feat(llm): Live-Parameter top_p + Live/Restart-Kennzeichnung — Plan-Schritt 4

- local_llm_top_p (Default 0.9) durch Config -> runtime_config (RUNTIME_SETTABLE,
  live ohne Neustart) -> local-openai-compatible Payload (top_p).
- Admin-Einstellungen: top_p als Slider (0–1); Hints "wirkt sofort (kein Neustart)"
  für temperature/top_p.
- Doku: top_p + Klarstellung Live-Parameter (temp/top_p/max_tokens) vs.
  Startup-Wert Kontextfenster (OLLAMA_CONTEXT_LENGTH / llama.cpp -c).

Verifiziert: top_p im Payload, Live-Config-Round-Trip (PUT/DELETE). 165 grün.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

BEDIENUNGSANLEITUNG.md

@@ -1027,6 +1027,12 @@ Diese Settings gelten nur für den Provider `local-openai-compatible`.
 | `LOCAL_LLM_SYSTEM_PROMPT` | Sprach-Prompt | Kurze, vorlesbare Antworten |
 | `LOCAL_LLM_MAX_TOKENS` | `0` (Server-Limit) | Optionaler Deckel, z. B. `256` |
 | `LOCAL_LLM_TEMPERATURE` | `0.3` | Sampling-Temperatur |
+| `LOCAL_LLM_TOP_P` | `0.9` | Nucleus-Sampling (0.0–1.0) |
+
+> Temperatur, Top-p und Max-Tokens sind **Live-Parameter**: im Admin-Panel →
+> Einstellungen änderbar und **ohne Neustart** sofort wirksam (pro Anfrage gesendet).
+> Das **Kontextfenster** ist dagegen ein Startup-Wert (Ollama: `OLLAMA_CONTEXT_LENGTH`,
+> llama.cpp: `-c`) und erfordert einen Backend-Neustart.
 
 Messung (Qwen3-35B, `va_llm`): Reasoning an → **5,5 s / 1433 Zeichen**;
 Reasoning aus + Sprach-Prompt → **0,7 s / ~190 Zeichen**.

app/config.py

@@ -131,6 +131,7 @@ class Settings(BaseSettings):
     local_llm_disable_reasoning: bool = True  # Qwen3 /no_think: spart die Denkphase
     local_llm_max_tokens: int = 0             # 0 = serverseitiges Limit (-n)
     local_llm_temperature: float = 0.3
+    local_llm_top_p: float = 0.9              # Nucleus-Sampling (0.0–1.0)
     faster_whisper_model: str = "base"        # tiny|base|small|medium|large-v3
     faster_whisper_device: str = "auto"       # auto|cpu|cuda
     faster_whisper_compute_type: str = "default"  # default|int8|float16|int8_float16

app/dependencies.py

@@ -63,6 +63,7 @@ LLM_REGISTRY = {
         disable_reasoning=s.local_llm_disable_reasoning,
         max_tokens=s.local_llm_max_tokens,
         temperature=s.local_llm_temperature,
+        top_p=s.local_llm_top_p,
     ),
 }
 

app/providers/llm/local_openai_compatible.py

@@ -20,6 +20,7 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         disable_reasoning: bool = True,
         max_tokens: int = 0,
         temperature: float = 0.3,
+        top_p: float = 0.9,
     ):
         self.base_url = base_url.rstrip("/")
         self.api_key = api_key
@@ -28,6 +29,7 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         self.disable_reasoning = disable_reasoning
         self.max_tokens = max_tokens
         self.temperature = temperature
+        self.top_p = top_p
 
     def _build_messages(
         self, text: str, history: list[dict] | None, language: str | None = None
@@ -66,6 +68,7 @@ class LocalOpenAICompatibleLLM(LLMProvider):
             "model": self.model,
             "messages": self._build_messages(text, history, language=language),
             "temperature": self.temperature,
+            "top_p": self.top_p,
         }
         if stream:
             payload["stream"] = True

app/runtime_config.py

@@ -26,7 +26,8 @@ RUNTIME_SETTABLE: dict[str, tuple[str, str, str]] = {
     "openrouter_tts_voice":          ("TTS-Stimme (OpenRouter)",       "str",   "z.B. Zephyr, Puck, Kore"),
     "piper_voice":                   ("TTS-Stimme (piper)",            "str",   "z.B. de_DE-thorsten-high"),
     "local_llm_system_prompt":       ("Systemprompt (lokal)",          "str",   "Freier Text"),
-    "local_llm_temperature":         ("Temperatur (lokal)",            "float", "0.0–2.0"),
+    "local_llm_temperature":         ("Temperatur (lokal)",            "float", "0.0–2.0 — wirkt sofort (kein Neustart)"),
+    "local_llm_top_p":               ("Top-p (lokal)",                 "float", "0.0–1.0 — wirkt sofort (kein Neustart)"),
     "local_llm_max_tokens":          ("Max. Tokens (lokal)",           "int",   "0 = kein Limit"),
     "tts_normalize_level":           ("TTS-Normalisierung",            "str",   "auto | full | light | off"),
     "audio_stream_default":          ("Audio-Streaming Standard",      "bool",  "true | false"),

app/web/app.js

@@ -1384,6 +1384,7 @@ const FIELD_META = {
   piper_voice:                     { ui: "combo",   opts: ["de_DE-thorsten-high","en_US-ryan-high","en_US-lessac-high","en_GB-cori-high","es_ES-sharvard-medium","fr_FR-siwis-medium","it_IT-paola-medium","nl_NL-mls-medium","ru_RU-irina-medium","zh_CN-huayan-medium"], test: "tts", testProvider: "piper" },
   local_llm_system_prompt:         { ui: "textarea", test: "llm" },
   local_llm_temperature:           { ui: "range",   min: 0, max: 2, step: 0.1, test: "llm" },
+  local_llm_top_p:                 { ui: "range",   min: 0, max: 1, step: 0.05, test: "llm" },
   local_llm_max_tokens:            { ui: "number",  min: 0, test: "llm" },
   tts_normalize_level:             { ui: "select",  opts: ["auto","full","light","off"], test: "tts" },
   audio_stream_default:            { ui: "select",  opts: ["true","false"],
@@ -1398,7 +1399,7 @@ const FIELD_GROUPS = [
   { label: "Provider & Sprache",        keys: ["default_stt_provider","default_llm_provider","default_tts_provider","default_language"] },
   { label: "OpenRouter-Modelle & Stimmen", keys: ["openrouter_llm_model","openrouter_tts_model","openrouter_tts_voice"] },
   { label: "Piper TTS",                 keys: ["piper_voice"] },
-  { label: "Lokales LLM",              keys: ["local_llm_system_prompt","local_llm_temperature","local_llm_max_tokens"] },
+  { label: "Lokales LLM",              keys: ["local_llm_system_prompt","local_llm_temperature","local_llm_top_p","local_llm_max_tokens"] },
   { label: "TTS-Verarbeitung",         keys: ["tts_normalize_level","audio_stream_default"] },
   { label: "Gedächtnis",               keys: ["memory_extraction_enabled","memory_extraction_every_n_turns"] },
   { label: "Limits",                    keys: ["daily_request_limit"] },

app/web/index.html

@@ -250,6 +250,6 @@
     <div id="status" class="w-full max-w-3xl mx-auto mt-1 min-h-[1rem] text-xs text-slate-500 dark:text-slate-400"></div>
   </footer>
 
-  <script src="/app.js?v=30"></script>
+  <script src="/app.js?v=31"></script>
 </body>
 </html>