christian/voice-agent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(model): wählbares LLM-Modell pro User & pro Job

Browse Source 
Mac-Worker:
- /api/summarize, /api/protocol, /api/preload akzeptieren optionales
  model-Feld → überschreibt den .env-Default pro Call, kein Restart nötig
- /api/models listet die lokal verfügbaren Ollama-Modelle + Default

LXC:
- User.default_model + Job.model (Migration für SQLite)
- /api/auth/me liefert default_model, PATCH speichert ihn sofort
- /api/mac/models proxiet die Liste an die UI
- Pipeline reicht job.model an Mac-Worker durch; leer = Worker-Default
- create_job & reprocess akzeptieren model-Override
- Snapshot-Meta nutzt job.model (genauer als /health-Snapshot)

UI:
- Header-Dropdown "Modell:" neben "Profil:" — sofort-speichert User-Default
- Jeder Job zeigt sein Modell als Tag (lila); leer bedeutet Worker-Default
- Reprocess öffnet einen Modal-Picker (statt simplem confirm), Auswahl
  aller verfügbaren Modelle inkl. "Worker-Default"

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
root
2026-05-16 16:48:11 +00:00
co-authored by Claude Opus 4.7
parent dbf7bca6c7
commit c6164d946c
13 changed files with 386 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mac-worker/app/core/ollama_client.py
+41 -9
View File
@@ -52,7 +52,15 @@ def render(template: str, **vars: str) -> str:
return out
async def _chat(prompt: str) -> str:
def _resolve_model(model: str | None) -> str:
"""Wählt das tatsächlich zu verwendende Ollama-Modell.
Leerer oder fehlender `model`-Parameter → Worker-Default aus .env.
"""
return (model or "").strip() or settings.ollama_model
async def _chat(prompt: str, model: str | None = None) -> str:
"""Ruft Ollama im Streaming-Modus auf.
Streaming hat zwei wichtige Vorteile:
@@ -64,15 +72,16 @@ async def _chat(prompt: str) -> str:
`keep_alive`: das Modell bleibt 30 min im RAM zwischen Calls,
damit nachfolgende Anfragen nicht erneut die Ladezeit zahlen.
"""
use_model = _resolve_model(model)
url = f"{settings.ollama_url.rstrip('/')}/api/generate"
payload = {
"model": settings.ollama_model,
"model": use_model,
"prompt": prompt,
"stream": True,
"keep_alive": "30m",
"options": {"temperature": 0.2},
}
log.info("Ollama generate model=%s prompt_len=%d (streaming)", settings.ollama_model, len(prompt))
log.info("Ollama generate model=%s prompt_len=%d (streaming)", use_model, len(prompt))
timeout = httpx.Timeout(
connect=15.0,
read=settings.ollama_timeout, # pro Chunk — großzügig wegen initial Modell-Load
@@ -100,15 +109,17 @@ async def _chat(prompt: str) -> str:
return "".join(chunks)
async def summarize(transcript: str, title: str = "", profile_name: str | None = None) -> dict[str, Any]:
async def summarize(
transcript: str, title: str = "", profile_name: str | None = None, model: str | None = None,
) -> dict[str, Any]:
profile = profiles.get(profile_name)
prompt = render(
profile.summary_prompt,
title=title or "(ohne Titel)",
transcript=_truncate(transcript),
)
log.info("Summarize profile=%s", profile.name)
raw = await _chat(prompt)
log.info("Summarize profile=%s model=%s", profile.name, _resolve_model(model))
raw = await _chat(prompt, model=model)
try:
return _parse_json(raw)
except Exception: # noqa: BLE001
@@ -117,7 +128,8 @@ async def summarize(transcript: str, title: str = "", profile_name: str | None =
async def make_protocol(
transcript: str, summary: dict, title: str = "", profile_name: str | None = None
transcript: str, summary: dict, title: str = "", profile_name: str | None = None,
model: str | None = None,
) -> dict:
profile = profiles.get(profile_name)
prompt = render(
@@ -126,8 +138,8 @@ async def make_protocol(
summary_json=json.dumps(summary, ensure_ascii=False),
transcript=_truncate(transcript),
)
log.info("Protocol profile=%s", profile.name)
raw = await _chat(prompt)
log.info("Protocol profile=%s model=%s", profile.name, _resolve_model(model))
raw = await _chat(prompt, model=model)
try:
return _parse_json(raw)
except Exception: # noqa: BLE001
@@ -143,3 +155,23 @@ async def health() -> dict:
r = await c.get(url)
r.raise_for_status()
return r.json()
async def list_models() -> list[dict]:
"""Liefert die lokal gepullten Ollama-Modelle inkl. Größe und Mod-Zeit.
Antwort-Form pro Eintrag: {"name", "size", "modified_at"} (Größe in Bytes).
"""
url = f"{settings.ollama_url.rstrip('/')}/api/tags"
async with httpx.AsyncClient(timeout=10) as c:
r = await c.get(url)
r.raise_for_status()
data = r.json()
out: list[dict] = []
for m in data.get("models", []):
out.append({
"name": m.get("name") or m.get("model") or "",
"size": m.get("size", 0),
"modified_at": m.get("modified_at", ""),
})
return [m for m in out if m["name"]]
mac-worker/app/main.py
+27 -5
View File
@@ -68,6 +68,7 @@ class SummarizeIn(BaseModel):
transcript: str
title: str = ""
profile: str | None = None
model: str | None = None
class ProtocolIn(BaseModel):
@@ -75,6 +76,7 @@ class ProtocolIn(BaseModel):
summary: dict
title: str = ""
profile: str | None = None
model: str | None = None
class DocxIn(BaseModel):
@@ -89,6 +91,7 @@ class PdfIn(BaseModel):
class PreloadIn(BaseModel):
profile: str | None = None
model: str | None = None
@app.get("/health")
@@ -114,6 +117,21 @@ def list_profiles():
return {"profiles": profiles.list_all()}
@app.get("/api/models")
async def list_models():
"""Liste der lokal verfügbaren Ollama-Modelle + Default aus der Worker-.env.
Das LXC-Frontend zeigt davon ein Dropdown — leerer Wert beim Job-Start =
der Default hier wird genommen.
"""
try:
models = await ollama_client.list_models()
except Exception as e: # noqa: BLE001
log.warning("list_models: Ollama nicht erreichbar: %s", e)
return {"models": [], "default": settings.ollama_model, "error": str(e)}
return {"models": models, "default": settings.ollama_model}
@app.post("/api/transcribe")
async def transcribe(audio: UploadFile = File(...), language: str = Form("de")):
if not audio.filename:
@@ -135,7 +153,9 @@ async def transcribe(audio: UploadFile = File(...), language: str = Form("de")):
async def summarize(payload: SummarizeIn):
if not payload.transcript.strip():
raise HTTPException(400, "Leerer Transkript")
return await ollama_client.summarize(payload.transcript, title=payload.title, profile_name=payload.profile)
return await ollama_client.summarize(
payload.transcript, title=payload.title, profile_name=payload.profile, model=payload.model
)
@app.post("/api/protocol")
@@ -143,7 +163,8 @@ async def protocol(payload: ProtocolIn):
if not payload.transcript.strip():
raise HTTPException(400, "Leerer Transkript")
return await ollama_client.make_protocol(
payload.transcript, payload.summary, title=payload.title, profile_name=payload.profile
payload.transcript, payload.summary, title=payload.title, profile_name=payload.profile,
model=payload.model,
)
@@ -176,10 +197,11 @@ async def preload(payload: PreloadIn):
(häufige Ursache von 'Server disconnected without sending a response').
"""
profile_obj = profiles.get(payload.profile)
log.info("preload model=%s profile=%s", settings.ollama_model, profile_obj.name)
use_model = (payload.model or "").strip() or settings.ollama_model
log.info("preload model=%s profile=%s", use_model, profile_obj.name)
url = f"{settings.ollama_url.rstrip('/')}/api/generate"
body = {
"model": settings.ollama_model,
"model": use_model,
"prompt": "ping",
"stream": False,
"keep_alive": "30m",
@@ -195,7 +217,7 @@ async def preload(payload: PreloadIn):
log.warning("preload failed: %s", e)
raise HTTPException(503, f"Ollama preload fehlgeschlagen: {e}") from e
log.info("preload ok load_duration=%dms", int(data.get("load_duration", 0) / 1_000_000))
return {"ok": True, "model": settings.ollama_model, "load_duration_ns": data.get("load_duration", 0)}
return {"ok": True, "model": use_model, "load_duration_ns": data.get("load_duration", 0)}
@app.get("/api/jobs/{job_id}/log")