nr-llm
EN DE

ADR 066 · Accepted

Criteria-Mode Configurations Resolve in the Service Layer

Context

ADR-055 and ADR-056 introduced LlmConfiguration records with two selection modes: fixed (a direct model_uid relation) and criteria (a stored ModelSelectionCriteria JSON; model_uid = 0, the concrete model chosen at call time). ModelSelectionService::resolveModel() was the resolver — but it had no production caller. Every *ForConfiguration() entry point (embedForConfiguration, chatWithConfiguration, chatWithToolsForConfiguration, completeWithConfiguration, streamChatWithConfiguration) obtained its adapter through getAdapterFromConfiguration(), which read $configuration->getLlmModel() — a plain getter — directly. For a criteria-mode record that relation is null, so the call threw ProviderException: Configuration "…" has no model assigned.

This surfaced live: nr_ai_search's embeddingConfiguration / chatConfiguration presets are criteria-mode, so the entire retrieval path failed the moment it reached the provider adapter.

Decision

getAdapterFromConfiguration() — the single adapter choke point for every *ForConfiguration() path — resolves the model through ModelSelectionService::resolveModel($configuration), which returns the directly configured model unchanged for fixed mode and selects from the stored criteria for criteria mode. It still throws the same ProviderException (code 1735300100) when resolution yields no model.

The resolved model is not written back onto the configuration. The configuration is a repository-managed Extbase entity; calling setLlmModel() would mark it dirty and Extbase would persist model_uid at request end, silently converting a criteria-mode record into a fixed-mode one. Per-model cost analytics for criteria configs (UsageMiddleware reads getLlmModel() directly) therefore remain a separate, non-destructive follow-up.

Consequences

  • Criteria-mode configurations work across embed / chat / tools / complete / stream without the caller pre-resolving a model.

  • ModelSelectionServiceInterface is a trailing, nullable constructor dependency of LlmServiceManager (autowired); when absent the method falls back to the raw getter, so existing 5-argument constructions keep compiling.

  • Cost analytics for criteria-mode configs fall back to the provider-reported model id (no per-model DB pricing) until the follow-up threads the resolved model through the call metadata.

See PR #372.

← All ADRs