Context is a curated knowledge graph of facts—each carrying its provenance, validity-in-time, and a sensitivity label. Human assertions outrank machine-derived ones, pipelines refresh it continuously, and agents query it at runtime instead of re-deriving the truth on every call.
A curator's value overrides the pipeline's extracted date—and agents read the resolved fact with its citation.
What a fact carries
Not a vector blob and not a raw API response—a typed assertion that knows where it came from, when it's true, who can see it, and who asserted it.
Context stores discrete, typed assertions—“Riverside's BAA renews 2026-03”—linked into a graph, not blobs of text an agent has to re-parse and re-interpret on every call.
Each fact records where it came from: which source system, which document, or which named person asserted it—so every downstream answer is traceable and citable.
Facts carry valid-from / valid-until and an as-of timestamp. The graph knows what was true last quarter and what holds today, so agents never act on silently expired data.
Every fact is tagged public, internal, confidential, or PHI. Retrieval filters by the caller's clearance, so an agent only ever sees facts it's authorized to use.
When a human-entered fact and a machine-derived one conflict, the human assertion takes precedence. Curated institutional knowledge overrides noisy automated extraction.
Pipelines keep facts current on a cadence or on source events; agents query the resolved graph at runtime. Reads are decoupled from the source systems behind them.
How a fact stays trusted
Context replaces brittle, per-request retrieval with a governed pipeline that runs from source system to resolved, cited fact.
Define the entities, relationships, and fact types each tool needs—then never re-map the hospital for the next project.
Point fact types at CRM, EHR, contract, and policy systems with named owners, refresh cadence, and event triggers.
Curators add and override facts with provenance and severity; these outrank machine-derived values when the two disagree.
On each refresh the pipeline reconciles conflicts, stamps validity windows, and versions the fact instead of overwriting history.
Governance reviews, the builder, and managed agents query the same graph—filtered by sensitivity and returned with citations.
Each fact type maps to a named source with its own sensitivity label and refresh cadence.
Refresh, don't re-ping
The naive approach hits every source system on every agent query. Context moves that work into pipelines that run on a cadence—so source load stays flat no matter how many agents query, and reads are fast, cited, and rate-limit-safe.
Pipelines pull from each governed source on a cadence (or when the source emits an event), resolve conflicts, and write versioned facts to the graph. Agents query that resolved graph—they never fan out to live systems on every request.
Drag to set how many agent queries your fleet runs per day.
Every query fans out to ~4 live systems. Load scales with traffic.
Flat, scheduled refreshes. Adding more agents doesn't touch the source systems.
One graph, every tool
Stop rebuilding local knowledge stores. Catalyst apps read the same governed graph, scoped to their clearance at query time.
Assessment evidence, risk drivers, and approval records stay attached to the entities they describe.
Builders request approved context packs and searches instead of standing up one-off knowledge stores.
Production agents run only against approved relationships, with PHI scope enforced at retrieval time.
Explore the suite
Open the Context workspace to inspect facts, their provenance and validity, the pipelines that refresh them, and the activity behind every answer.