WORLDINTELLIGENCE
.live · global situation
Language
UTC--:--:--
● CRITICALEastern EuropeSustained energy-infrastructure strike pattern observed● HIGHRed SeaVessel-traffic anomaly continues in southern transit lane● HIGHEast AsiaMultiple grey-zone incursions across median line in 24h window● MEDIUMSahelCivilian-protection deterioration in northern provinces● MEDIUMAndesMagnitude 6.1 seismic event, shallow depth● LOWGlobalCrypto OTC desks observe elevated cross-border settlement volumes● CRITICALEastern EuropeSustained energy-infrastructure strike pattern observed● HIGHRed SeaVessel-traffic anomaly continues in southern transit lane● HIGHEast AsiaMultiple grey-zone incursions across median line in 24h window● MEDIUMSahelCivilian-protection deterioration in northern provinces● MEDIUMAndesMagnitude 6.1 seismic event, shallow depth● LOWGlobalCrypto OTC desks observe elevated cross-border settlement volumes
2026-05-02

How AI-Powered News Aggregation Works at WorldIntelligence

From RSS ingestion to embedding-based clustering to LLM-written briefings: the pipeline behind every intelligence summary on the site.

The pipeline

  1. Ingest. Roughly 435 RSS, Atom and JSON feeds are polled on category-specific cadences.
  2. Normalise. Items are de-duplicated, language-tagged, and stripped of tracking parameters.
  3. Cluster. Stories from multiple outlets are merged into events using cosine-similarity over sentence embeddings, then thresholded against a recency-decay window.
  4. Summarise. A free-tier LLM (currently Gemma-3-27B via OpenRouter) produces a 3-paragraph intelligence brief at the cluster level, with full source attribution.
  5. Translate. Briefings are translated on demand into the seven non-English locales, cached to disk for cost efficiency.

Hard rate-limiting

We never exceed the free-tier rate limit. The OpenRouter client uses a token-bucket algorithm (1 token per 4.2 seconds, capped at 15 tokens/minute, hard-capped at 1000 requests/day) and an exponential-backoff retry loop for transient failures. If the daily quota is exhausted, the site falls back to pre-cached summaries rather than dropping requests.

Why we publish methodology

Intelligence products that hide their pipeline cannot be challenged. Every claim on the site should be traceable to a citable upstream source or to a transparent aggregation step. If it is not, it is a bug — please report it.