mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-17 23:51:30 +00:00
1ebd3c656d
#35855 lets classifier_llm_config.system_prompt replace the classifier's entire system role, with its own classifier_fallback='default_model' escape hatch since the heuristic cannot score a repurposed taxonomy. This branch's classification_prompt replaces only the rubric's opening instructions and always keeps the per-tier bullets and the trust boundary. Both knobs now live on the unified builder: custom_prompt returns verbatim, preamble composes with the active tier entries, and the exception path prefers a custom set's fallback_tier before consulting classifier_fallback. Three combinations are rejected at config write because each pair claims the same seat: system_prompt with classification_prompt, system_prompt with tier_definitions, and classifier_fallback='default_model' with tier_definitions. The default-prompt endpoint keeps serving labeled rubrics through the new canonical_rubric_entries helper, and RoutingDecisionCause carries both branches' new causes. |
||
|---|---|---|
| .. | ||
| litellm-dashboard | ||
| Dockerfile | ||
| nginx.conf | ||