From 9cbdd2e28cde3e3329c4c3d5cc02b87692ef3cdd9930f98fc616887fd810c8ce Mon Sep 17 00:00:00 2001
From: tlg <thomas.langer@destengs.com>
Date: Thu, 9 Jul 2026 10:55:23 +0200
Subject: [PATCH] fix(projekt-matching): use response_format json_schema (vLLM
 0.22 ignores guided_json)

---
 projekt-matching/projektmatch/llm.py | 15 +++++++--------
 projekt-matching/tests/test_llm.py   | 16 ++++++----------
 2 files changed, 13 insertions(+), 18 deletions(-)

diff --git a/projekt-matching/projektmatch/llm.py b/projekt-matching/projektmatch/llm.py
index f9592d5..ac555ef 100644
--- a/projekt-matching/projektmatch/llm.py
+++ b/projekt-matching/projektmatch/llm.py
@@ -123,24 +123,23 @@ z. B. vLLM, TGI, Triton" ab).
 - reason: EIN kurzer deutscher Satz mit der Begründung."""
 
 
-def _post(base, model, messages, body_extra):
-    body = {"model": model, "messages": messages, "temperature": 0.1}
-    body.update(body_extra)
+def _post(base, model, messages, schema):
+    body = {"model": model, "messages": messages, "temperature": 0.1,
+            "response_format": {"type": "json_schema", "json_schema": {
+                "name": "out", "schema": schema}}}
     return requests.post(f"{base.rstrip('/')}/chat/completions", json=body,
                          timeout=1500)
 
 
 def chat_json(base, model, messages, schema):
-    resp = _post(base, model, messages, {"guided_json": schema})
-    if resp.status_code == 400 and "guided_json" in getattr(resp, "text", ""):
-        resp = _post(base, model, messages, {"response_format": {
-            "type": "json_schema",
-            "json_schema": {"name": "out", "schema": schema}}})
+    resp = _post(base, model, messages, schema)
     if resp.status_code != 200:
         raise LlmError(f"vLLM HTTP {resp.status_code}: "
                        f"{getattr(resp, 'text', '')[:300]}")
     content = resp.json()["choices"][0]["message"]["content"] or ""
     content = THINK_RE.sub("", content).strip()
+    if content.startswith("```"):
+        content = re.sub(r"^```[a-z]*\s*|\s*```$", "", content).strip()
     try:
         return json.loads(content)
     except json.JSONDecodeError as exc:
diff --git a/projekt-matching/tests/test_llm.py b/projekt-matching/tests/test_llm.py
index 0f3f83f..6535f52 100644
--- a/projekt-matching/tests/test_llm.py
+++ b/projekt-matching/tests/test_llm.py
@@ -25,18 +25,14 @@ def test_chat_json_strips_think_and_parses():
     assert out == {"a": 1}
     body = p.call_args.kwargs["json"]
     assert "max_tokens" not in body
-    assert body["guided_json"] == {"type": "object"}
+    assert body["response_format"]["json_schema"]["schema"] == {"type": "object"}
 
 
-def test_chat_json_fallback_to_response_format():
-    bad = mock.Mock(status_code=400,
-                    text="Unknown parameter: 'guided_json'")
-    good = mock.Mock(status_code=200)
-    good.json.return_value = {"choices": [{"message": {"content": "{}"}}]}
-    with mock.patch("projektmatch.llm.requests.post",
-                    mock.Mock(side_effect=[bad, good])) as p:
-        assert llm.chat_json("http://v/v1", "m", [], {"type": "object"}) == {}
-    assert "response_format" in p.call_args.kwargs["json"]
+def test_chat_json_strips_markdown_fences():
+    content = "```json\n{\"a\": 2}\n```"
+    with mock.patch("projektmatch.llm.requests.post", fake_post([content])):
+        out = llm.chat_json("http://v/v1", "m", [], {"type": "object"})
+    assert out == {"a": 2}
 
 
 def test_match_cv_retries_on_missing_nr_then_raises():