Option classes 

Shared call controls 

Provider-specific option support still depends on the adapter and operation. The service forwards call metadata separately; toArray() is not a complete serialization of all policy and actor context.

ChatOptions 

class ChatOptions

Fully qualified name

\Netresearch\NrLlm\Service\Option\ChatOptions

Typed options for chat operations.

static factual()

Create options optimized for factual responses (temperature: 0.1).

returns

ChatOptions

static creative()

Create options for creative content (temperature: 1.2).

returns

ChatOptions

static balanced()

Create balanced options (temperature: 0.7).

returns

ChatOptions

static json()

Create options for JSON output format.

returns

ChatOptions

static code()

Create options optimized for code generation.

returns

ChatOptions

withTemperature(float $temperature): self

Set temperature (0.0 - 2.0).

withMaxTokens(int $maxTokens): self

Request a maximum output-token budget. For recognized OpenAI reasoning models, the adapter raises values below 16,000 to 16,000 unless the applied reasoning effort is None. Thinking consumes that budget too. This option is therefore not a universal hard output or cost cap.

withTopP(float $topP): self

Set nucleus sampling parameter.

withFrequencyPenalty(float $penalty): self

Set frequency penalty (-2.0 to 2.0).

withPresencePenalty(float $penalty): self

Set presence penalty (-2.0 to 2.0).

withSystemPrompt(string $prompt): self

Set system prompt.

withResponseFormat(string $format): self

Request an output format: text, json or markdown. Adapter support differs. The current Claude adapter does not implement bare response_format=json. Its object-schema path uses a forced tool instead. For validated structured output, use completeStructured(); local schema validation remains authoritative (ADR-128).

withResponseSchema(array $schema): self

Attach a strict-subset JSON schema (ADR-126) the provider should enforce natively where it can. Set automatically by completeStructured(); the local strict validation remains authoritative either way.

withStopSequences(array $sequences): self

Set stop sequences the model must not generate past. The OpenAI provider does not send them to a reasoning model (GPT-6, GPT-5, o-series), because OpenAI refuses the parameter there; the answer is then not cut at the sequence.

withProvider(string $provider): self

Set an explicit registered provider key on the unbound call path. This is not a substitute for a database configuration identifier.

withModel(string $model): self

Set specific model.

withReasoningEffort(ReasoningEffort $effort): self

Ask a reasoning model for an amount of thinking: one of the ReasoningEffort cases None, Minimal, Low, Medium, High, XHigh and Max (ADR-204). It wins over the coarse think switch. The adapter maps it to a bundled model profile, not a live provider capability query. Profiles cover the declared family rules. GPT-6.1 Sol, GPT-6 Astra and their dated snapshots accept low/medium/high/xhigh and map None and think=false to Low. GPT-6 Sol, GPT-6 Luna and GPT-6.1 Terra accept none to xhigh. No GPT-6 model accepts Max; for these five models a request for it is clamped to XHigh (measured against the OpenAI API on 2026-11-09). Any other gpt-6 identifier has no effort scale, so it receives no effort at all, Max included, and unknown gateway identifiers receive no guessed effort scale.

nrllm_reasoning_effort records an applied effort when available. Its absence does not prove the provider has no effort setting: the model may have no local profile or may use the provider default.

toArray(): array

Convert to array format.

ToolOptions 

Extends ChatOptions. toolChoice and parallelToolCalls default to null (no explicit override); captureRaw defaults to false. ToolOptions::auto() selects automatic tool choice, but does not choose a default provider. Use the configuration-bound service or set an explicit registered provider key. withTaskUid() supplies task attribution; withTextSink() attaches a live text callback on supported paths.

EmbeddingOptions 

model, provider and dimensions default to null. cacheTtl defaults to 86400 seconds. Use withCacheTtl() or noCache() to change caching, withDimensions() to request a supported vector size, and the compact() / highPrecision() factories for the offered presets. Model limits and adapter support still apply; see EmbeddingService.

VisionOptions 

detailLevel, maxTokens, temperature, provider and model default to null. withDetailLevel() asks for the supported image-detail mode; it does not make unsupported image input available on another adapter. Budget attribution uses the shared controls. See VisionService.

TranslationOptions 

preserveFormatting defaults to true. Provider/model/configuration, languages' glossary/context controls, tagHandling and cacheTtl are unset by default; polishing is opt-in. withConfiguration() selects the translator configuration on the specialized path, not automatically the polisher's configuration. See Translation options for the full method-specific options and LLM and specialized translator paths for result types.