Limit hinzugefügt

math-tutor/backend/.env-example

@@ -2,6 +2,8 @@ MATHPIX_APP_ID=""
 MATHPIX_APP_KEY=""
 
 POSTGRES_URL=""
+DAILY_LLM_CALL_LIMIT="100"
+DAILY_LLM_TOKEN_LIMIT="50000"
 
 FRONTEND_URL="http://localhost:5173"
 
@@ -28,4 +30,4 @@ OLLAMA_TIMEOUT="60"
 MISTRAL_CHAT_MODEL="mistral-large-3-675b-instruct-2512"
 MISTRAL_API_KEY=""
 MISTRAL_CHAT_TIMEOUT="60"
-MISTRAL_CHAT_TEMPERATURE="0.2"
\ No newline at end of file
+MISTRAL_CHAT_TEMPERATURE="0.2"

math-tutor/backend/app/api/health.py

@@ -9,6 +9,7 @@ import psycopg
 from fastapi import APIRouter
 
 import app.config as config
+from app.deterministic_services import llm_quota
 
 router = APIRouter()
 logger = logging.getLogger(__name__)
@@ -98,12 +99,36 @@ def _check_postgres() -> dict:
         return {"status": "error", "detail": str(exc)}
 
 
+def _check_llm_quota() -> dict:
+    try:
+        pg_url = config.get_postgres_url()
+    except ValueError:
+        return {"status": "missing_config"}
+
+    try:
+        today_usage = llm_quota.get_today_usage(pg_url)
+        limits = config.get_llm_quota_settings()
+        return {
+            "status": "ok",
+            "usage_date": today_usage.usage_date.isoformat(),
+            "call_count": today_usage.call_count,
+            "token_count": today_usage.token_count,
+            "limits": {
+                "daily_call_limit": limits.daily_call_limit,
+                "daily_token_limit": limits.daily_token_limit,
+            },
+        }
+    except Exception as exc:
+        return {"status": "error", "detail": str(exc)}
+
+
 @router.get("/api/health")
 def health() -> Dict[str, Any]:
     services = {
         "ollama": _check_ollama(),
         "openai": _check_openai(),
         "postgres": _check_postgres(),
+        "llm_quota": _check_llm_quota(),
     }
     required_statuses = {"ok"}
     overall = "ok" if all(

math-tutor/backend/app/config.py

@@ -92,6 +92,12 @@ class MathpixSettings:
     app_key: str
 
 
+@dataclass(frozen=True)
+class LLMQuotaSettings:
+    daily_call_limit: int | None
+    daily_token_limit: int | None
+
+
 def _read_float(value: str | None) -> float | None:
     if value is None or value == "":
         return None
@@ -101,6 +107,15 @@ def _read_float(value: str | None) -> float | None:
         return None
 
 
+def _read_int(value: str | None) -> int | None:
+    if value is None or value == "":
+        return None
+    try:
+        return int(value)
+    except ValueError:
+        return None
+
+
 def get_mathpix_settings() -> MathpixSettings:
     return MathpixSettings(
         app_id=os.getenv("MATHPIX_APP_ID"),
@@ -172,3 +187,10 @@ def get_postgres_url() -> str:
     if not pg_url:
         raise ValueError("Missing POSTGRES_URL")
     return pg_url
+
+
+def get_llm_quota_settings() -> LLMQuotaSettings:
+    return LLMQuotaSettings(
+        daily_call_limit=_read_int(os.getenv("DAILY_LLM_CALL_LIMIT")),
+        daily_token_limit=_read_int(os.getenv("DAILY_LLM_TOKEN_LIMIT")),
+    )

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

 import inspect
 import json
+from datetime import date
 from typing import Any, Callable
 
 import ollama
@@ -7,6 +8,7 @@ from openai import OpenAI
 from mistralai.client import Mistral
 
 from app import config
+from app.deterministic_services import llm_quota
 
 
 def _filter_kwargs(func, kwargs: dict) -> dict:
@@ -23,6 +25,39 @@ def _extract_message(response) -> object:
         return response.message
     return {}
 
+
+def _extract_total_tokens(response: object) -> int:
+    usage = None
+    if isinstance(response, dict):
+        usage = response.get("usage")
+    elif hasattr(response, "usage"):
+        usage = getattr(response, "usage")
+
+    if usage is None:
+        return 0
+
+    if isinstance(usage, dict):
+        total = usage.get("total_tokens")
+        if total is not None:
+            return int(total or 0)
+        prompt = usage.get("prompt_tokens") or usage.get("input_tokens") or 0
+        completion = usage.get("completion_tokens") or usage.get("output_tokens") or 0
+        return int(prompt) + int(completion)
+
+    total = getattr(usage, "total_tokens", None)
+    if total is not None:
+        return int(total or 0)
+    prompt = getattr(usage, "prompt_tokens", None) or getattr(usage, "input_tokens", None) or 0
+    completion = getattr(usage, "completion_tokens", None) or getattr(usage, "output_tokens", None) or 0
+    return int(prompt) + int(completion)
+
+
+def _record_call(result: dict, tokens: int | None = None) -> dict:
+    pg_url = config.get_postgres_url()
+    token_count = _extract_total_tokens(result.get("raw")) if tokens is None else tokens
+    llm_quota.record_usage(pg_url, date.today(), calls=1, tokens=token_count)
+    return result
+
 def _chat_openai(messages: list[dict]) -> dict:
     settings = config.get_openai_chat_settings()
     if not settings:
@@ -77,21 +112,72 @@ def chat(
     use_ollama: bool = False,
     use_mistral: bool = False,
 ) -> dict:
+    quota_settings = config.get_llm_quota_settings()
+    pg_url = config.get_postgres_url()
     if use_mistral and tools:
         raise RuntimeError("Mistral chat is currently only implemented for calls without tools.")
 
     if use_mistral:
-        mistral_result = _chat_mistral(messages)
+        mistral_settings = config.get_mistral_chat_settings()
+        if not mistral_settings:
+            return {}
+        llm_quota.ensure_within_limits(
+            pg_url,
+            date.today(),
+            call_limit=quota_settings.daily_call_limit,
+            token_limit=quota_settings.daily_token_limit,
+            add_calls=1,
+            add_tokens=0,
+        )
+        try:
+            mistral_result = _chat_mistral(messages)
+        except Exception:
+            _record_call({"raw": None}, tokens=0)
+            raise
         if mistral_result:
-            return mistral_result
+            return _record_call(mistral_result)
+        _record_call({"raw": None}, tokens=0)
+        return mistral_result
 
     if not tools and not use_ollama:
-        openai_result = _chat_openai(messages)
-        if openai_result:
+        openai_settings = config.get_openai_chat_settings()
+        if openai_settings:
+            llm_quota.ensure_within_limits(
+                pg_url,
+                date.today(),
+                call_limit=quota_settings.daily_call_limit,
+                token_limit=quota_settings.daily_token_limit,
+                add_calls=1,
+                add_tokens=0,
+            )
+            try:
+                openai_result = _chat_openai(messages)
+            except Exception:
+                _record_call({"raw": None}, tokens=0)
+                raise
+            if openai_result:
+                return _record_call(openai_result)
+            _record_call({"raw": None}, tokens=0)
             return openai_result
 
-        mistral_result = _chat_mistral(messages)
-        if mistral_result:
+        mistral_settings = config.get_mistral_chat_settings()
+        if mistral_settings:
+            llm_quota.ensure_within_limits(
+                pg_url,
+                date.today(),
+                call_limit=quota_settings.daily_call_limit,
+                token_limit=quota_settings.daily_token_limit,
+                add_calls=1,
+                add_tokens=0,
+            )
+            try:
+                mistral_result = _chat_mistral(messages)
+            except Exception:
+                _record_call({"raw": None}, tokens=0)
+                raise
+            if mistral_result:
+                return _record_call(mistral_result)
+            _record_call({"raw": None}, tokens=0)
             return mistral_result
 
     settings = config.get_ollama_settings()

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

+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import date
+from typing import Any
+
+import psycopg
+from psycopg.rows import dict_row
+
+
+DATABASE_CREATION_SQL = """
+CREATE TABLE IF NOT EXISTS llm_daily_usage (
+  usage_date DATE PRIMARY KEY,
+  call_count BIGINT NOT NULL DEFAULT 0,
+  token_count BIGINT NOT NULL DEFAULT 0,
+  updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
+);
+"""
+
+
+@dataclass(frozen=True)
+class LLMDailyUsage:
+    usage_date: date
+    call_count: int
+    token_count: int
+    updated_at: str | None = None
+
+
+def init_db(pg_url: str) -> None:
+    with psycopg.connect(pg_url) as conn:
+        with conn.cursor() as cur:
+            cur.execute(DATABASE_CREATION_SQL)
+        conn.commit()
+
+
+def _ensure_row(conn: psycopg.Connection[Any], usage_date: date) -> None:
+    with conn.cursor() as cur:
+        cur.execute(
+            """
+            INSERT INTO llm_daily_usage (usage_date)
+            VALUES (%s)
+            ON CONFLICT (usage_date) DO NOTHING
+            """,
+            (usage_date,),
+        )
+
+
+def get_daily_usage(pg_url: str, usage_date: date) -> LLMDailyUsage:
+    init_db(pg_url)
+    with psycopg.connect(pg_url, row_factory=dict_row) as conn:
+        with conn.cursor() as cur:
+            cur.execute(
+                """
+                SELECT usage_date, call_count, token_count, updated_at
+                FROM llm_daily_usage
+                WHERE usage_date = %s
+                """,
+                (usage_date,),
+            )
+            row = cur.fetchone()
+    if not row:
+        return LLMDailyUsage(usage_date=usage_date, call_count=0, token_count=0)
+    return LLMDailyUsage(
+        usage_date=row["usage_date"],
+        call_count=int(row["call_count"]),
+        token_count=int(row["token_count"]),
+        updated_at=row.get("updated_at").isoformat() if row.get("updated_at") else None,
+    )
+
+
+def ensure_within_limits(
+    pg_url: str,
+    usage_date: date,
+    *,
+    call_limit: int | None,
+    token_limit: int | None,
+    add_calls: int = 1,
+    add_tokens: int = 0,
+) -> None:
+    init_db(pg_url)
+    with psycopg.connect(pg_url) as conn:
+        with conn.transaction():
+            _ensure_row(conn, usage_date)
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    SELECT call_count, token_count
+                    FROM llm_daily_usage
+                    WHERE usage_date = %s
+                    FOR UPDATE
+                    """,
+                    (usage_date,),
+                )
+                row = cur.fetchone() or (0, 0)
+                next_calls = int(row[0]) + add_calls
+                next_tokens = int(row[1]) + add_tokens
+
+                if call_limit is not None and next_calls > call_limit:
+                    raise QuotaExceededError("daily LLM call limit exceeded")
+                if token_limit is not None and next_tokens > token_limit:
+                    raise QuotaExceededError("daily LLM token limit exceeded")
+
+
+def record_usage(
+    pg_url: str,
+    usage_date: date,
+    *,
+    calls: int = 1,
+    tokens: int = 0,
+) -> None:
+    init_db(pg_url)
+    with psycopg.connect(pg_url) as conn:
+        with conn.transaction():
+            _ensure_row(conn, usage_date)
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    UPDATE llm_daily_usage
+                    SET call_count = call_count + %s,
+                        token_count = token_count + %s,
+                        updated_at = NOW()
+                    WHERE usage_date = %s
+                    """,
+                    (calls, tokens, usage_date),
+                )
+
+
+def get_today_usage(pg_url: str) -> LLMDailyUsage:
+    return get_daily_usage(pg_url, date.today())
+
+
+class QuotaExceededError(RuntimeError):
+    pass

math-tutor/backend/app/main.py

@@ -5,7 +5,8 @@ from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 from app.api import canvas, chat, context, health, orchestrator, tasks
 from app.config import get_frontend_url
-from app.deterministic_services import embedding_provider
+from app.deterministic_services import embedding_provider, llm_quota
+from app import config
 
 
 logger = logging.getLogger(__name__)
@@ -23,6 +24,15 @@ async def lifespan(_app: FastAPI):
         yield
         return
 
+    try:
+        llm_quota.init_db(config.get_postgres_url())
+        logger.info("LLM quota table ensured")
+    except Exception:
+        logger.exception("LLM quota init failed")
+        health.set_readiness_failed("LLM quota init failed")
+        yield
+        return
+
     try:
         warmup_timing = embedding_provider.warmup_embedder()
         logger.info("Embedding warmup finished: %s", warmup_timing)