feat: Ollama und Modell vor dem Start prüfen, kurze Meldung statt Traceback

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

README.md

@@ -21,6 +21,8 @@ Die Werkzeuge sind einzeln oder hintereinander nutzbar:
 - für den Download: [yt-dlp](https://github.com/yt-dlp/yt-dlp) (`pip install yt-dlp`)
 - Ollama auf `127.0.0.1:11434` mit dem Modell `qwen3:32b`
   (`ollama pull qwen3:32b`)
+  `rede_korrigieren.py` und `glossar_vorschlag.py` prüfen beides vor dem Start und
+  brechen sonst mit einem Hinweis ab (`sudo systemctl start ollama` bzw. `ollama pull …`)
 
 Für gute Geschwindigkeit muss das Modell komplett in den Grafikspeicher passen.
 Mit `OLLAMA_NUM_PARALLEL=1` braucht `qwen3:32b` bei 16k Kontext rund 24 GB

glossar_vorschlag.py

@@ -9,11 +9,11 @@ import hashlib
 import io
 import json
 import re
-import urllib.request
 
 from aenderungen import compare, read_raw
 from rede_korrigieren import (
-    MODEL, apply_replacements, chat_request, read_glossary, split_chunks,
+    MODEL, apply_replacements, chat_request, check_ollama, ollama_open, read_glossary,
+    split_chunks,
 )
 
 SYSTEM = """Du suchst Hörfehler in der automatischen Abschrift einer deutschen Rede.
@@ -77,7 +77,7 @@ def suggest_chunk(chunk, model):
         ],
         "options": {"temperature": 0, "num_ctx": 16384, "num_predict": 4096},
     }
-    with urllib.request.urlopen(chat_request(payload), timeout=600) as response:
+    with ollama_open(chat_request(payload)) as response:
         result = json.load(response)
     try:
         entries = json.loads(result["message"]["content"])["eintraege"]
@@ -227,6 +227,7 @@ def main():
     with redirect_stdout(io.StringIO()):
         text = apply_replacements(text, replacements)
 
+    check_ollama(args.model)
     cache_dir = args.output.with_suffix(".parts")
     cache_dir.mkdir(parents=True, exist_ok=True)
     suggestions = model_suggestions(text, args.model, cache_dir, known)

pytest.ini

@@ -3,3 +3,5 @@ testpaths = tests
 pythonpath = .
 # Project has no async code; skip the globally installed asyncio plugin
 addopts = -p no:asyncio
+markers =
+    real_ollama_check: run the real Ollama server check (urlopen still faked)

rede_korrigieren.py

@@ -8,6 +8,7 @@ import re
 import shutil
 import sys
 import time
+import urllib.error
 import urllib.request
 
 from aenderungen import REVIEW_MARK, print_summary, report_path, write_report
@@ -17,6 +18,7 @@ from yt_untertitel import YT_DLP, base_language, default_name, original_language
 
 MODEL = "qwen3:32b"
 OLLAMA_URL = "http://127.0.0.1:11434/api/chat"
+OLLAMA_TAGS_URL = "http://127.0.0.1:11434/api/tags"
 
 CHUNK_SIZE = 1200          # Zielgröße eines Abschnitts in Wörtern
 MAX_EXTRA_WORDS = 300      # so weit darf ein Abschnitt bis zum nächsten Satzende wachsen
@@ -237,6 +239,31 @@ def chat_request(payload):
     )
 
 
+def ollama_open(request, timeout=600):
+    """urlopen() for Ollama; a missing server ends the program with a short message."""
+    try:
+        return urllib.request.urlopen(request, timeout=timeout)
+    except urllib.error.HTTPError:
+        raise
+    except urllib.error.URLError as error:
+        raise SystemExit(
+            f"Ollama nicht erreichbar ({error.reason}). Läuft der Dienst? "
+            "Starten mit: sudo systemctl start ollama (oder: ollama serve)"
+        ) from None
+
+
+def check_ollama(model):
+    """Stop early, before any download, if Ollama is down or lacks the model."""
+    with ollama_open(OLLAMA_TAGS_URL, timeout=10) as response:
+        installed = {m.get("name") for m in json.load(response).get("models", [])}
+    wanted = model if ":" in model else f"{model}:latest"
+    if wanted not in installed:
+        raise SystemExit(
+            f"Modell „{model}“ ist in Ollama nicht installiert. "
+            f"Installieren mit: ollama pull {model}"
+        )
+
+
 def correct_chunk(chunk, index, system):
     payload = {
         "model": MODEL,
@@ -263,7 +290,7 @@ def correct_chunk(chunk, index, system):
     words = 0
     result = {}
 
-    with urllib.request.urlopen(request, timeout=600) as response:
+    with ollama_open(request) as response:
         for line in response:
             if not line.strip():
                 continue
@@ -328,7 +355,7 @@ def find_paragraph_starts(sentences, first, system):
             "num_predict": 2048,
         },
     }
-    with urllib.request.urlopen(chat_request(payload), timeout=600) as response:
+    with ollama_open(chat_request(payload)) as response:
         result = json.load(response)
 
     content = result["message"]["content"]
@@ -542,6 +569,7 @@ def main():
     # reuses cached parts of another
     MODEL = args.model
     print(f"Modell: {MODEL}")
+    check_ollama(MODEL)
     raw, target, lang = read_source(args)
     print(f"Sprache der Rede: {lang}")
     parts_dir = target.with_suffix(".parts")

tests/conftest.py

@@ -1,4 +1,4 @@
-"""Shared fixtures: yt-dlp is always replaced by fakes (no network)."""
+"""Shared fixtures: yt-dlp and Ollama are always replaced by fakes (no network)."""
 import json
 import shutil
 import subprocess
@@ -7,6 +7,17 @@ from pathlib import Path
 import pytest
 
 
+@pytest.fixture(autouse=True)
+def ollama_available(request, monkeypatch):
+    """Skip the Ollama server check; tests marked real_ollama_check test it."""
+    if "real_ollama_check" in request.keywords:
+        return
+    import glossar_vorschlag
+    import rede_korrigieren
+    for module in (rede_korrigieren, glossar_vorschlag):
+        monkeypatch.setattr(module, "check_ollama", lambda model: None)
+
+
 @pytest.fixture
 def fake_yt_dlp():
     """Factory for a subprocess.run stand-in.

tests/test_glossar_vorschlag.py

@@ -2,6 +2,7 @@
 import io
 import json
 import sys
+import urllib.request
 
 import pytest
 
@@ -63,7 +64,7 @@ def run_main(monkeypatch, *args):
 
 def test_main_filters_and_formats(raw_file, glossary, monkeypatch, capsys):
     model = fake_model(ENTRIES)
-    monkeypatch.setattr(gv.urllib.request, "urlopen", model)
+    monkeypatch.setattr(urllib.request, "urlopen", model)
     run_main(monkeypatch, "-i", raw_file, "-g", glossary, "-m", "test:1b")
 
     draft = raw_file.with_name("rede.glossar.entwurf.txt").read_text(encoding="utf-8")
@@ -84,12 +85,12 @@ def test_main_filters_and_formats(raw_file, glossary, monkeypatch, capsys):
     assert "Vermutete Hörfehler: 2 (1 sicher, 1 unsicher)" in out
 
     # Second run uses the cache
-    monkeypatch.setattr(gv.urllib.request, "urlopen", no_model_calls)
+    monkeypatch.setattr(urllib.request, "urlopen", no_model_calls)
     run_main(monkeypatch, "-i", raw_file, "-g", glossary, "-m", "test:1b")
 
 
 def test_correction_changes_are_commented_out(raw_file, glossary, tmp_path, monkeypatch):
-    monkeypatch.setattr(gv.urllib.request, "urlopen", fake_model(ENTRIES[:2]))
+    monkeypatch.setattr(urllib.request, "urlopen", fake_model(ENTRIES[:2]))
     corrected = tmp_path / "rede.korrigiert.txt"
     # Corrected without glossary: "Kremmel" became "Kremml" (a model error)
     corrected.write_text(
@@ -110,7 +111,7 @@ def test_correction_changes_are_commented_out(raw_file, glossary, tmp_path, monk
 
 
 def test_drafted_entries_are_valid_glossary(raw_file, monkeypatch, tmp_path):
-    monkeypatch.setattr(gv.urllib.request, "urlopen", fake_model(ENTRIES[:2]))
+    monkeypatch.setattr(urllib.request, "urlopen", fake_model(ENTRIES[:2]))
     out = tmp_path / "entwurf.txt"
     run_main(monkeypatch, "-i", raw_file, "-o", out)
     assert rk.read_glossary(out) == ([], [])  # nothing active
@@ -126,7 +127,7 @@ def test_unreadable_model_answer(raw_file, monkeypatch, tmp_path, capsys):
     def urlopen(request, timeout=None):
         return io.BytesIO(json.dumps({"message": {"content": "kein JSON"}}).encode())
 
-    monkeypatch.setattr(gv.urllib.request, "urlopen", urlopen)
+    monkeypatch.setattr(urllib.request, "urlopen", urlopen)
     run_main(monkeypatch, "-i", raw_file, "-o", tmp_path / "e.txt")
     assert "unlesbar" in capsys.readouterr().out
 

tests/test_rede_korrigieren.py

@@ -566,3 +566,52 @@ def test_cli_time_errors(sample_copy, monkeypatch, capsys, args, message):
         run_main(monkeypatch, *args)
     assert exc.value.code == 2
     assert message in capsys.readouterr().err
+
+
+# === Ollama availability ===
+
+def tags_urlopen(*names):
+    body = json.dumps({"models": [{"name": n} for n in names]}).encode("utf-8")
+    return fake_urlopen(body)
+
+
+@pytest.mark.real_ollama_check
+@pytest.mark.parametrize("model", ["qwen3:32b", "llama3"])
+def test_check_ollama_model_installed(monkeypatch, model):
+    monkeypatch.setattr(rk.urllib.request, "urlopen", tags_urlopen("qwen3:32b", "llama3:latest"))
+    rk.check_ollama(model)
+
+
+@pytest.mark.real_ollama_check
+def test_check_ollama_model_missing(monkeypatch):
+    monkeypatch.setattr(rk.urllib.request, "urlopen", tags_urlopen("qwen3:8b"))
+    with pytest.raises(SystemExit, match="ollama pull qwen3:32b"):
+        rk.check_ollama("qwen3:32b")
+
+
+def refused(request, timeout=None):
+    raise rk.urllib.error.URLError(ConnectionRefusedError(111, "Connection refused"))
+
+
+@pytest.mark.real_ollama_check
+def test_main_ollama_down_stops_before_download(fake_yt_dlp, yt_dlp_installed, monkeypatch):
+    run = fake_yt_dlp(VIDEO, vtt=VTT)
+    monkeypatch.setattr(subprocess, "run", run)
+    monkeypatch.setattr(rk.urllib.request, "urlopen", refused)
+    with pytest.raises(SystemExit, match="Ollama nicht erreichbar.*systemctl start ollama"):
+        run_main(monkeypatch, "-u", URL)
+    assert run.calls == []
+
+
+def test_correct_chunk_ollama_gone_midway(monkeypatch):
+    monkeypatch.setattr(rk.urllib.request, "urlopen", refused)
+    with pytest.raises(SystemExit, match="Ollama nicht erreichbar"):
+        rk.correct_chunk("Text.", 1, "system")
+
+
+def test_ollama_open_passes_http_errors(monkeypatch):
+    def not_found(request, timeout=None):
+        raise rk.urllib.error.HTTPError(rk.OLLAMA_URL, 404, "Not Found", {}, None)
+    monkeypatch.setattr(rk.urllib.request, "urlopen", not_found)
+    with pytest.raises(rk.urllib.error.HTTPError):
+        rk.ollama_open(rk.OLLAMA_URL)