feat(backend,frontend,mcp): any voice or round video can be transcribed on request, media is found by its bytes

backend/src/utils/read/media.py

@@ -3,7 +3,7 @@ import json
 import asyncpg
 
 from utils.files import media_file_name
-from utils.read.models import MediaVersionView, MediaView
+from utils.read.models import MediaPlaceView, MediaVersionView, MediaView
 
 MEDIA_COLS = (
     "m.id, m.account_id, m.chat_id, m.message_id, m.kind, m.storage_key, "
@@ -23,6 +23,12 @@ LEFT JOIN LATERAL (
 
 _VERSION_COLS = "id, kind, storage_key, file_size, mime, observed_at"
 
+_PLACES = """
+SELECT account_id, chat_id, message_id, kind, extracted_text FROM media
+WHERE storage_key = $1 AND ($2::bigint IS NULL OR account_id = $2)
+ORDER BY created_at DESC LIMIT 100
+"""
+
 _WEB_PAGE_MEDIA_KINDS = ("photo", "video", "animation", "document", "audio")
 
 
@@ -100,6 +106,13 @@ async def get_message_media(
     return await _web_page_media_stub(pool, account_id, chat_id, message_id)
 
 
+async def find_by_content(
+    pool: asyncpg.Pool, storage_key: str, account_id: int | None = None
+) -> list[MediaPlaceView]:
+    rows = await pool.fetch(_PLACES, storage_key, account_id)
+    return [MediaPlaceView(**dict(row)) for row in rows]
+
+
 async def get_media_versions(
     pool: asyncpg.Pool, account_id: int, chat_id: int, message_id: int
 ) -> list[MediaVersionView]: