pre-labelled tickers, but the labelling is still string matching underneath, so you get every article that happens to contain a word and miss every article that means the thing without saying it. Here is how Clair works under the hood: Every article is split into chunks of a few sentences, and each chunk is embedded with bge-m3 (open weights, 1024-dim). Your query is embedded the same way, and matching is cosine similarity, you provide a threshold in the request, if the meaning is closer than your threshold the article is included in the response. An article scores as its single best-matching chunk, never the average. A long report that spends one paragraph squarely on your topic still surfaces, and a rambling one can't dilute its way past your threshold. Every result carries the chunk that produced it, so you can see exactly what matched and re-embed it to check the number yourself. There is no LLM anywhere in the serving path, which is the part I actually care about. The same article and the same query always produce the same number. Nothing is generated, summarized, or re-ranked by a model that might change under you. It can surface a weak match, but it cannot invent an article or a quote.

Clair News API

Semantic news API that matches meaning not keywords

Semantic news API that matches meaning not keywords