Configuration parameters and policies

Configurations define use-case presets with model selection and parameters.

Configuration list with model assignments — Configuration list showing linked model, selection mode and parameters.

Required

identifier (config)

Type: string
Required: true

Unique slug (e.g., blog-summarizer).

name (config)

Type: string
Required: true

Display name (e.g., Blog Post Summarizer).

model

Type: reference

Stored as model_uid. Needed for execution in fixed mode; criteria mode resolves a model without this reference.

system_prompt

Type: text

Optional system message that sets the AI's behavior.

Optional

temperature

Type: float
Default: 0.7

Creativity (0.0 = deterministic, 2.0 = creative).

max_tokens (config)

Type: integer
Default: 1000

Maximum response length in tokens. Set 0 to inherit the model's max_output_tokens. Precedence: per-call option > configuration value (> 0) > model max_output_tokens (> 0) > provider default.

A reasoning model of OpenAI counts its thinking against this limit, so nr-llm sends at least 16,000 tokens for such a model unless the call switches thinking off. The value bounds the length of an answer only on a model that does not think. See ADR-241.

top_p

Type: float
Default: 1.0

Nucleus sampling (0.0-1.0).

frequency_penalty

Type: float
Default: 0.0

Reduces word repetition (-2.0 to 2.0).

presence_penalty

Type: float
Default: 0.0

Encourages topic diversity (-2.0 to 2.0).

fallback_chain

Type: JSON (text column)
Default: (empty)

JSON object with a single key, configurationIdentifiers, whose value is the ordered list of other configuration identifiers to retry against when the primary fails with a retryable error (connection error, HTTP 5xx, or HTTP 429 rate-limit). Non-retryable errors bubble up unchanged. Streaming requests do not trigger fallback after visible output. Before the first chunk, eligible failures can fall back; see Streaming support.

Example payload:

{"configurationIdentifiers": ["claude-sonnet", "ollama-local"]}

Identifiers are matched case-insensitively; leave empty to disable fallback. See Fallback chain.

answer_cache_ttl

Type: int (seconds)
Default: 0 (off)

How long the same request to this configuration gets the same answer again without a provider call, at most 86400 (one day). It applies to a plain chat or completion. A call of an agent run, a tool call, a call bound to a response schema and a streamed answer are never answered from the cache. Only a complete text answer is stored, in JSON mode only one that parses. Evaluation runs always ask the model, and a caller can decline the cache for one call with ChatOptions::withoutAnswerCache().

The entry is the answer verbatim, so it is stored and served only where the content privacy level is full, globally and for the provider that served the answer (ADR-242). Below that the field has no effect.

The key covers the configuration and its last edit, the messages after skill injection, the per-call options, the site and the language. Saving the configuration therefore starts afresh, and flushing all caches empties it (cache nrllm_responses, tag nrllm_answers). A served answer consumes no tokens, is not charged against a budget and counts as a cache hit in telemetry. See ADR-249.

system_prompt_data_class

Type: select
Default: (empty — undeclared)

The sensitivity ceiling for confval-config-system-prompt. Declaring one refuses any send whose serving model sits in a weaker trust zone than the class allows, the same axis snippets and skills carry (ADR-144, ADR-155).

Empty means no statement was made and constrains nothing, so existing configurations behave exactly as before. The class describes the text: a configuration whose system prompt is blank declares nothing whatever this field says.

Enforcement follows the instance-wide tools.dataClassEnforcement switch — see The four keys.