Skip to content
OpenPond
Join community
Cloud/Model deployments

Model deployments

Review personalized model usage, deployments, and API access.

Open Ponder → Inference → Deployments in the Console for personalized model inference. Workspace permissions control whether you can view models, invoke inference, or manage deployments.

From result to serving

Prepare a deployment

  1. Review the completed training result and its evaluation in Models.
  2. Click the Model's Serve toggle and select its completed run. The Console saves the exact eligible version internally.
  3. Review the actual serving offer. A resource-time offer shows its per-minute rate and optional spending limit. Choose Serve to approve activation.
  4. Wait for the deployment's ready state before directing traffic to it.
  5. Use API → Your models for the exact endpoint and authentication contract.

Use Overview → Your models to inspect personalized inference activity. Check deployment status and request errors when traffic fails; a model artifact alone does not establish that an endpoint is available.

Trained deployments and the OpenPond Chat catalog share the Inference area. Select Your models or OpenPond Chat in the API tab and use the endpoint and model identity shown for that source. Keep API credentials on the server.

Call the Model from your product

Keep a configured OpenPond client in your product's server code. Supply your workspace API key and give a readable alias to the exact callable Model ID shown under API → Your models. The SDK connects to OpenPond automatically:

ts
import { createOpenPondClient } from "openpond-sdk"; const openpond = createOpenPondClient({ apiKey: process.env.OPENPOND_API_KEY!, teamId: process.env.OPENPOND_TEAM_ID!, datasetId: process.env.OPENPOND_DATASET_ID!, models: { support: process.env.OPENPOND_MODEL_ID! }, }); const answer = await openpond.models.support.generate({ prompt: question }); console.log(answer.text, answer.requestId, answer.usage);

Use a matching configured-client SDK release. generate and stream use the existing serving API; configuration does not activate a deployment. Named handles are text-generation APIs; Decision scorers keep their separate scoring contract. Results retain server-provided request, usage and model-version evidence where available. An interrupted billed request is surfaced without automatic replay.

Successful saved chat-model requests automatically collect their model-visible messages and recorded response into the model's existing accessible Dataset. This includes Chat, the model's Playground, and external API/SDK requests. The server resolves and pins the destination; a client datasetId does not redirect this automatic collection. Delivery retries keep the original request identity and cannot create a second task. Missing, inaccessible or ambiguous model/Dataset links are not replaced by a newly created Dataset.

The recorded response is attempt evidence, not an approved reference answer. Collection does not publish a Dataset version, approve training or change a learning policy's pinned release. File-backed Datasets retain collected conversations alongside the imported source rows. Supported Chat attachments keep their original file references and the context actually sent to the model.

Use await openpond.addTask({ id: event.id, name, input }) for additional application-authored tasks. Do not submit the same model request again merely to collect its conversation. A Dataset producer credential cannot invoke serving.

Python Client and AsyncClient provide corresponding named Model handles. For React, keep credentials on an authenticated backend. See the SDK guide and maintained product integration examples for streaming, supported Vercel versions and capability checks.

Read exact serving state with the CLI

API and CLI clients retain the explicit result-save operation. For a saved training Model, put the exact current IDs from its saved version in selection.json:

json
{ "modelProjectId": "<model-project-id>", "versionHash": "<saved-version-content-hash>" }
bash
openpond training hosted-model-serving selection.json --team-id "$TEAM_ID" --json

This read does not activate capacity. Use each modification's own serving operation for an Engram, Decision or neuron-training Model; see its method guide.

A status excerpt illustration, not a real deployment or qualification:

json
{ "state": "inactive", "usageBilling": { "schemaVersion": "openpond.managedServingUsageReviewResult.v1", "availability": "unavailable", "reason": "model_serving_native_profile_unqualified", "review": null, "usage": null } }

Review costs and stop

  • Review the actual offered billing contract. Existing token-priced serving and an available reviewed deployment-time offer have different terms.
  • Deployment-time terms, when qualified, disclose the rate, billable startup and idle periods, a five-minute idle timeout and an optional cumulative cap. An unavailable review cannot be bypassed by adding credits.
  • Opening or cancelling a dialog does not provision resources. Activation is a separate explicit action bound to the exact saved version and review.
  • Stop closes new admission. Wait for positive physical cleanup and final settlement; a stopped switch alone does not prove resource release.
  • After final closure, Previous serving capacity retains the deployment's final charged amount and quantity. A new activation gets a fresh review; it does not replace the old deployment's cap or charges.

Sources

Compatibility

A trained artifact must be deployed to a supported serving target with available capacity before it can receive inference traffic.

View as Markdown