Six capability tracks, all already proven against a live 6.4M-article Benzinga corpus — not a proposal built from slideware.
Any chatbot can fetch today's headlines. Almost none of them can tell you which headline actually moved the stock.
Retrieval is the one thing LLMs already do better than any wire product — a frontier model with web search pulls recent news on any ticker for free. That makes raw "we have the news" a shrinking asset. The durable value sits one layer up: things that require holding the whole corpus at once and measuring it against what actually happened next. That's the layer described below, and it's already running.
Collapses N wire hits covering one event — syndicated reprints, law-firm class-action pile-ons, multi-source earnings recaps — into one canonical story plus a coverage count. Cleans the feed and gives you a native "trending" signal for free.
A 9-class taxonomy (earnings, M&A, legal, personnel, regulatory, capital action, product/partnership, promotional, other) trained on labeled samples — replaces or audits broad channel tags, and flags promotional/spam wire filler for automatic exclusion.
Every article joined to split-correct forward returns (1d/3d/5d) and scored for P(the stock actually moved ≥5%). This is the piece a chatbot structurally cannot produce — it requires the full corpus and the price tape at once. Natural premium-tier differentiator.
Natural-language search over every sentence and article back to 2016 — "management commentary on margin pressure," not just ticker + keyword. Sits underneath any agent or product surface you want to expose it through.
Questions no single article answers: which names had the most material news this quarter, event-type frequency by sector over time, stories that were under-covered but moved the stock anyway, and where realized returns diverged from predicted materiality (the market got surprised).
The same extraction technique we run in production elsewhere — pulling structured deal events (acquirer, target, amount, type) out of raw wire text at scale, cross-lingual — applied to your corpus turns "articles that mention a company" into a linked graph: company ↔ people ↔ deals ↔ every classified event, with a real timeline instead of a pile of headlines.
| Capability | A chatbot with web search | Measured over the full corpus |
|---|---|---|
| "What's the latest on TICKER" | Does this trivially | — |
| "Which of today's stories actually moved a stock" | Cannot compute | Materiality + return join |
| "What kind of event is this, systematically" | Inconsistent, per-call | Fixed 9-class taxonomy |
| "Show me every M&A event for this company since 2020" | Recency-biased, incomplete | Structured, longitudinal timeline |
| "Which under-covered story just moved the stock 8%" | Can't scan the archive | Coverage × materiality join |
Track 06 is the connective layer over everything above: instead of a company's news being a flat list of headlines, it becomes a node connected to the people, deals and filings mentioned across the whole corpus — queryable as a timeline, not re-derived per chat session.