Resources
Deployment
Hosted service, provisioned endpoint controls, and operational limits.
The hosted Integrity service is the production deployment. SDK and gateway clients support base-URL overrides for an endpoint provisioned for your environment.
Deployment topologies#
| Topology | Status | Boundary |
|---|---|---|
| Triage hosted | Available | Out-of-the-box SDK and redirectable gateway clients. Model-provider endpoints remain outside the customer tenant; use only for non-ITAR/CUI workloads approved for this hosted boundary. |
| Customer AWS / GovCloud | Under development; enterprise-provisioned only | Designed for sensitive workloads with Integrity and Bedrock endpoints controlled inside the customer AWS boundary. It is not a generally available or self-service deployment today. |
Do not send ITAR, CUI, or other contractually restricted data through the hosted topology unless your security and compliance owners have approved that boundary. A diagram or roadmap item is not evidence that the customer-tenant deployment has been provisioned.
Hosted (default)#
The hosted service at https://integrity.triage-sec.com is the default and needs no setup beyond a project API key. It terminates TLS at the edge and keeps GPU inference warm so classifier latency stays low. The SDKs and REST endpoints target it out of the box. Redirectable Codex and Claude Code clients can use the same endpoint through the gateway instructions; fully hosted agent surfaces without a base-URL override are outside this path.
Dedicated deployment#
The customer-controlled AWS/GovCloud deployment is being built for sensitive-data use cases where the model endpoint (including Bedrock) and Integrity boundary are controlled inside the customer tenant. It is not a self-service or generally available product today. Identity translation, storage, networking, updates, rollback, evidence, and cutover criteria must be validated for each enterprise deployment. This page does not claim general VPC, Kubernetes, on-premises, Bedrock SigV4, or Vertex OAuth availability on the hosted gateway.
Dedicated deployments are an enterprise engagement. We size and provision them with your team. Reach out at info@triage-sec.com to scope one, including data-residency and provider requirements.
Repointing clients#
After Triage provisions a compatible endpoint, point a Python SDK client at it with base_url:
triage_sdk.init(api_key="tsk_...", base_url="https://integrity.your-org.example")Per-classifier SDK URL overrides are also available for migrations. See Configuration & endpoints.
Request and timeout limits#
Current hosted gateway runtime defaults are:
| Limit | Default | Behavior |
|---|---|---|
| Max request body | 10 MiB | Larger requests are rejected before any guard or upstream call. |
| Max decoded upstream body | 16 MiB | Buffered bodies and incremental streams are terminated before unbounded accumulation. |
| Upstream I/O timeout | 60s | Connect, read-inactivity, write, and pool timeout per operation; not an overall wall-clock deadline. |
| Gateway response-start deadline | 110s | Cancels provider and guard work and returns a traced HTTP 504 if the proxy has not started a response. |
| Stream gap timeout | 110s | Maximum upstream gap between streamed chunks, bounded below the edge idle timeout. |
| Edge origin idle timeout | 120s gateway / 60s classifiers | Bounds the wait for the first origin packet and each later idle gap; it is not a total SSE duration cap. |
Per-project custom upstreams are restricted to operator-approved HTTPS hosts on port 443. The gateway rejects non-public DNS answers immediately before connecting. Customer-tenant deployments must also enforce outbound network policy because application DNS preflight alone cannot eliminate a re-resolution race.
Classifier calls made directly through the REST endpoints or SDKs are small and fast; these bounds mainly matter when you route full chat completions through the gateway.