fix(projekt-matching): retry chat_json on empty/unparseable vLLM completion

projekt-matching/projektmatch/llm.py

@@ -7,6 +7,7 @@ from __future__ import annotations
 
 import json
 import re
+import time
 
 import requests
 
@@ -131,7 +132,9 @@ def _post(base, model, messages, schema):
                          timeout=1500)
 
 
-def chat_json(base, model, messages, schema):
+def chat_json(base, model, messages, schema, attempts=3):
+    last = None
+    for attempt in range(attempts):
         resp = _post(base, model, messages, schema)
         if resp.status_code != 200:
             raise LlmError(f"vLLM HTTP {resp.status_code}: "
@@ -143,7 +146,9 @@ def chat_json(base, model, messages, schema):
         try:
             return json.loads(content)
         except json.JSONDecodeError as exc:
-        raise LlmError(f"LLM lieferte kein JSON: {exc}: {content[:200]}")
+            last = LlmError(f"LLM lieferte kein JSON: {exc}: {content[:200]}")
+            time.sleep(10 * (attempt + 1))
+    raise last
 
 
 def extract_project(base, model, page_text):

projekt-matching/tests/test_llm.py

@@ -35,6 +35,20 @@ def test_chat_json_strips_markdown_fences():
     assert out == {"a": 2}
 
 
+def test_chat_json_retries_on_empty_content():
+    good = json.dumps({"a": 1})
+    with mock.patch("projektmatch.llm.time.sleep", lambda s: None), \
+         mock.patch("projektmatch.llm.requests.post", fake_post(["", good])):
+        assert llm.chat_json("http://v/v1", "m", [], {"type": "object"}) == {"a": 1}
+
+
+def test_chat_json_raises_after_exhausted_retries():
+    with mock.patch("projektmatch.llm.time.sleep", lambda s: None), \
+         mock.patch("projektmatch.llm.requests.post", fake_post(["", "", ""])):
+        with pytest.raises(llm.LlmError, match="kein JSON"):
+            llm.chat_json("http://v/v1", "m", [], {"type": "object"})
+
+
 def test_match_cv_retries_on_missing_nr_then_raises():
     reqs = [{"nr": 1, "text": "Python"}, {"nr": 2, "text": "K8s"}]
     partial = json.dumps({"ratings": [