Skip to content

Technical overview

Implementation model for shells, packs, intake channels, workflows, policy overlays, integrations, evaluation gates, and governed execution.

Platform architecture

  • Channel adapters normalize inbound payloads into workflow-compatible intake events.
  • Workflow engines produce schema-valid WorkItems with decision-step state transitions.
  • Evidence services attach citations and retrieval traces to response and packet generation.
  • Integration runtime executes actions asynchronously with retry and idempotency controls.
  • Telemetry and audit streams provide end-to-end traceability across the lifecycle.

Knowledge and ingestion

Supported ingestion modes

  • Refresh pipelines for knowledge sources
  • Upload pipelines for documents and policy bundles
  • Connector-sync pipelines for managed external systems

Ingestion behavior

  • Ingestion jobs emit status, latency, and failure diagnostics
  • Pipelines support safe retries and backoff strategy
  • Asset-level policies define freshness and retention behavior

Render and extraction settings are configurable per source profile.

Freshness operations

  • Scheduled refresh and on-demand reprocessing
  • Diff visibility to evaluate source changes
  • Staleness alerts tied to workflow risk profile

Evidence retrieval

Retrieval profiles: Profiles define recall, precision, and citation strictness by workflow and risk class.

Decision thresholds: Confidence, evidence availability, and policy checks determine whether to answer, review, or escalate.

Response and packet generation

WorkPackets are generated from canonical WorkItem state and evidence bundles.

Destination renderers map canonical packet content to channel-specific payload shapes.

Safety boundaries

  • Strict citation mode for policy-critical workflows
  • Explicit refusal and escalation when evidence is insufficient
  • PII-aware handling and configurable redaction policies

Channel activation

Activation setup: Each intake channel ships activation checks, health signals, and scoped security bindings.

Activation checks

  • Origin binding and channel verification
  • Performance-conscious integration loading behavior
  • Accessible interaction and localization support

Operational controls

  • Enable, disable, and monitor intake channels by environment
  • Apply channel-specific policy overrides
  • Inspect health state and telemetry by internal channel identifier

Workflow and action APIs

Public API intake channels support WorkItem operations, policy-aware execution, and telemetry export.

Authentication

  • Scoped credentials for server-to-server access
  • Role and tenant context validation for each request
  • Fail-closed behavior on missing authorization context

Authorization scopes

  • Read and manage WorkItems
  • Manage policy and decision-step state transitions
  • Execute governed actions with explicit permission

Core endpoint groups

  • WorkItem lifecycle operations
  • Decision-step and approval operations
  • Integration execution and retry operations

Response model

  • Typed success and failure envelopes
  • Machine-readable reason codes for policy outcomes
  • Trace and correlation identifiers for debugging

Telemetry dimensions

  • channel identifier and workflow_id
  • work_item_id and policy_version
  • integration_id and action outcome state

Integration orchestration patterns

Integrations can be connected through native adapters or middleware where required.

Execution patterns

  • Create and update destination resources
  • Attach packet context and evidence references
  • Retry with error classification and dead-letter handling

Ingestion patterns

  • Receive request events from external tools
  • Normalize into workflow schema
  • Route to decision steps and action proposals

Governance controls

  • Action allowlists by workflow and role
  • Approval requirements by risk profile
  • Audit event capture for each mutation

Security model

Security controls are applied at intake, decisioning, execution, and telemetry layers with clear ownership boundaries.

Data handling model

Stored artifacts

  • WorkItem state and lifecycle metadata
  • Evidence references and packet render metadata
  • Execution outcomes and audit event records

Not retained by default

  • Unscoped credential material in plaintext form
  • Unbounded raw payload logs
  • Cross-tenant shared request state

Retention: Retention and archival policies are configurable by policy scope and data classification.

Model training posture: Data handling policies should explicitly control whether workflow artifacts are eligible for model improvement workflows.

FAQ

Can we launch with only web intake?
Yes. Web can be your first channel while you keep the same workflow primitives for future channels.
Can integrations run asynchronously?
Yes. Integration execution is asynchronous and supports retries with auditable error classes.
How do we trace a failed outcome?
Use correlation IDs, WorkItem history, decision events, and execution logs to reconstruct end-to-end flow.