How it works
From public source to citation-validated brief.
- 01
Continuous ingestion
We have catalogued 13,000+ public sources across 12 sectors (as of July 2026) — SEC EDGAR, government APIs, trade press, academic preprints, GitHub telemetry, USPTO, PACER court records, Wayback. Update cadence ranges from real-time (CISA advisories) to weekly (annual filings).
- 02
Chunk + embed + index
Documents are deduped by body-hash, chunked at multiple granularities, embedded via voyage-3 (1024d), and indexed in pgvector with HNSW — catalogued sources are embedded and indexed as capacity allows, highest-signal first. Sector-tuned embeddings and reranking are designed to improve retrieval precision over generic embeddings; measured gains are published as our benchmark suite matures.
- 03
Per-persona retrieval + synthesis
At 06:00 customer local time, the coordinator agent retrieves the day's top chunks per persona's tuned source weights. Specialist sub-agents (legal, financial, technical, geospatial) contribute where the section calls for it.
- 04
Citation validation
The Validator agent verifies every claim against retrieved chunk substrings. Failures retry once with a stricter prompt; on second failure the paragraph renders 'insufficient evidence' rather than fabricating.
- 05
Delivery + entity-graph augmentation
Brief is delivered to in-app + email + voice (optional). Every named entity is resolved against the cross-sector entity graph and surfaced as a hover-linkable chip. Forum Forecast probability shifts are computed and revision-tracked.