f6d3ac05a5
Backend:
- Fix TicketInput ticket_id type (int|str -> str) — Pydantic 500 on numeric IDs
- Fix LLM module: add missing logger import, handle 429 and non-200 responses
- Switch LLM model to poolside/laguna-m.1:free (avoid OpenRouter rate limits)
- Add GET /api/collections/{id} endpoint to fetch single collection
- Deduplicate tickets: replace randomized hash() with deterministic
hashlib.md5 for Qdrant point IDs
- Add get_existing_ticket_ids() — scroll Qdrant before processing,
skip already-imported tickets, avoid redundant embeddings
- Multi-tenant isolation: tenant_id + collection_id in all Qdrant filters
- Fix embedding: drop incompatible openai SDK, use raw httpx
- Fix reranking: access ScoredPoint.payload via .get(), not []
- Fix chunking: expand role detection (support/operator/agent + client/user/customer)
- Fix bcrypt: downgrade to 4.0.1 (incompatible 5.x with passlib)
Frontend:
- Register: add plan selection (free/starter/pro)
- CollectionView: fetch ticket_count from backend on mount
(not only from upload response — survived page refresh)
- uploadFile: use getToken() instead of stale module-level var,
show real HTTP status in error messages
- Nginx: increase client_max_body_size to 50M for file uploads
65 lines
2.5 KiB
Python
65 lines
2.5 KiB
Python
import logging
|
||
|
||
import httpx
|
||
|
||
from app.config import settings
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# Increase log level for httpx to reduce noise
|
||
logging.getLogger("httpx").setLevel(logging.WARNING)
|
||
|
||
OPENROUTER_BASE = "https://openrouter.ai/api/v1"
|
||
|
||
_client: httpx.AsyncClient | None = None
|
||
|
||
|
||
def get_client() -> httpx.AsyncClient:
|
||
global _client
|
||
if _client is None:
|
||
_client = httpx.AsyncClient(
|
||
base_url=OPENROUTER_BASE,
|
||
headers={
|
||
"Authorization": f"Bearer {settings.openrouter_api_key}",
|
||
"Content-Type": "application/json",
|
||
},
|
||
timeout=60,
|
||
)
|
||
return _client
|
||
|
||
|
||
SYSTEM_PROMPT = """Ты — специалист техподдержки. Отвечай клиенту, используя ТОЛЬКО информацию из переданных тикетов (контекст).
|
||
|
||
Правила:
|
||
- Если контекст содержит подходящее решение — напиши ответ своими словами, адаптируя под вопрос
|
||
- Если контекст не относится к вопросу — напиши: «Недостаточно информации в истории обращений»
|
||
- НЕ придумывай ответы, НЕ используй общие знания
|
||
- НЕ говори «обратитесь в службу поддержки» — ты сам и есть поддержка
|
||
- Укажи в конце: «Основано на тикете №...»"""
|
||
|
||
|
||
async def generate_answer(query: str, context: str) -> str | None:
|
||
client = get_client()
|
||
response = await client.post(
|
||
"/chat/completions",
|
||
json={
|
||
"model": settings.openrouter_llm_model,
|
||
"messages": [
|
||
{"role": "system", "content": SYSTEM_PROMPT},
|
||
{
|
||
"role": "user",
|
||
"content": f"Контекст (история обращений):\n{context}\n\nВопрос клиента:\n{query}",
|
||
},
|
||
],
|
||
"temperature": 0.1,
|
||
"max_tokens": 2000,
|
||
},
|
||
)
|
||
if response.status_code != 200:
|
||
logger.error("LLM API error [%s]: %s", response.status_code, await response.aread())
|
||
return None
|
||
data = response.json()
|
||
if "choices" not in data or not data["choices"]:
|
||
logger.error("LLM response missing choices: %s", data.get("error", data))
|
||
return None
|
||
return data["choices"][0]["message"]["content"]
|