Skip to content
KinetixAI
Start Free Trial

Documentation

Documentation

Quickstart, SDKs, API reference, architecture, integrations, and security for the Kinetix control plane.

Everything you need to deploy a routed, cached, deterministic inference pipeline. The SDKs are open source; the control plane is managed.

Quickstart

Install the SDK, export an API key, and initialise a pipeline. The CLI writes a kinetix.config.ts file with routing and cache defaults.

npm install @kinetix/sdk
export KINETIX_KEY="kx_live_..."
npx kinetix init
npx kinetix deploy --prod
The Developer tier includes 500,000 requests per month and a 14-day trial. No credit card is required to create keys.

SDKs

First-party clients cover the languages most teams use for production inference. Community SDKs are listed in the project gallery.

  • Python — pip install kinetix, typed responses, async support.
  • Node.js — npm install @kinetix/sdk, ESM and CJS builds.
  • Go — go get github.com/kinetix-ai/kinetix-go.

API reference

All endpoints are JSON over HTTPS and authenticated with a bearer key. Requests are idempotent when an Idempotency-Key header is present.

EndpointPurpose
POST /v2/runExecute a pipeline with routing and caching.
GET /v2/pipelinesList pipelines and their deployment status.
POST /v2/cache/queryQuery the semantic cache directly.
GET /v2/usageRequest counts, spend, and routing splits.

Architecture

Kinetix is a control plane, not a model host. Your application calls one endpoint; the router chooses a provider, checks the semantic cache, and streams the result back. Provider credentials never leave the control plane.

Adaptive routing

Each request is scored for reasoning depth, context size, and latency sensitivity. Cheap, well-cached queries go to edge models; complex reasoning goes to frontier models. Fallback chains are deterministic, so a provider outage does not change your output schema.

Deterministic context engine

Embeddings are cached by content hash, not by conversation ID. Retrieval is sub-10ms at p50 for warm indexes, and every cached response carries the schema version that produced it.

Integrations

Use Kinetix with the model providers and data stores you already run.

  • Models: OpenAI, Anthropic, Google, Mistral, Cohere, and 35+ OpenAI-compatible endpoints.
  • Vector stores: pgvector, Pinecone, Weaviate, Qdrant, and Redis Vector.
  • Observability: OpenTelemetry, Datadog, Grafana, and S3-compatible log export.
  • Identity: SAML SSO and SCIM provisioning on Enterprise.

Security & compliance

Enterprise instances run in a dedicated VPC with customer-managed encryption keys. Prompts, completions, and embeddings are never used for model training, and retention can be set to zero.

  • SOC 2 Type II controls, with reports available under NDA.
  • HIPAA BAA available on Enterprise plans.
  • Zero data retention and zero-training commitments.
  • SSO, SCIM, audit logs, and per-key spend limits.

Status

All systems operational. Routing, semantic cache, and the management API are healthy across the US, EU, and APAC regions.

Changelog

  • v2.4 — Sub-millisecond vector query routing and per-pipeline cache budgets.
  • v2.3 — Node.js SDK, OpenTelemetry spans, and SCIM provisioning.
  • v2.2 — Enterprise VPC isolation and customer-managed keys.

About this demo

Kinetix AI is a fictional B2B SaaS brand built by Rokhsana as an enterprise WordPress showcase. The design system, performance budget, accessibility behaviour, and client editing workflow are production patterns — the product metrics are illustrative.

Contact

For a walkthrough of the build, email hello@rokhsanak.xyz or open the project gallery.