feat(diarization-ui): multi-provider LLM support via OpenAI-compatible API

- Replace Ollama-native API (/api/generate) with OpenAI-compatible
  /v1/chat/completions — works with Ollama, OpenAI, Groq, LM Studio,
  LocalAI, and any other OpenAI-compatible endpoint
- New providers table: name, base_url, optional API key, default flag
- Settings page: add/delete/set-default providers; rename LLM params
  from ollama_* to llm_* settings keys
- Per-prompt: choose provider from dropdown, model load button fetches
  from that provider's /v1/models
- <think>…</think> tags extracted for live thinking display (works with
  Ollama thinking models and DeepSeek-R1 style responses)
- Existing OLLAMA_BASE_URL env var seeds the initial provider on first
  start for backward compat

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -5,6 +5,7 @@ import email.mime.multipart
 import email.mime.text
 import json
 import os
+import re
 import smtplib
 import sqlite3
 import threading
@@ -19,10 +20,10 @@ from fastapi import FastAPI, File, Form, HTTPException, UploadFile
 from fastapi.responses import HTMLResponse, PlainTextResponse, Response, JSONResponse, RedirectResponse
 
 API_BASE = os.getenv("API_BASE", "http://gx10.aquantico.lan:8093").rstrip("/")
-OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://gx10.aquantico.lan:11434").rstrip("/")
+# Kept for backward-compat seeding of the initial provider on first start
-OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen3.5:9b")
+_SEED_PROVIDER_URL = os.getenv("OLLAMA_BASE_URL", "http://gx10.aquantico.lan:11434").rstrip("/")
-OLLAMA_NUM_PREDICT = int(os.getenv("OLLAMA_NUM_PREDICT", "16384"))
+_SEED_MODEL = os.getenv("OLLAMA_MODEL", "qwen3.5:9b")
-OLLAMA_THINK = os.getenv("OLLAMA_THINK", "true").lower() in ("1", "true", "yes")
+_SEED_NUM_PREDICT = int(os.getenv("OLLAMA_NUM_PREDICT", "16384"))
 DB_PATH = os.getenv("DB_PATH", "/data/ui.db")
 
 app = FastAPI(title="Diarization UI")
@@ -156,6 +157,35 @@ def init_db():
             except Exception:
                 pass
 
+        try:
+            c.execute(
+                """
+                CREATE TABLE IF NOT EXISTS providers (
+                  id INTEGER PRIMARY KEY AUTOINCREMENT,
+                  name TEXT NOT NULL,
+                  base_url TEXT NOT NULL,
+                  api_key TEXT,
+                  is_default INTEGER NOT NULL DEFAULT 0,
+                  created_at TEXT NOT NULL
+                )
+                """
+            )
+        except Exception:
+            pass
+
+        try:
+            c.execute("ALTER TABLE prompts ADD COLUMN provider_id INTEGER REFERENCES providers(id)")
+        except Exception:
+            pass
+
+        # Seed default provider from env if table is empty
+        count = c.execute("SELECT COUNT(*) FROM providers").fetchone()[0]
+        if count == 0:
+            c.execute(
+                "INSERT INTO providers(name, base_url, api_key, is_default, created_at) VALUES (?,?,?,1,?)",
+                ("Ollama (Standard)", _SEED_PROVIDER_URL, None, now_iso()),
+            )
+
         try:
             c.execute(
                 """
@@ -203,6 +233,32 @@ def set_setting(key: str, value: str):
         c.execute("INSERT OR REPLACE INTO settings(key, value) VALUES (?,?)", (key, value))
 
 
+def _get_default_provider() -> dict | None:
+    with db() as c:
+        row = c.execute("SELECT * FROM providers WHERE is_default=1 LIMIT 1").fetchone()
+        if not row:
+            row = c.execute("SELECT * FROM providers ORDER BY id LIMIT 1").fetchone()
+    return dict(row) if row else None
+
+
+def _get_provider(provider_id: int | None) -> dict | None:
+    if not provider_id:
+        return _get_default_provider()
+    with db() as c:
+        row = c.execute("SELECT * FROM providers WHERE id=?", (provider_id,)).fetchone()
+    return dict(row) if row else _get_default_provider()
+
+
+def _split_thinking(text: str) -> tuple[str, str]:
+    """Split <think>…</think> prefix from the rest of the content."""
+    if text.startswith("<think>"):
+        end = text.find("</think>")
+        if end == -1:
+            return text[7:], ""
+        return text[7:end], text[end + 8:].lstrip("\n")
+    return "", text
+
+
 def _send_email(to_addrs: list[str], subject: str, body_md: str, attachment_name: str):
     host = get_setting("smtp_host", "")
     if not host:
@@ -480,68 +536,98 @@ def _process_analysis_job(job_id: int):
         if not (doc["content_md"] or "").strip():
             raise RuntimeError("Dokument hat keinen Inhalt – bitte zuerst das Transkript prüfen")
 
-        user_extra = (j.get("user_prompt") or "").strip()
-        llm_prompt = (
-            "Du bist ein präziser Assistent. Antworte auf Deutsch.\\n"
-            f"AUFTRAG:\\n{prm['prompt']}\\n"
-            + (f"\\nZUSATZINFOS:\\n{user_extra}\\n" if user_extra else "")
-            + f"\\nTEXT:\\n{doc['content_md']}\\n"
-        )
-
         use_default = int(prm["llm_use_default"] if prm["llm_use_default"] is not None else 1)
         if use_default:
-            model = get_setting("ollama_model", OLLAMA_MODEL)
+            provider = _get_default_provider()
-            think = get_setting("ollama_think", "true" if OLLAMA_THINK else "false") == "true"
+            model = get_setting("llm_model", _SEED_MODEL)
-            num_ctx_setting = get_setting("ollama_num_ctx", "auto")
+            num_ctx_setting = get_setting("llm_num_ctx", "auto")
-            num_predict = int(get_setting("ollama_num_predict", str(OLLAMA_NUM_PREDICT)))
+            num_predict = int(get_setting("llm_num_predict", str(_SEED_NUM_PREDICT)))
-            repeat_penalty = float(get_setting("ollama_repeat_penalty", "1.15"))
+            repeat_penalty = float(get_setting("llm_repeat_penalty", "1.15"))
-            repeat_last_n = int(get_setting("ollama_repeat_last_n", "128"))
+            repeat_last_n = int(get_setting("llm_repeat_last_n", "128"))
         else:
-            model = prm["llm_model"] or get_setting("ollama_model", OLLAMA_MODEL)
+            provider = _get_provider(prm.get("provider_id"))
-            think = (prm["llm_think"] or get_setting("ollama_think", "true" if OLLAMA_THINK else "false")) == "true"
+            model = prm["llm_model"] or get_setting("llm_model", _SEED_MODEL)
-            num_ctx_setting = prm["llm_num_ctx"] or get_setting("ollama_num_ctx", "auto")
+            num_ctx_setting = prm["llm_num_ctx"] or get_setting("llm_num_ctx", "auto")
-            num_predict = int(prm["llm_num_predict"] or get_setting("ollama_num_predict", str(OLLAMA_NUM_PREDICT)))
+            num_predict = int(prm["llm_num_predict"] or get_setting("llm_num_predict", str(_SEED_NUM_PREDICT)))
-            repeat_penalty = float(prm["llm_repeat_penalty"] or get_setting("ollama_repeat_penalty", "1.15"))
+            repeat_penalty = float(prm["llm_repeat_penalty"] or get_setting("llm_repeat_penalty", "1.15"))
-            repeat_last_n = int(prm["llm_repeat_last_n"] or get_setting("ollama_repeat_last_n", "128"))
+            repeat_last_n = int(prm["llm_repeat_last_n"] or get_setting("llm_repeat_last_n", "128"))
-        num_ctx = _estimate_num_ctx(llm_prompt) if num_ctx_setting == "auto" else int(num_ctx_setting)
+
-        _job_set(job_id, llm_prompt=f"[num_ctx={num_ctx}]\n\n{llm_prompt}")
+        if not provider:
+            raise RuntimeError("Kein LLM-Anbieter konfiguriert")
+
+        user_extra = (j.get("user_prompt") or "").strip()
+        system_msg = "Du bist ein präziser Assistent. Antworte auf Deutsch."
+        user_msg = (
+            f"AUFTRAG:\n{prm['prompt']}\n"
+            + (f"\nZUSATZINFOS:\n{user_extra}\n" if user_extra else "")
+            + f"\nTEXT:\n{doc['content_md']}\n"
+        )
+
+        ctx_hint = f"[num_ctx=auto]" if num_ctx_setting == "auto" else f"[num_ctx={num_ctx_setting}]"
+        _job_set(job_id, llm_prompt=f"{ctx_hint} [Anbieter: {provider['name']} | Modell: {model}]\n\nSYSTEM:\n{system_msg}\n\nUSER:\n{user_msg}")
+
         with _JOB_STREAM_LOCK:
             _JOB_STREAMS[job_id] = {"thinking": "", "response": ""}
-        r = requests.post(
+
-            f"{OLLAMA_BASE_URL}/api/generate",
+        headers = {"Content-Type": "application/json"}
-            json={"model": model, "prompt": llm_prompt, "stream": True, "think": think, "options": {
+        if provider.get("api_key"):
-                "num_ctx": num_ctx,
+            headers["Authorization"] = f"Bearer {provider['api_key']}"
+
+        req_body: dict = {
+            "model": model,
+            "messages": [
+                {"role": "system", "content": system_msg},
+                {"role": "user", "content": user_msg},
+            ],
+            "stream": True,
+            "max_tokens": num_predict,
+        }
+        if num_ctx_setting != "auto":
+            req_body["options"] = {
+                "num_ctx": int(num_ctx_setting),
                 "num_predict": num_predict,
                 "repeat_penalty": repeat_penalty,
                 "repeat_last_n": repeat_last_n,
-            }},
+            }
+
+        r = requests.post(
+            f"{provider['base_url'].rstrip('/')}/v1/chat/completions",
+            json=req_body,
+            headers=headers,
             stream=True,
             timeout=1200,
         )
         r.raise_for_status()
 
-        acc_thinking = ""
+        acc = ""
-        acc_response = ""
-        ollama_final = {}
         chunk_count = 0
+        finish_reason = None
         for line in r.iter_lines():
             if not line:
                 continue
-            chunk = json.loads(line)
+            line_str = line.decode() if isinstance(line, bytes) else line
-            acc_thinking += chunk.get("thinking") or ""
+            if not line_str.startswith("data: "):
-            acc_response += chunk.get("response") or ""
+                continue
+            data_str = line_str[6:]
+            if data_str.strip() == "[DONE]":
+                break
+            try:
+                chunk = json.loads(data_str)
+            except Exception:
+                continue
+            choice = (chunk.get("choices") or [{}])[0]
+            acc += choice.get("delta", {}).get("content", "") or ""
+            finish_reason = choice.get("finish_reason") or finish_reason
+            t_part, r_part = _split_thinking(acc)
             with _JOB_STREAM_LOCK:
-                _JOB_STREAMS[job_id] = {"thinking": acc_thinking, "response": acc_response}
+                _JOB_STREAMS[job_id] = {"thinking": t_part, "response": r_part}
             chunk_count += 1
             if chunk_count % 100 == 0:
                 j_check = _job_get(job_id)
                 if not j_check or j_check["status"] == "cancelled":
                     return
-            if chunk.get("done"):
-                ollama_final = chunk
-                break
 
-        answer = acc_response
+        acc_thinking, answer = _split_thinking(acc)
 
         j = _job_get(job_id)
         if not j or j["status"] == "cancelled":
@@ -560,7 +646,7 @@ def _process_analysis_job(job_id: int):
                     answer,
                     doc["id"],
                     prm["id"],
-                    json.dumps({"ollama_response": ollama_final}, ensure_ascii=False),
+                    json.dumps({"provider": provider["name"], "model": model, "finish_reason": finish_reason}, ensure_ascii=False),
                     now_iso(),
                 ),
             )
@@ -667,15 +753,22 @@ self.addEventListener('fetch', (event) => {
 
 @app.get("/healthz")
 def healthz():
-    return {"ok": True, "api_base": API_BASE, "ollama_base_url": OLLAMA_BASE_URL, "ollama_model": OLLAMA_MODEL, "db_path": DB_PATH}
+    return {"ok": True, "api_base": API_BASE, "db_path": DB_PATH}
 
 
 @app.get("/settings/models")
-def settings_models():
+def settings_models(provider_id: Optional[int] = None):
+    provider = _get_provider(provider_id)
+    if not provider:
+        return JSONResponse({"ok": False, "error": "Kein Anbieter konfiguriert", "models": []})
+    headers = {}
+    if provider.get("api_key"):
+        headers["Authorization"] = f"Bearer {provider['api_key']}"
     try:
-        r = requests.get(f"{OLLAMA_BASE_URL}/api/tags", timeout=5)
+        r = requests.get(f"{provider['base_url'].rstrip('/')}/v1/models", headers=headers, timeout=5)
         r.raise_for_status()
-        names = sorted(m["name"] for m in r.json().get("models", []))
+        data = r.json()
+        names = sorted(m["id"] for m in data.get("data", []))
         return JSONResponse({"ok": True, "models": names})
     except Exception as e:
         return JSONResponse({"ok": False, "error": str(e), "models": []})
@@ -683,12 +776,11 @@ def settings_models():
 
 @app.get("/settings", response_class=HTMLResponse)
 def settings_page(msg: str = ""):
-    current_model = get_setting("ollama_model", OLLAMA_MODEL)
+    current_model = get_setting("llm_model", _SEED_MODEL)
-    current_think = get_setting("ollama_think", "true" if OLLAMA_THINK else "false") == "true"
+    current_num_ctx = get_setting("llm_num_ctx", "auto")
-    current_num_ctx = get_setting("ollama_num_ctx", "auto")
+    current_num_predict = get_setting("llm_num_predict", str(_SEED_NUM_PREDICT))
-    current_num_predict = get_setting("ollama_num_predict", str(OLLAMA_NUM_PREDICT))
+    current_repeat_penalty = get_setting("llm_repeat_penalty", "1.15")
-    current_repeat_penalty = get_setting("ollama_repeat_penalty", "1.15")
+    current_repeat_last_n = get_setting("llm_repeat_last_n", "128")
-    current_repeat_last_n = get_setting("ollama_repeat_last_n", "128")
 
     smtp_host = get_setting("smtp_host", "")
     smtp_port = get_setting("smtp_port", "587")
@@ -699,8 +791,25 @@ def settings_page(msg: str = ""):
     smtp_tls_checked = "checked" if smtp_tls else ""
 
     with db() as c:
+        providers = c.execute("SELECT * FROM providers ORDER BY name").fetchall()
         email_addrs = c.execute("SELECT id, label, email FROM email_addresses ORDER BY label").fetchall()
 
+    provider_rows = "".join(
+        f"""<tr>
+          <td>{p["name"]}</td>
+          <td><code class='small'>{p["base_url"]}</code></td>
+          <td class='text-center'><i class='bi bi-{'key-fill text-success' if p["api_key"] else 'key text-secondary'}'></i></td>
+          <td class='text-center'>{'<span class=\'badge bg-primary\'>Standard</span>' if p["is_default"] else f'<form method=\'post\' action=\'/providers/{p["id"]}/set-default\' class=\'d-inline\'><button class=\'btn btn-sm btn-outline-secondary py-0 px-2\'>Standard</button></form>'}</td>
+          <td class='text-end'>
+            <form method='post' action='/providers/{p["id"]}/delete' class='d-inline'
+                  onsubmit='return confirm("Anbieter löschen?")'>
+              <button class='btn btn-sm btn-outline-danger py-0 px-2'><i class='bi bi-trash'></i></button>
+            </form>
+          </td>
+        </tr>"""
+        for p in providers
+    )
+
     email_addr_rows = "".join(
         f"""<tr>
           <td>{a["label"]}</td>
@@ -715,7 +824,9 @@ def settings_page(msg: str = ""):
         for a in email_addrs
     )
 
-    think_checked = "checked" if current_think else ""
+    provider_id_opts = "".join(
+        f"<option value='{p['id']}'>{p['name']}</option>" for p in providers
+    )
 
     ctx_options = [("auto", "Automatisch (empfohlen)"), ("4096", "4 096"), ("8192", "8 192"),
                    ("16384", "16 384"), ("32768", "32 768"), ("65536", "65 536"), ("131072", "131 072")]
@@ -729,16 +840,48 @@ def settings_page(msg: str = ""):
   <h2 class='h4 mb-0'>Einstellungen</h2>
 </div>
 {f"<div class='alert alert-success py-2'>{msg}</div>" if msg else ""}
-<div class='card'>
+
-  <h5 class='h6 fw-semibold mb-3'><i class='bi bi-robot text-primary me-1'></i> Ollama / LLM</h5>
+<div class='card' id='anbieter'>
+  <h5 class='h6 fw-semibold mb-3'><i class='bi bi-hdd-network text-primary me-1'></i> LLM-Anbieter</h5>
+  <p class='text-secondary small mb-2'>OpenAI-kompatible Endpunkte (Ollama, OpenAI, Groq, LM Studio, …)</p>
+  {"<table class='table table-sm mb-3'><thead><tr><th>Name</th><th>Base-URL</th><th class='text-center'>API-Key</th><th class='text-center'>Standard</th><th></th></tr></thead><tbody>" + provider_rows + "</tbody></table>" if providers else "<p class='text-secondary small'>Noch kein Anbieter konfiguriert.</p>"}
+  <form method='post' action='/providers/add' class='row g-2 align-items-end'>
+    <div class='col-12 col-md-3'>
+      <label class='form-label'>Name</label>
+      <input class='form-control' name='name' placeholder='z.B. Lokales Ollama' required>
+    </div>
+    <div class='col-12 col-md-4'>
+      <label class='form-label'>Base-URL</label>
+      <input class='form-control font-monospace' name='base_url' placeholder='http://localhost:11434' required>
+    </div>
+    <div class='col-12 col-md-3'>
+      <label class='form-label'>API-Key <span class='text-secondary fw-normal'>(optional)</span></label>
+      <input class='form-control' type='password' name='api_key' autocomplete='new-password'>
+    </div>
+    <div class='col-auto d-flex align-items-end gap-2'>
+      <div class='form-check mb-2'>
+        <input class='form-check-input' type='checkbox' name='is_default' value='1' id='provDefault'>
+        <label class='form-check-label' for='provDefault'>Standard</label>
+      </div>
+      <button class='btn btn-outline-primary mb-0' type='submit'><i class='bi bi-plus-lg'></i> Hinzufügen</button>
+    </div>
+  </form>
+</div>
+
+<div class='card mt-3'>
+  <h5 class='h6 fw-semibold mb-3'><i class='bi bi-sliders text-primary me-1'></i> Standard-LLM-Parameter</h5>
+  <p class='text-secondary small mb-2'>Gelten für alle Prompts ohne eigene Einstellungen. Anbieter = markierter Standard-Anbieter oben.</p>
   <form method='post' action='/settings'>
     <div class='row g-3'>
 
       <div class='col-12'>
-        <label class='form-label'>Modell</label>
+        <label class='form-label'>Standard-Modell</label>
         <div class='input-group'>
-          <input id='modelInput' class='form-control font-monospace' name='ollama_model'
+          <select id='providerSelect' class='form-select' style='max-width:220px' onchange='updateProviderForLoad()'>
-                 value='{current_model}' placeholder='z.B. qwen3.5:9b' required>
+            {provider_id_opts}
+          </select>
+          <input id='modelInput' class='form-control font-monospace' name='llm_model'
+                 value='{current_model}' placeholder='z.B. qwen3:8b' required>
           <button type='button' class='btn btn-outline-secondary' id='loadModelsBtn' onclick='loadModels()'>
             <i class='bi bi-arrow-clockwise' id='loadIcon'></i> Modelle laden
           </button>
@@ -751,38 +894,28 @@ def settings_page(msg: str = ""):
 
       <div class='col-12 col-md-4'>
         <label class='form-label'>Kontextgröße (num_ctx)</label>
-        <select class='form-select' name='ollama_num_ctx'>{ctx_opts_html}</select>
+        <select class='form-select' name='llm_num_ctx'>{ctx_opts_html}</select>
         <div class='form-text'>Automatisch wählt passend zur Promptgröße.</div>
       </div>
 
       <div class='col-12 col-md-4'>
         <label class='form-label'>Max. Ausgabe-Tokens (num_predict)</label>
-        <input class='form-control' type='number' name='ollama_num_predict' min='256' max='131072' step='256'
+        <input class='form-control' type='number' name='llm_num_predict' min='256' max='131072' step='256'
                value='{current_num_predict}'>
       </div>
 
       <div class='col-12 col-md-2'>
         <label class='form-label'>Repeat Penalty</label>
-        <input class='form-control' type='number' name='ollama_repeat_penalty' min='1.0' max='2.0' step='0.01'
+        <input class='form-control' type='number' name='llm_repeat_penalty' min='1.0' max='2.0' step='0.01'
                value='{current_repeat_penalty}'>
       </div>
 
       <div class='col-12 col-md-2'>
         <label class='form-label'>Repeat Last N</label>
-        <input class='form-control' type='number' name='ollama_repeat_last_n' min='0' max='2048' step='1'
+        <input class='form-control' type='number' name='llm_repeat_last_n' min='0' max='2048' step='1'
                value='{current_repeat_last_n}'>
       </div>
 
-      <div class='col-12'>
-        <div class='form-check form-switch'>
-          <input class='form-check-input' type='checkbox' role='switch' id='thinkSwitch'
-                 name='ollama_think' value='true' {think_checked}>
-          <label class='form-check-label' for='thinkSwitch'>
-            Thinking aktivieren <span class='text-secondary small'>(Extended Thinking / Chain-of-Thought)</span>
-          </label>
-        </div>
-      </div>
-
     </div>
     <div class='mt-3'>
       <button class='btn btn-primary' type='submit'><i class='bi bi-check2-circle'></i> Speichern</button>
@@ -851,13 +984,18 @@ def settings_page(msg: str = ""):
   <table class='table table-sm mb-0'>
     <tbody>
       <tr><td class='text-secondary' style='width:220px'>API_BASE</td><td><code>{API_BASE}</code></td></tr>
-      <tr><td class='text-secondary'>OLLAMA_BASE_URL</td><td><code>{OLLAMA_BASE_URL}</code></td></tr>
       <tr><td class='text-secondary'>DB_PATH</td><td><code>{DB_PATH}</code></td></tr>
     </tbody>
   </table>
 </div>
 
 <script>
+let _loadProviderId = null;
+function updateProviderForLoad() {{
+  _loadProviderId = document.getElementById('providerSelect')?.value || null;
+  document.getElementById('modelSelectWrap').classList.add('d-none');
+}}
+
 async function loadModels() {{
   const btn = document.getElementById('loadModelsBtn');
   const icon = document.getElementById('loadIcon');
@@ -866,11 +1004,12 @@ async function loadModels() {{
   const inp = document.getElementById('modelInput');
   btn.disabled = true;
   icon.className = 'bi bi-hourglass-split';
+  const pid = _loadProviderId || document.getElementById('providerSelect')?.value || '';
   try {{
-    const res = await fetch('/settings/models');
+    const res = await fetch('/settings/models' + (pid ? '?provider_id=' + pid : ''));
     const data = await res.json();
     if (!data.ok || !data.models.length) {{
-      alert('Keine Modelle gefunden oder Ollama nicht erreichbar:\\n' + (data.error || ''));
+      alert('Keine Modelle gefunden:\\n' + (data.error || ''));
       return;
     }}
     const current = inp.value.trim();
@@ -897,19 +1036,17 @@ async function loadModels() {{
 
 @app.post("/settings", response_class=HTMLResponse)
 def settings_save(
-    ollama_model: str = Form(...),
+    llm_model: str = Form(...),
-    ollama_think: Optional[str] = Form(None),
+    llm_num_ctx: str = Form("auto"),
-    ollama_num_ctx: str = Form("auto"),
+    llm_num_predict: str = Form("16384"),
-    ollama_num_predict: str = Form("16384"),
+    llm_repeat_penalty: str = Form("1.15"),
-    ollama_repeat_penalty: str = Form("1.15"),
+    llm_repeat_last_n: str = Form("128"),
-    ollama_repeat_last_n: str = Form("128"),
 ):
-    set_setting("ollama_model", ollama_model.strip())
+    set_setting("llm_model", llm_model.strip())
-    set_setting("ollama_think", "true" if ollama_think == "true" else "false")
+    set_setting("llm_num_ctx", llm_num_ctx.strip())
-    set_setting("ollama_num_ctx", ollama_num_ctx.strip())
+    set_setting("llm_num_predict", str(int(float(llm_num_predict))))
-    set_setting("ollama_num_predict", str(int(float(ollama_num_predict))))
+    set_setting("llm_repeat_penalty", str(float(llm_repeat_penalty)))
-    set_setting("ollama_repeat_penalty", str(float(ollama_repeat_penalty)))
+    set_setting("llm_repeat_last_n", str(int(float(llm_repeat_last_n))))
-    set_setting("ollama_repeat_last_n", str(int(float(ollama_repeat_last_n))))
     return RedirectResponse("/settings?msg=Einstellungen+gespeichert.", status_code=303)
 
 
@@ -949,6 +1086,38 @@ def delete_email_address(addr_id: int):
     return RedirectResponse("/settings?msg=E-Mail-Adresse+gelöscht.#email-adressen", status_code=303)
 
 
+@app.post("/providers/add", response_class=HTMLResponse)
+def add_provider(
+    name: str = Form(...),
+    base_url: str = Form(...),
+    api_key: str = Form(""),
+    is_default: Optional[str] = Form(None),
+):
+    with db() as c:
+        if is_default == "1":
+            c.execute("UPDATE providers SET is_default=0")
+        c.execute(
+            "INSERT INTO providers(name, base_url, api_key, is_default, created_at) VALUES (?,?,?,?,?)",
+            (name.strip(), base_url.strip().rstrip("/"), api_key.strip() or None, 1 if is_default == "1" else 0, now_iso()),
+        )
+    return RedirectResponse("/settings?msg=Anbieter+hinzugefügt.#anbieter", status_code=303)
+
+
+@app.post("/providers/{provider_id}/delete", response_class=HTMLResponse)
+def delete_provider(provider_id: int):
+    with db() as c:
+        c.execute("DELETE FROM providers WHERE id=?", (provider_id,))
+    return RedirectResponse("/settings?msg=Anbieter+gelöscht.#anbieter", status_code=303)
+
+
+@app.post("/providers/{provider_id}/set-default", response_class=HTMLResponse)
+def set_default_provider(provider_id: int):
+    with db() as c:
+        c.execute("UPDATE providers SET is_default=0")
+        c.execute("UPDATE providers SET is_default=1 WHERE id=?", (provider_id,))
+    return RedirectResponse("/settings?msg=Standard-Anbieter+gesetzt.#anbieter", status_code=303)
+
+
 @app.get("/", response_class=HTMLResponse)
 def upload_page(msg: str = ""):
     projects = get_projects()
@@ -1598,21 +1767,24 @@ def bulk_delete_documents(ids: List[int] = Form(...)):
     return HTMLResponse("<meta http-equiv='refresh' content='0; url=/library'>")
 
 
-def _prompt_llm_block(p, def_model, def_think, def_num_ctx, def_num_predict, def_repeat_penalty, def_repeat_last_n):
+def _prompt_llm_block(p, providers, def_model, def_num_ctx, def_num_predict, def_repeat_penalty, def_repeat_last_n):
     pid = p["id"]
     use_default = int(p["llm_use_default"]) if p["llm_use_default"] is not None else 1
     m = p["llm_model"] or def_model
-    th = p["llm_think"] or def_think
     nc = p["llm_num_ctx"] or def_num_ctx
     np_ = p["llm_num_predict"] or def_num_predict
     rp = p["llm_repeat_penalty"] or def_repeat_penalty
     rn = p["llm_repeat_last_n"] or def_repeat_last_n
+    cur_provider_id = p.get("provider_id")
     disabled = "disabled" if use_default else ""
     opacity = "opacity:.45;pointer-events:none" if use_default else ""
     ctx_opts = [("auto","Automatisch"), ("4096","4 096"), ("8192","8 192"), ("16384","16 384"),
                 ("32768","32 768"), ("65536","65 536"), ("131072","131 072")]
     ctx_html = "".join(f"<option value='{v}'{' selected' if v==str(nc) else ''}>{l}</option>" for v,l in ctx_opts)
-    think_checked = "checked" if str(th) == "true" else ""
+    provider_opts = "".join(
+        f"<option value='{pv['id']}'{' selected' if pv['id'] == cur_provider_id else ''}>{pv['name']}</option>"
+        for pv in providers
+    )
     return (
         f"<hr class='my-3'>"
         f"<div class='d-flex align-items-center gap-3 mb-2'>"
@@ -1625,11 +1797,18 @@ def _prompt_llm_block(p, def_model, def_think, def_num_ctx, def_num_predict, def
         f"</div>"
         f"<input type='hidden' id='llm_use_default_{pid}' name='llm_use_default' value='{use_default}'>"
         f"<div id='llmfields_{pid}' class='row g-2' style='{opacity}'>"
-        f"  <div class='col-12'>"
+        f"  <div class='col-12 col-md-4'>"
+        f"    <label class='form-label small text-secondary mb-1'>Anbieter</label>"
+        f"    <select class='form-select form-select-sm' name='provider_id' id='llmprov_{pid}'"
+        f"      onchange='loadModelsInto(\"llmmodel_{pid}\",this.value)' {disabled}>"
+        f"      <option value=''>– Standard –</option>{provider_opts}"
+        f"    </select>"
+        f"  </div>"
+        f"  <div class='col-12 col-md-8'>"
         f"    <label class='form-label small text-secondary mb-1'>Modell</label>"
         f"    <div class='input-group input-group-sm'>"
         f"      <input id='llmmodel_{pid}' class='form-control form-control-sm font-monospace' name='llm_model' value='{m}' {disabled}>"
-        f"      <button type='button' class='btn btn-outline-secondary btn-sm' onclick='loadModelsInto(\"llmmodel_{pid}\")' {'disabled' if use_default else ''}>"
+        f"      <button type='button' class='btn btn-outline-secondary btn-sm' onclick='loadModelsInto(\"llmmodel_{pid}\",document.getElementById(\"llmprov_{pid}\").value)' {'disabled' if use_default else ''}>"
         f"        <i class='bi bi-arrow-clockwise'></i>"
         f"      </button>"
         f"    </div>"
@@ -1650,12 +1829,6 @@ def _prompt_llm_block(p, def_model, def_think, def_num_ctx, def_num_predict, def
         f"    <label class='form-label small text-secondary mb-1'>Repeat Last N</label>"
         f"    <input class='form-control form-control-sm' type='number' name='llm_repeat_last_n' min='0' max='2048' step='1' value='{rn}' {disabled}>"
         f"  </div>"
-        f"  <div class='col-12'>"
-        f"    <div class='form-check form-switch'>"
-        f"      <input class='form-check-input' type='checkbox' role='switch' id='llmthink_{pid}' name='llm_think' value='true' {think_checked} {disabled}>"
-        f"      <label class='form-check-label small' for='llmthink_{pid}'>Thinking aktivieren</label>"
-        f"    </div>"
-        f"  </div>"
         f"</div>"
     )
 
@@ -1665,12 +1838,12 @@ def prompts_page():
     with db() as c:
         prompts = c.execute("SELECT * FROM prompts ORDER BY name").fetchall()
         projects = c.execute("SELECT id,name FROM projects ORDER BY name").fetchall()
-    def_model = get_setting("ollama_model", OLLAMA_MODEL)
+        providers = c.execute("SELECT * FROM providers ORDER BY name").fetchall()
-    def_think = get_setting("ollama_think", "true" if OLLAMA_THINK else "false")
+    def_model = get_setting("llm_model", _SEED_MODEL)
-    def_num_ctx = get_setting("ollama_num_ctx", "auto")
+    def_num_ctx = get_setting("llm_num_ctx", "auto")
-    def_num_predict = get_setting("ollama_num_predict", str(OLLAMA_NUM_PREDICT))
+    def_num_predict = get_setting("llm_num_predict", str(_SEED_NUM_PREDICT))
-    def_repeat_penalty = get_setting("ollama_repeat_penalty", "1.15")
+    def_repeat_penalty = get_setting("llm_repeat_penalty", "1.15")
-    def_repeat_last_n = get_setting("ollama_repeat_last_n", "128")
+    def_repeat_last_n = get_setting("llm_repeat_last_n", "128")
 
     project_opts = "".join([f"<option value='{p['name']}'>{p['name']}</option>" for p in projects])
 
@@ -1706,7 +1879,7 @@ def prompts_page():
             f"<textarea id='prompt_text_{p['id']}' class='form-control' name='prompt' style='min-height:110px; resize:vertical'>{p['prompt']}</textarea>"
             f"</div>"
             f"</div>"
-            + _prompt_llm_block(p, def_model, def_think, def_num_ctx, def_num_predict, def_repeat_penalty, def_repeat_last_n) +
+            + _prompt_llm_block(p, providers, def_model, def_num_ctx, def_num_predict, def_repeat_penalty, def_repeat_last_n) +
             f"<div class='mt-3 d-flex flex-wrap gap-2'>"
             f"<button class='btn btn-primary btn-sm' type='submit'><i class='bi bi-check2-circle'></i> Speichern</button>"
             f"<button class='btn btn-outline-secondary btn-sm' type='button' onclick='previewPrompt({p['id']})'><i class='bi bi-eye'></i> Anzeigen</button>"
@@ -1833,7 +2006,6 @@ document.addEventListener('DOMContentLoaded', () => {{
 
 const _llmDefaults = {{
   model: {json.dumps(def_model)},
-  think: {json.dumps(def_think)},
   num_ctx: {json.dumps(def_num_ctx)},
   num_predict: {json.dumps(def_num_predict)},
   repeat_penalty: {json.dumps(def_repeat_penalty)},
@@ -1858,19 +2030,19 @@ function toggleLlmDef(pid, isDefault) {{
   }}
 }}
 
-async function loadModelsInto(inputId) {{
+async function loadModelsInto(inputId, providerId) {{
   const inp = document.getElementById(inputId);
   if (!inp) return;
   const btn = inp.nextElementSibling;
   if (btn) btn.disabled = true;
   try {{
-    const res = await fetch('/settings/models');
+    const pid = providerId || '';
+    const res = await fetch('/settings/models' + (pid ? '?provider_id=' + pid : ''));
     const data = await res.json();
     if (!data.ok || !data.models.length) {{
       alert('Keine Modelle gefunden: ' + (data.error || ''));
       return;
     }}
-    const current = inp.value.trim();
     const v = await window.uiSelect('Modell wählen',
       data.models.map(m => ({{value: m, label: m}})),
       '– Modell auswählen –'
@@ -1913,27 +2085,28 @@ def prompt_update(
     prompt: str = Form(...),
     llm_use_default: int = Form(1),
     llm_model: Optional[str] = Form(None),
-    llm_think: Optional[str] = Form(None),
     llm_num_ctx: Optional[str] = Form(None),
     llm_num_predict: Optional[str] = Form(None),
     llm_repeat_penalty: Optional[str] = Form(None),
     llm_repeat_last_n: Optional[str] = Form(None),
+    provider_id: Optional[int] = Form(None),
 ):
     with db() as c:
         c.execute(
             """UPDATE prompts SET name=?, prompt=?, updated_at=?,
-               llm_use_default=?, llm_model=?, llm_think=?,
+               llm_use_default=?, llm_model=?,
-               llm_num_ctx=?, llm_num_predict=?, llm_repeat_penalty=?, llm_repeat_last_n=?
+               llm_num_ctx=?, llm_num_predict=?, llm_repeat_penalty=?, llm_repeat_last_n=?,
+               provider_id=?
                WHERE id=?""",
             (
                 name.strip(), prompt.strip(), now_iso(),
                 llm_use_default,
                 (llm_model or "").strip() or None,
-                "true" if llm_think == "true" else "false",
                 (llm_num_ctx or "auto").strip(),
                 int(float(llm_num_predict)) if llm_num_predict else None,
                 float(llm_repeat_penalty) if llm_repeat_penalty else None,
                 int(float(llm_repeat_last_n)) if llm_repeat_last_n else None,
+                provider_id or None,
                 id,
             ),
         )