fix: think=false en Open WebUI (el SYSTEM /no_think no basta)

El SYSTEM /no_think del Modelfile no frena a Qwen3 en tareas complejas
(la pregunta de code generaba 700+ tokens de razonamiento a pesar del
SYSTEM y del soft-switch en el prompt).

Via oficial: la API de Ollama 0.9+ acepta 'think': false en /api/chat,
y Open WebUI 0.11.3 mapea el param del modelo directamente al payload
(payload.py: 'think': lambda x: x).

DEFAULT_MODEL_PARAMS ahora: {"function_calling":"legacy","think":false}

Medido via API con think:false: 208 tokens, sin campo thinking, contenido
limpio (vs 689-725 con thinking).
This commit is contained in:
hermes committed 2026-09-12 23:06:22 +02:00
1 parent c68fe80484
commit d6b6323ea7
2 files changed
+2 -2

No files matched your search

+1 -1
View File
@@ -48,7 +48,7 @@ services:
ENABLE_WEB_SEARCH: "false" ENABLE_WEB_SEARCH: "false"
ENABLE_DIRECT_CONNECTIONS: "false" ENABLE_DIRECT_CONNECTIONS: "false"
CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS: "1" CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS: "1"
DEFAULT_MODEL_PARAMS: '{"function_calling": "legacy"}' DEFAULT_MODEL_PARAMS: '{"function_calling": "legacy", "think": false}'
depends_on: depends_on:
ollama: ollama:
condition: service_healthy condition: service_healthy
+1 -1
View File
@@ -78,7 +78,7 @@ services:
WEBUI_AUTH: "false" WEBUI_AUTH: "false"
DEFAULT_LOCALE: "es" DEFAULT_LOCALE: "es"
CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS: "1" CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS: "1"
DEFAULT_MODEL_PARAMS: '{"function_calling": "legacy"}' DEFAULT_MODEL_PARAMS: '{"function_calling": "legacy", "think": false}'
depends_on: depends_on:
ollama: ollama:
condition: service_healthy condition: service_healthy