Configuration

Everything below is set once on the VernLLM instance and applies to every call() unless overridden per request. Nothing is required beyond client and model — every other option has a sensible default.

const llm = new VernLLM({
  client: openai,
  model: 'gpt-4o',
  maxRetries: 3,
  timeoutMs: 10_000,
  baseDelayMs: 500,
  defaultMaxTokens: 1000,
  nonRetryableStatus: [400, 401, 403],
  circuitBreaker: true,
  debug: false,
});

Required

client

An LLMClient — an OpenAI SDK instance, or the result of wrapping another provider with an adapter (fromAnthropic, fromGemini, fromBedrock, fromFetch, etc). See Adapters.

model

The default model ID for calls that do not override it via call({ model: '...' }).

Retry & timeout

Option	Default	Notes
`maxRetries`	`1`	Retries after the first attempt, so 2 attempts total by default. `maxRetries: 3` means up to 4 attempts.
`timeoutMs`	`25000`	Per attempt timeout, not a total call timeout. Each retry gets its own fresh `timeoutMs` window.
`baseDelayMs`	`500`	Base for exponential backoff between retries. Actual delay grows per attempt with jitter.
`nonRetryableStatus`	`[400, 401, 403]`	HTTP status codes that fail immediately instead of retrying. Bad request, unauthorized, and forbidden errors are treated as non transient.

maxRetries defaults to 1, not 3. See Error Handling for exactly which failures get retried.

Request defaults

Option	Default	Notes
`defaultMaxTokens`	`1000`	Used for any call that does not set `maxTokens` itself.
`parseJson`	`defaultParseJson`	Custom parser for non standard JSON responses. It should return `undefined` when parsing fails, which signals a parse error to VernLLM.

Caching

Option	Default	Notes
`cache`	`new InMemoryCacheAdapter()`	Used only by `cachedCall` and `cachedLLMCall`, not plain `call()`. The default in memory cache is not shared across processes. Use a shared adapter such as Redis or Upstash for multi process deployments. See Caching.

Observability

Option	Default	Notes
`onUsage`	none	Called after every successful call with token counts, but only if the provider actually returned usage data on the response.
`logger`	console based `ConsoleLogger`	Implement the `Logger` interface to route output elsewhere, such as pino or winston. See Pluggable Logger.
`debug`	`NODE_ENV !== 'production'`	Controls `logger.debug` calls such as raw model output and retry attempts. `warn` and `error` always fire regardless of this flag.

Circuit breaker

Option	Default	Notes
`circuitBreaker`	disabled	Pass `true` for default circuit breaker settings, or `{ threshold, cooldownMs }` to tune them. See Circuit Breaker.

Per call defaults and overrides

A handful of options are set per call() rather than on the instance. Two have their own defaults worth knowing:

Option	Default	Notes
`temperature`	`0.2`	Low by design. This library targets structured and deterministic extraction more than open ended generation. Override freely per call.
`jsonMode`	`true`	Responses are parsed as JSON by default. Set `jsonMode: false` to get the raw string back with no parsing.

The rest, including model, maxTokens, reasoningEffort, schema, jsonSchema, signal, and requestId, have no call level default beyond falling back to the instance level option or being undefined when omitted.