Ideal routing: voice transcribed into scenario routing, OR qwen newcomer in light chain

bot.py

@@ -245,6 +245,8 @@ OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
 # Groq — прямой провайдер лёгкого текста (калибровка 21.09.2026): 1000 запросов/день против 50 у OpenRouter.
 GROQ_API_KEY = os.getenv("GROQ_API_KEY", "").strip()
 GROQ_BASE_URL = "https://api.groq.com/openai/v1"
+# Голосовые длиннее не транскрибируем (Whisper API всё равно режет 25 МБ) — такие идут прежним путём через Gemini-аудио.
+VOICE_TRANSCRIBE_MAX_BYTES = int(os.getenv("VOICE_TRANSCRIBE_MAX_BYTES", str(10 * 1024 * 1024)))
 # Раньше у OpenRouter был свой отдельный лимит истории (30), меньший, чем у Gemini
 # (100) — при переключении провайдера (/provider или /model) ощущалось резкое
 # "обнуление" контекста разговора. Теперь история ОБЩАЯ (см. state["history"] в
@@ -713,6 +715,7 @@ __all__ = [
     "ask_openrouter_text",
     "ask_openrouter_multimodal",
     "ask_groq_text",
+    "_transcribe_audio",
     "_gemini_history_contents",
     "_build_gemma_identity_contents",
     "_build_gemini_call_config",
@@ -945,6 +948,7 @@ from lumen_routes import (
     ask_openrouter_text,
     ask_openrouter_multimodal,
     ask_groq_text,
+    _transcribe_audio,
     _gemini_history_contents,
     _build_gemma_identity_contents,
     _build_gemini_call_config,

docs/ARCHITECTURE.md

@@ -15,7 +15,7 @@ This document goes one level deeper than the main [README](../README.md) into ho
 Routing is fully automatic: the router builds a fresh candidate list for every message:
 
 - **A YouTube or website link** → Gemini only. It's the only provider that can read page content (`url_context`) or analyze a video by URL.
-- **A video or audio attachment** → Gemini only. OpenRouter's multimodal models only accept images as base64.
+- **A video or audio attachment** → video goes to Gemini only (OpenRouter's multimodal models only accept images as base64). Voice/audio is first transcribed cheaply (Groq Whisper) and the text joins the normal routing below; if transcription fails or is unavailable, the audio goes to Gemini as before.
 - **An image attachment, no live-info need** → free OpenRouter vision models first, Gemini as a reserve.
 - **Needs current information** (a lightweight keyword heuristic: "now," "today," "price," "who is currently...") → Gemini, prioritizing the models with a real search-grounding quota, then OpenRouter as a reserve, Groq last (knowledge-only answer if both are down).
 - **Plain text, no attachments, no freshness need** (the most common case) → Groq first (1000 free requests/day: Qwen, then gpt-oss), then OpenRouter. Heavier requests (code, multi-step reasoning, caught by another lightweight heuristic) get routed to the stronger free OpenRouter models first.

docs/ENVIRONMENT.md

@@ -48,10 +48,11 @@ Before publishing, configure the same `LUMEN_PROXY_SECRET` on HF and every Deno
 | Variable | Default | Purpose |
 |---|---|---|
 | `OPENROUTER_API_KEY` / `OPENROUTER_KEY` | — | OpenRouter API key. Without it, the router sends everything to Gemini, which burns through its much smaller quota fast. |
-| `GROQ_API_KEY` | — | Groq API key (free plan, no card: 1000 req/day). Heads plain-text routes; without it OpenRouter heads them as before. |
+| `GROQ_API_KEY` | — | Groq API key (free plan, no card: 1000 req/day). Heads plain-text routes; without it OpenRouter heads them as before. Also powers voice transcription (Whisper). |
 | `OPENROUTER_HTTP_REFERER` | `https://t.me/{BOT_USERNAME}` | `HTTP-Referer` header sent with OpenRouter requests. |
 | `OPENROUTER_TITLE` | `BOT_USERNAME` | App title header sent with OpenRouter requests. |
 | `POLLINATIONS_IMAGE_MODEL` | `flux` | Default image-generation model (used only if the prompt doesn't match a more specific style). Renamed from `HF_IMAGE_MODEL` (Sept 2026) — the backend is Pollinations.ai, not Hugging Face; the old name is no longer read. |
+| `VOICE_TRANSCRIBE_MAX_BYTES` | `10485760` (10 MB) | Voice/audio attachments larger than this skip Whisper transcription and go straight to Gemini audio, as before. |
 
 ## Timeouts, rate limiting & routing budgets
 

lumen_message_core.py

@@ -262,6 +262,16 @@ async def _handle_message_core(message: Message, extra_media: list[tuple[bytes,
         message, state, clean_prompt, is_private=is_private,
     )
 
+    # Голос/аудио: сначала дешёвая транскрибация — дальше текст идёт общим роутингом по
+    # сценарию (тяжесть/свежесть определяются по сказанному). Не вышло — падает в прежний
+    # путь: аудио напрямую в Gemini (см. is_video_or_audio_media в роутере).
+    if media_tuple and media_tuple[1].startswith("audio/"):
+        transcript = await bot._transcribe_audio(media_tuple[0], message.chat.id)
+        if transcript:
+            marked = "[Расшифровка голосового]: " + transcript
+            clean_prompt = (clean_prompt + "\n" + marked).strip() if clean_prompt else marked
+            media_tuple = None
+
     if media_tuple and not clean_prompt:
          clean_prompt = _ensure_prompt_text(None, media_tuple[1])
     if youtube_url_to_analyze and not clean_prompt:

lumen_router_config.py

@@ -180,6 +180,7 @@ _KNOWN_MODEL_IDS_FOR_LEAK_DETECTION: list[str] = [
     "inclusionai/ling-3.0-flash-fin:free",
     "inclusionai/ling-3.0-flash-vl:free",
     "liquid/lfm-2.5-2.6b:free",
+    "qwen/qwen3.8-27b:free",
     "gemini-3.8-flash",
     "openrouter/free",
     # Groq-ID без :free-суффикса — те же семейства, что выше через OpenRouter; дословно в ответе им тоже не место.
@@ -320,6 +321,9 @@ _OR_LIGHT_ORDER: list[str] = [
     "inclusionai/ling-3.0-flash-sante:free",
     "inclusionai/ling-3.0-flash-fin:free",
     "liquid/lfm-2.5-2.6b:free",
+    # Аудит 21.09.2026 (живой каталог): qwen3.8-27b:free — то же семейство, что калиброванный
+    # Groq-Qwen, но другой эндпоинт: некалиброван, поэтому после проверенных, перед резервом.
+    "qwen/qwen3.8-27b:free",
     "openrouter/free",
 ]
 

lumen_routes.py

@@ -281,6 +281,46 @@ async def ask_groq_text(chat_id: int, user_text: str, model_chain: list[str], *,
     bot._record_quota_usage("groq", model_trial)
     return answer
 
+async def _transcribe_audio(audio_bytes: bytes, chat_id: int) -> str | None:
+    """Голос/аудио в текст через Groq Whisper (дешевле Gemini-квоты на порядок). None при
+    любой неудаче — вызывающий код молча идёт прежним путём (аудио напрямую в Gemini)."""
+    import bot
+    if not bot.GROQ_API_KEY or not audio_bytes:
+        return None
+    if len(audio_bytes) > bot.VOICE_TRANSCRIBE_MAX_BYTES:
+        log.warning("[groq] Voice message too big for transcription (%d bytes), leaving to Gemini.", len(audio_bytes))
+        return None
+    lang = bot._chat_lang(chat_id)
+    form = aiohttp.FormData()
+    form.add_field("file", audio_bytes, filename="voice.ogg", content_type="audio/ogg")
+    form.add_field("model", "whisper-large-v3-turbo")
+    if isinstance(lang, str) and len(lang) == 2:
+        form.add_field("language", lang)
+    try:
+        session = await bot._get_http_session()
+        async with session.post(
+            f"{bot.GROQ_BASE_URL}/audio/transcriptions",
+            headers={"Authorization": f"Bearer {bot.GROQ_API_KEY}"},
+            data=form,
+            timeout=aiohttp.ClientTimeout(total=bot.ROUTE_MODEL_TIMEOUT_SEC, connect=10.0),
+        ) as resp:
+            if resp.status >= 400:
+                body = await resp.read()
+                log.warning("[groq] Whisper HTTP %s: %r", resp.status, body[:200])
+                return None
+            data = await resp.json(content_type=None)
+    except Exception as exc:
+        exc_str = str(exc) or repr(exc) or exc.__class__.__name__
+        if bot.GROQ_API_KEY:
+            exc_str = exc_str.replace(bot.GROQ_API_KEY, "<KEY>")
+        log.warning("[groq] Whisper request failed: %s", exc_str)
+        return None
+    text = (data.get("text") or "").strip() if isinstance(data, dict) else ""
+    if not text:
+        return None
+    bot._record_quota_usage("groq", "whisper-large-v3-turbo")
+    return text
+
 async def ask_openrouter_multimodal(
     chat_id: int, user_text: str, media_tuple: tuple[bytes, str], media_filename: str,
     model_chain: list[str], *, deadline: float | None = None,

tests/test_bot_routes.py

@@ -1020,3 +1020,65 @@ def test_route_error_reply_text_maps_groq_error():
     assert "Rate limit reached" not in text
     assert text.strip() != ""
 
+
+def _fake_groq_audio_session(response_json, status=200):
+    class _FakeResp:
+        def __init__(self):
+            self.status = status
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *args):
+            return False
+
+        async def read(self):
+            return b"{}"
+
+        async def json(self, content_type=None):
+            return response_json
+
+    class _FakeSession:
+        def post(self, *args, **kwargs):
+            return _FakeResp()
+
+    async def fake_get_http_session():
+        return _FakeSession()
+
+    return fake_get_http_session
+
+
+def test_transcribe_audio_success_records_quota(monkeypatch):
+    # Groq Whisper отдал текст — пишем расход groq/whisper и возвращаем текст.
+    monkeypatch.setattr(bot, "_get_http_session", _fake_groq_audio_session({"text": "  привет мир  "}))
+    monkeypatch.setattr(bot, "GROQ_API_KEY", "fake-key")
+    recorded = {}
+    monkeypatch.setattr(bot, "_record_quota_usage", lambda provider, model: recorded.setdefault("v", (provider, model)))
+    assert asyncio.run(bot._transcribe_audio(b"ogg-bytes", 123)) == "привет мир"
+    assert recorded["v"] == ("groq", "whisper-large-v3-turbo")
+
+
+def test_transcribe_audio_empty_result_is_none(monkeypatch):
+    # Пустая расшифровка — как неудача: вызывающий код идёт прежним путём.
+    monkeypatch.setattr(bot, "_get_http_session", _fake_groq_audio_session({"text": "   "}))
+    monkeypatch.setattr(bot, "GROQ_API_KEY", "fake-key")
+    assert asyncio.run(bot._transcribe_audio(b"ogg-bytes", 123)) is None
+
+
+def test_transcribe_audio_no_key_or_oversize_skips_network(monkeypatch):
+    async def must_not_be_called():
+        raise AssertionError("no network without key or for oversize audio")
+
+    monkeypatch.setattr(bot, "_get_http_session", must_not_be_called)
+    monkeypatch.setattr(bot, "GROQ_API_KEY", "")
+    assert asyncio.run(bot._transcribe_audio(b"ogg-bytes", 123)) is None
+    monkeypatch.setattr(bot, "GROQ_API_KEY", "fake-key")
+    monkeypatch.setattr(bot, "VOICE_TRANSCRIBE_MAX_BYTES", 4)
+    assert asyncio.run(bot._transcribe_audio(b"too-big-bytes", 123)) is None
+
+
+def test_transcribe_audio_http_error_is_none(monkeypatch):
+    monkeypatch.setattr(bot, "_get_http_session", _fake_groq_audio_session({}, status=503))
+    monkeypatch.setattr(bot, "GROQ_API_KEY", "fake-key")
+    assert asyncio.run(bot._transcribe_audio(b"ogg-bytes", 123)) is None
+

tests/test_bot_state.py

@@ -561,6 +561,59 @@ def test_media_question_without_file_gets_no_file_notice(rate_guard_setup, monke
     assert "не выдумывай" in prompt.lower()
 
 
+def test_voice_message_transcribed_into_normal_routing(rate_guard_setup, monkeypatch):
+    # Войс: транскрибация уходит в общий роутинг как текст, аудио дальше не едет.
+    message = rate_guard_setup()
+    message.text = ""
+    captured = {}
+
+    async def fake_resolve(message, state, clean_prompt, *, is_private):
+        return None, "", "", (b"ogg-bytes", "audio/ogg")
+
+    async def fake_transcribe(audio_bytes, chat_id):
+        assert audio_bytes == b"ogg-bytes"
+        return "текст из войса"
+
+    async def fake_run_route(chat_id, ai_prompt, route, message, **kwargs):
+        captured["prompt"] = ai_prompt
+        captured["media"] = kwargs.get("media")
+        return "ok", False
+
+    monkeypatch.setattr(bot, "_resolve_incoming_media", fake_resolve)
+    monkeypatch.setattr(bot, "_transcribe_audio", fake_transcribe)
+    monkeypatch.setattr(bot, "_run_route", fake_run_route)
+    monkeypatch.setattr(bot, "_safe_reply", AsyncMock())
+    asyncio.run(bot._handle_message_core(message))
+    assert "текст из войса" in captured["prompt"]
+    assert captured["media"] is None
+    bot.chat_state.pop(123, None)
+
+
+def test_voice_message_falls_back_to_gemini_audio_on_transcribe_failure(rate_guard_setup, monkeypatch):
+    # Whisper упал — войс идёт прежним путём (аудио в Gemini), а не в пустоту.
+    message = rate_guard_setup()
+    message.text = ""
+    captured = {}
+
+    async def fake_resolve(message, state, clean_prompt, *, is_private):
+        return None, "", "", (b"ogg-bytes", "audio/ogg")
+
+    async def fake_transcribe(audio_bytes, chat_id):
+        return None
+
+    async def fake_run_route(chat_id, ai_prompt, route, message, **kwargs):
+        captured["media"] = kwargs.get("media")
+        return "ok", False
+
+    monkeypatch.setattr(bot, "_resolve_incoming_media", fake_resolve)
+    monkeypatch.setattr(bot, "_transcribe_audio", fake_transcribe)
+    monkeypatch.setattr(bot, "_run_route", fake_run_route)
+    monkeypatch.setattr(bot, "_safe_reply", AsyncMock())
+    asyncio.run(bot._handle_message_core(message))
+    assert captured["media"] == [(b"ogg-bytes", "audio/ogg")]
+    bot.chat_state.pop(123, None)
+
+
 def test_ordinary_question_gets_no_file_notice(rate_guard_setup, monkeypatch):
     message = rate_guard_setup()
     message.text = "столица Венгрии?"