src/inference_service.py

"""Small, fail-closed provider adapter shared by REST and Gradio.

Configured routes are not health claims. Explicit model choices never fall back
to another provider; ``auto`` may choose from configured routes.
"""
import json
import os
import re
import time
import urllib.error
import urllib.parse
import urllib.request


ROUTES = {
    "qwen": ("QWEN", "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "qwen-plus"),
    "llama": ("GROQ", "https://api.groq.com/openai/v1", "llama-3.3-70b-versatile"),
    "deepseek": ("DEEPSEEK", "https://api.deepseek.com", "deepseek-chat"),
    "gemini": ("GEMINI", "https://generativelanguage.googleapis.com/v1beta/openai", "gemini-2.5-flash"),
    "nemotron": ("OPENROUTER", "https://openrouter.ai/api/v1", "nvidia/nemotron-3-nano-30b-a3b:free"),
    "gpt-oss": ("OPENROUTER", "https://openrouter.ai/api/v1", "openai/gpt-oss-20b:free"),
    "lfm": ("OPENROUTER", "https://openrouter.ai/api/v1", "liquid/lfm-2.5-1.2b-instruct:free"),
    "openrouter-free": ("OPENROUTER", "https://openrouter.ai/api/v1", "openrouter/free"),
}
MODEL_ENV = {
    "nemotron": "OPENROUTER_NEMOTRON_MODEL_NAME",
    "gpt-oss": "OPENROUTER_GPT_OSS_MODEL_NAME",
    "lfm": "OPENROUTER_LFM_MODEL_NAME",
    "openrouter-free": "OPENROUTER_FREE_MODEL_NAME",
}
ERROR_TEXT = re.compile(
    r"^(?:AI[_ ]GENERATION[_ ]FAILED|AI generation failed|Transcription (?:failed|unavailable)|"
    r"Translation (?:failed|unavailable)|Error\s*:|Service unavailable)", re.I
)


def failure(code, message):
    return {"ok": False, "error": {"code": code, "message": message}}


def valid_content(text):
    return isinstance(text, str) and 0 < len(text.strip()) <= 4000 and not ERROR_TEXT.search(text.strip())


def _route(choice):
    prefix, default_url, default_model = ROUTES[choice]
    key = os.getenv(prefix + "_API_KEY", "").strip()
    if prefix == "GEMINI":
        key = key or os.getenv("GOOGLE_API_KEY", "").strip()
    if not key or key.lower() in {"your-api-key-here", "none", "null"}:
        return None
    model = os.getenv(MODEL_ENV.get(choice, prefix + "_MODEL_NAME"), default_model).strip()
    base_url = os.getenv(prefix + "_BASE_URL", default_url).strip().rstrip("/")
    parsed = urllib.parse.urlsplit(base_url)
    # Only operator-configured endpoints are used. Never put keys in URLs.
    if parsed.scheme not in {"https", "http"} or not parsed.hostname or parsed.username or parsed.password or parsed.query or parsed.fragment:
        return None
    if parsed.scheme == "http" and parsed.hostname not in {"localhost", "127.0.0.1", "::1"}:
        return None
    if not model:
        return None
    return {"key": key, "url": base_url + "/chat/completions", "model": model, "route": choice}


def available_routes():
    """Return non-secret configuration metadata, without claiming availability."""
    return [
        {"route": name, "model": route["model"], "configured": True, "verified": False}
        for name in ROUTES if (route := _route(name))
    ]


class _NoRedirect(urllib.request.HTTPRedirectHandler):
    def redirect_request(self, req, fp, code, msg, headers, newurl):
        # Do not forward Authorization to a redirected host.
        return None


class _OutputFailure(Exception):
    def __init__(self, code, message):
        super().__init__(message)
        self.code = code
        self.message = message


def _post(route, prompt, timeout=10):
    body = json.dumps({
        "model": route["model"],
        "messages": [{"role": "user", "content": prompt}],
        "temperature": 0.3,
        "max_tokens": 768,
    }).encode("utf-8")
    request = urllib.request.Request(route["url"], data=body, headers={
        "Authorization": "Bearer " + route["key"],
        "Content-Type": "application/json",
    }, method="POST")
    with urllib.request.build_opener(_NoRedirect).open(request, timeout=timeout) as response:
        raw = response.read(1024 * 1024 + 1)
    if len(raw) > 1024 * 1024:
        raise ValueError("Response exceeded limit")
    payload = json.loads(raw)
    choice = payload["choices"][0]
    finish_reason = choice.get("finish_reason")
    if finish_reason in {"length", "max_tokens"}:
        raise _OutputFailure("OUTPUT_TRUNCATED", "The model response was cut short. Shorten the text or retry.")
    if finish_reason != "stop" or choice.get("message", {}).get("tool_calls"):
        raise _OutputFailure("INVALID_OUTPUT", "The model did not return a complete usable response.")
    return choice["message"]["content"]


def generate_text(prompt, model_choice="auto", *, deadline=None):
    if not isinstance(prompt, str) or not prompt.strip() or len(prompt) > 24000:
        return failure("INVALID_REQUEST", "A non-empty, bounded inference prompt is required.")
    choice = str(model_choice or "auto").strip().lower()
    if choice != "auto" and choice not in ROUTES:
        return failure("MODEL_UNAVAILABLE", "The selected model route is not supported.")
    names = list(ROUTES) if choice == "auto" else [choice]
    configured = [route for name in names if (route := _route(name))]
    if not configured:
        return failure("MODEL_UNAVAILABLE", "The selected model route is not configured.")
    deadline = min(deadline or (time.monotonic() + 30), time.monotonic() + 30)
    attempts = []
    result = failure("PROVIDER_UNAVAILABLE", "The selected provider is unavailable. Please retry.")
    for route in configured[:3]:
        remaining = deadline - time.monotonic()
        if remaining <= 0:
            break
        try:
            content = _post(route, prompt, timeout=min(10, remaining))
            if valid_content(content):
                return {"ok": True, "text": content.strip(), "model": route["model"], "route": route["route"], "attempts": attempts}
            result = failure("INVALID_OUTPUT", "The model did not return usable content.")
        except _OutputFailure as exc:
            result = failure(exc.code, exc.message)
        except urllib.error.HTTPError as exc:
            if exc.code in {401, 403}:
                result = failure("PROVIDER_AUTH_FAILED", "The selected provider rejected its credentials.")
            elif exc.code == 429:
                result = failure("PROVIDER_RATE_LIMITED", "The selected provider is rate limited or has no remaining quota.")
            elif exc.code in {400, 404, 422}:
                result = failure("PROVIDER_CONFIGURATION_ERROR", "The selected provider rejected the configured model or request.")
            else:
                result = failure("PROVIDER_UNAVAILABLE", "The selected provider is unavailable. Please retry.")
        except (TimeoutError, urllib.error.URLError):
            result = failure("PROVIDER_UNAVAILABLE", "The selected provider could not be reached. Please retry.")
        except Exception:
            result = failure("INVALID_OUTPUT", "The model did not return usable content.")
        attempts.append({"route": route["route"], "code": result["error"]["code"]})
    return {**result, "attempts": attempts}