Go · Inference Implementation

Go with KERNSTACK inference

Go handlers calling catalogue models is the shortest path from an API to an intelligent API on KERNSTACK: embed text with text-embed-v1 (3 credits per 1,000 tokens), classify tone with sentiment-multilingual-v1 (2 credits per 1,000), or score an image with a vision model (3 credits per image) — invoked by name from the same project, with no model vendor’s key in your secrets.

Model Inference stack.json IaC

Talk to us Read Technical Documentation → Software Lifecycle Cloud →

Go Inference

Two invocation styles cover most shapes. Direct calls from a Go handler suit request-time intelligence (embed the question, retrieve, answer). Declared ai steps in a workflow suit pipeline intelligence (score every incoming document), carrying context from earlier steps and an output schema so downstream steps consume structured results.

The economics are posted and small enough to reason about in a design review: a support assistant embedding two million tokens a month spends about 6,000 credits on embeddings; a moderation pass over ten thousand images spends about 30,000. Both land on the same balance as the Go compute around them, attributed to the calling code.

When the catalogue lacks the model, the Go code does not change its posture: a custom inference_endpoint (your runtime, your model bucket) is invoked the same way, so experiments graduate from catalogue to custom without an integration project.

CLI

Plan Go Inference, then deploy.

Terminal — CLI workflow kernstack cli

kernstack add function --name api --blueprint func-go-minimal
kernstack plan
kernstack deploy

Docs for Go Inference. Inference on the platform.

go-inference.txt Go

kernstack add function --name ask
# call catalogue models by name from the handler;
# or declare an ai step in a workflow (below) and let the platform call it
kernstack plan && kernstack deploy

stack.json — Go + inference plan · deploy

{
  "resources": {
    "functions": [{ "name": "ask" }],
    "search_indexes": [{ "name": "answers", "index_type": "vector",
      "sources": [{ "type": "buckets", "name": "help-center" }] }],
    "buckets": [{ "name": "help-center" }],
    "workflows": [{ "name": "triage", "steps": [
      { "id": "score", "ai": { "model": "sentiment-multilingual-v1",
          "prompt": "Classify the tone of this support message." } }
    ] }]
  }
}

FAQ

Frequently asked questions

Do we need an external AI provider account?

Not for catalogue models. They are platform resources invoked by name and metered in credits. External providers remain possible from worker code, but they become a choice, not a prerequisite.

Can a workflow call a model without code?

Yes. An ai step names the model (or endpoint), the prompt, optional context from earlier steps, and an output schema. The run history shows the step’s result like any other step.

How do we switch models later?

Change the model name at the call site or step and redeploy; corpora and indexes do not move. For embedding-model changes, the vector index re-embeds from its sources as a declared operation.

Connected in the Project Graph

Related capabilities, architectures, and comparisons

Ship your stack on the Software Lifecycle Cloud.

Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.

Talk to us See pricing & credits →