naeel
a684d51169
h2 — pure Python, ставится без компиляции в slim-контейнерах. ssl.create_default_context() + ALPN h2 + h2.connection.H2Connection.
148 lines
4.9 KiB
Python
148 lines
4.9 KiB
Python
"""
|
||
llm_client.py — Тонкий HTTP-клиент к LLM API.
|
||
|
||
Только отправить промпт → получить ответ. Никакой бизнес-логики.
|
||
Бизнес-логика (промпты, парсинг ответа) → extractor.py.
|
||
|
||
Использует h2 + ssl + socket — HTTP/2 на чистом Python.
|
||
api.aillm.ru за ddos-guard, требует HTTP/2.
|
||
h2 — pure Python, ставится без компиляции в slim-контейнерах.
|
||
|
||
Переменные окружения:
|
||
LLM_API_URL — URL API (по умолчанию https://api.aillm.ru/v1/chat/completions)
|
||
LLM_API_KEY — ключ авторизации
|
||
"""
|
||
|
||
import os
|
||
import json
|
||
import ssl
|
||
import socket
|
||
import h2.connection
|
||
import h2.events
|
||
from urllib.parse import urlparse
|
||
|
||
# ── Конфигурация ────────────────────────────────────────────────
|
||
|
||
DEFAULT_URL = "https://api.aillm.ru/v1/chat/completions"
|
||
DEFAULT_MODEL = "gpt-oss-120b" # 120B модель через LiteLLM-прокси
|
||
|
||
|
||
def _http2_request(url: str, headers: dict, body: str, timeout: int = 120) -> dict:
|
||
"""
|
||
HTTP/2 POST-запрос на чистом Python (h2 + ssl + socket).
|
||
Возвращает (status_code, response_body) или бросает исключение.
|
||
"""
|
||
parsed = urlparse(url)
|
||
host = parsed.hostname
|
||
port = parsed.port or 443
|
||
|
||
# SSL-контекст с ALPN = h2 (HTTP/2)
|
||
ctx = ssl.create_default_context()
|
||
ctx.set_alpn_protocols(["h2"])
|
||
|
||
sock = socket.create_connection((host, port), timeout=timeout)
|
||
ssl_sock = ctx.wrap_socket(sock, server_hostname=host)
|
||
|
||
conn = h2.connection.H2Connection()
|
||
conn.initiate_connection()
|
||
ssl_sock.sendall(conn.data_to_send())
|
||
|
||
# Формируем заголовки
|
||
request_headers = [
|
||
(":method", "POST"),
|
||
(":path", parsed.path + ("?" + parsed.query if parsed.query else "")),
|
||
(":authority", host),
|
||
(":scheme", "https"),
|
||
("content-type", "application/json"),
|
||
("accept", "application/json"),
|
||
]
|
||
for k, v in headers.items():
|
||
request_headers.append((k.lower(), v))
|
||
|
||
conn.send_headers(1, request_headers, end_stream=False)
|
||
body_bytes = body.encode("utf-8")
|
||
conn.send_data(1, body_bytes, end_stream=True)
|
||
ssl_sock.sendall(conn.data_to_send())
|
||
|
||
# Читаем ответ
|
||
response_data = b""
|
||
response_headers = {}
|
||
|
||
while True:
|
||
data = ssl_sock.recv(65535)
|
||
if not data:
|
||
break
|
||
|
||
events = conn.receive_data(data)
|
||
for event in events:
|
||
if isinstance(event, h2.events.ResponseReceived):
|
||
response_headers = dict(event.headers)
|
||
elif isinstance(event, h2.events.DataReceived):
|
||
response_data += event.data
|
||
conn.acknowledge_received_data(event.flow_controlled_length, event.stream_id)
|
||
elif isinstance(event, h2.events.StreamEnded):
|
||
ssl_sock.sendall(conn.data_to_send())
|
||
ssl_sock.close()
|
||
status = int(response_headers.get(b":status", b"0"))
|
||
return status, response_data.decode("utf-8")
|
||
|
||
ssl_sock.sendall(conn.data_to_send())
|
||
|
||
ssl_sock.close()
|
||
raise Exception("HTTP/2 stream ended without response")
|
||
|
||
|
||
def ask(prompt: str, model: str = None, max_tokens: int = 8000) -> dict:
|
||
"""
|
||
Отправить промпт к LLM API через HTTP/2 (h2).
|
||
|
||
Args:
|
||
prompt: текст промпта
|
||
model: модель (None = DEFAULT_MODEL)
|
||
max_tokens: максимум токенов в ответе
|
||
|
||
Returns:
|
||
{"text": "ответ модели", "model": "...", "usage": {...}}
|
||
или
|
||
{"error": "описание ошибки"}
|
||
"""
|
||
api_url = os.getenv("LLM_API_URL", DEFAULT_URL)
|
||
api_key = os.getenv("LLM_API_KEY", "")
|
||
|
||
if not api_key:
|
||
return {"error": "LLM_API_KEY not set"}
|
||
|
||
model = model or DEFAULT_MODEL
|
||
|
||
payload = json.dumps({
|
||
"model": model,
|
||
"messages": [{"role": "user", "content": prompt}],
|
||
"max_tokens": max_tokens,
|
||
"temperature": 0.1,
|
||
})
|
||
|
||
headers = {"authorization": f"Bearer {api_key}"}
|
||
|
||
try:
|
||
status, body = _http2_request(api_url, headers, payload)
|
||
|
||
if status != 200:
|
||
return {"error": f"LLM HTTP {status}: {body[:500]}"}
|
||
|
||
data = json.loads(body)
|
||
|
||
choice = data.get("choices", [{}])[0]
|
||
message = choice.get("message", {})
|
||
text = message.get("content", "")
|
||
|
||
return {
|
||
"text": text,
|
||
"model": data.get("model", model),
|
||
"usage": data.get("usage", {}),
|
||
}
|
||
|
||
except socket.timeout:
|
||
return {"error": "LLM timeout"}
|
||
except Exception as e:
|
||
return {"error": f"LLM unexpected: {e}"}
|