@interview-sdk/adapter-openai

OpenAI provider adapter for @interview-sdk/core, built on the official openai SDK's Responses API (client.responses.create) — OpenAI's current recommended interface, though the older Chat Completions API remains fully supported indefinitely if you need it instead.

Install

npm install @interview-sdk/core @interview-sdk/adapter-openai

Usage

import { OpenAIAdapter } from '@interview-sdk/adapter-openai';

const adapter = new OpenAIAdapter({ apiKey: process.env.OPENAI_API_KEY });
// <InterviewWidget adapter={adapter} mode="client" ... />
// or: new ServerAnswerProcessor({ questions, rubric, adapter })

(AdapterRegistry from @interview-sdk/core is a separate, optional utility for looking an adapter up by string id at runtime — most apps just pass the adapter directly like above.)

OpenAIAdapter accepts an optional model (defaults to gpt-5.4-mini — a cost-effective choice for structured scoring tasks; pass gpt-5.5 for the frontier model) and an optional pre-configured client for testing or custom base URLs.

Behavior

• Maps AIMessage[] directly onto the Responses API's input array — system/user/assistant roles all map 1:1.
• Requests text: { format: { type: 'json_object' } } when responseFormat: 'json' is set. Deliberately not json_schema (OpenAI's stricter, preferred structured-output mode): core's evaluation/follow-up response shapes carry dynamic keys (developer-defined rubric dimension ids), which a fixed JSON Schema can't express.
• Normalizes every OpenAI SDK exception onto the shared Provider*Error taxonomy exported by @interview-sdk/core, the same as every other adapter, so withRetry and FailoverAdapter work without any OpenAI-specific knowledge.
• Context-length overflow and a deprecated/retired model id both surface as plain 4xx errors from OpenAI with no distinct error type — context-length is detected heuristically from the message text; a missing/retired model normalizes to ProviderInvalidRequestError.

Verified against

[email protected], current as of 2026-07-04 — model names, the Responses vs. Chat Completions guidance, and the error taxonomy shift quickly; re-check before assuming this stays accurate for long.