GWDG jetzt auch für chat

math-tutor/backend/app/LLM_services/hint_LLM.py

@@ -5,7 +5,7 @@ def generate_hint(
     task: str,
     solution: str,
     history: str | None = None,
-    context_sheet: str | None = None,
+    retrival: str | None = None,
 ) -> str:
     prompt = (
         "Du bist ein didaktischer Tutor. "
@@ -20,9 +20,9 @@ def generate_hint(
         + "\n"
     )
     if history:
-        prompt += "\nHistorie:\n" + history + "\n"
-    if context_sheet:
-        prompt += "\nKontext:\n" + context_sheet + "\n"
+        prompt = "\nHistorie:\n" + history + "\n" + prompt
+    if retrival:
+        prompt = "\nKontext:\n" + retrival + "\n" + prompt
     prompt += "Halte dich kurz und klar. Gibt nicht die Lösung aus."
 
     result = llm_client.chat(
@@ -51,7 +51,7 @@ TOOL_SPEC = {
                     "type": "string",
                     "description": "Optionaler Verlauf, kann leer sein",
                 },
-                "context_sheet": {
+                "retrival": {
                     "type": "string",
                     "description": "Optionales Kontextblatt mit Werkzeug- und Retrieval-Infos",
                 },

math-tutor/backend/app/config.py

@@ -21,6 +21,13 @@ class EmbeddingSettings:
     model: str
     target_dim: int
 
+@dataclass(frozen=True)
+class OpenAIChatSettings:
+    base_url: str
+    api_key: str
+    model: str
+    timeout: float | None
+
 @dataclass(frozen=True)
 class MathpixSettings:
     app_id: str
@@ -63,6 +70,21 @@ def get_embedding_settings() -> EmbeddingSettings:
         target_dim=1024,
     )
 
+def get_openai_chat_settings() -> OpenAIChatSettings | None:
+    model = os.getenv("OPENAI_CHAT_MODEL")
+    if not model:
+        return None
+    base_url = os.getenv("OPENAI_BASE_URL")
+    api_key = os.getenv("OPENAI_API_KEY")
+    if not base_url or not api_key:
+        raise ValueError("Missing OPENAI_BASE_URL or OPENAI_API_KEY for chat")
+    return OpenAIChatSettings(
+        base_url=base_url,
+        api_key=api_key,
+        model=model,
+        timeout=_read_float(os.getenv("OPENAI_CHAT_TIMEOUT")),
+    )
+
 
 def get_postgres_url() -> str:
     pg_url = os.getenv("POSTGRES_URL")

math-tutor/backend/app/deterministic_services/context_store.py

@@ -163,6 +163,12 @@ def format_sheet(sheet: dict[str, Any]) -> str:
 
     return "\n\n".join(parts)
 
+def get_retrival(sheet: dict[str, Any]) -> str:
+    retrievals = sheet.get("retrieval_contexts", [])
+    if not retrievals:
+        return ""
+    latest = retrievals[-1]
+    return latest.get("context", "")
 
 def save_sheet(sheet: dict[str, Any]) -> None:
     os.makedirs(_LOG_DIR, exist_ok=True)

math-tutor/backend/app/deterministic_services/llm_client.py

 import inspect
+
 import ollama
+from openai import OpenAI
 
 from app import config
 
@@ -12,7 +14,27 @@ def _filter_kwargs(func, kwargs: dict) -> dict:
     return {key: value for key, value in kwargs.items() if key in signature.parameters}
 
 
+def _chat_openai(messages: list[dict]) -> dict:
+    settings = config.get_openai_chat_settings()
+    if not settings:
+        return {}
+
+    timeout = settings.timeout or 60.0
+    client = OpenAI(api_key=settings.api_key, base_url=settings.base_url, timeout=timeout)
+    response = client.chat.completions.create(
+        messages=messages,
+        model=settings.model,
+    )
+    message = response.choices[0].message if response.choices else {}
+    return {"raw": response, "message": message}
+
+
 def chat(messages: list[dict], tools: list[dict] | None = None) -> dict:
+    if not tools:
+        openai_result = _chat_openai(messages)
+        if openai_result:
+            return openai_result
+
     settings = config.get_ollama_settings()
     client = ollama.Client(host=settings.base_url, timeout=settings.timeout)
 

math-tutor/backend/app/deterministic_services/orchestrator.py

@@ -62,7 +62,7 @@ def run_chat(messages: list[dict], draft: str | None = None) -> dict:
         "task": last_user.get("content", ""),
         "solution": context_store.first_math_solution(sheet),
         "history": history_text,
-        "context_sheet": sheet_text,
+        "retrival": context_store.get_retrival(sheet),
     }
     reply = hint_LLM.generate_hint(**hint_args)
     _append_tool_log(tool_log, "generate_hint", hint_args, reply)

math-tutor/backend/requirements.txt

@@ -5,6 +5,7 @@ mpxpy
 pillow
 httpx
 ollama
+openai
 sympy
 psycopg[binary]
 pgvector