feat(multimodal): 实现本地模型管线与请求用量配置

backend/app/agent/runtime.py

@@ -562,6 +562,7 @@ class AgentRuntime:
     @staticmethod
     def _request_metadata(record: RunRecord) -> dict[str, object]:
         metadata = dict(record.request.metadata)
+        metadata["run_id"] = record.run.run_id
         if record.skill_config is not None:
             metadata["skill_id"] = record.skill_config.skill_id
             metadata["retrieval"] = record.skill_config.retrieval.model_dump(mode="json")

backend/app/benchmarks/service.py

@@ -116,7 +116,12 @@ async def _validate_index_compatibility(request: RAGRunRequest) -> None:
     reasons: list[str] = []
     if stats["blocks"] == 0:
         reasons.append("index is empty (no indexed blocks; run /api/index/rebuild first)")
-    if needs_vector:
+    from app.local_models.runtime import LocalEmbedding
+    if needs_vector and isinstance(engine.embedding, LocalEmbedding):
+        from app.retrieval import routed_vectors
+        if await routed_vectors.search_remote("索引可用性检查", top_k=1, accept_local=True) is None:
+            reasons.append("current semantic model space has no complete index")
+    elif needs_vector:
         if meta.get("embedding_model") != engine.embedding.model_id:
             reasons.append(
                 f"embedding model mismatch: index={meta.get('embedding_model')!r}, "

backend/app/container.py

@@ -88,7 +88,7 @@ def build_container() -> ApplicationContainer:
     return ApplicationContainer(
         providers=providers,
         provider_factory=provider_factory,
-        model_routing=ModelRoutingService(providers, provider_factory.credentials),
+        model_routing=_local_model_routing(providers, provider_factory.credentials),
         credentials=credentials,
         tools=tools,
         permissions=permissions,
@@ -99,4 +99,9 @@ def build_container() -> ApplicationContainer:
     )
 
 
+def _local_model_routing(providers, credentials):
+    from app.local_models.runtime import LocalEmbedding, LocalSpeech
+    return ModelRoutingService(providers, credentials, local_embedding=LocalEmbedding(), local_speech=LocalSpeech())
+
+
 container = build_container()

backend/app/contracts.py

@@ -2,7 +2,8 @@ from datetime import datetime
 from enum import Enum
 from typing import Annotated, Any, Literal
 
-from pydantic import BaseModel, ConfigDict, Field, SecretStr, field_validator
+from pydantic import BaseModel, ConfigDict, Field, SecretStr, field_validator, model_validator
+from app.request_overrides import RequestOverride
 
 
 class Contract(BaseModel):
@@ -783,6 +784,8 @@ class ProviderConnectionFields(Contract):
 
 
 class ProviderConfig(ProviderConnectionFields):
+    version: int = Field(default=1, ge=1)
+    request_overrides: list[RequestOverride] = Field(default_factory=list, max_length=32)
     provider_id: str
     provider_type: ProviderType
     name: str
@@ -794,6 +797,7 @@ class ProviderConfig(ProviderConnectionFields):
 
 
 class ProviderCreateRequest(ProviderConnectionFields):
+    request_overrides: list[RequestOverride] = Field(default_factory=list, max_length=32)
     provider_type: ProviderType
     name: str
     base_url: str | None = None
@@ -803,6 +807,8 @@ class ProviderCreateRequest(ProviderConnectionFields):
 
 
 class ProviderUpdateRequest(ProviderConnectionFields):
+    version: int | None = Field(default=None, ge=1)
+    request_overrides: list[RequestOverride] | None = Field(default=None, max_length=32)
     provider_type: ProviderType | None = None
     name: str | None = None
     base_url: str | None = None
@@ -890,6 +896,7 @@ class EmbeddingResult(Contract):
 class SpeakerMatchRequest(Contract):
     attachment_id: str
     reference_attachment_id: str
+    local_only: bool = False
 
 
 class SpeakerMatchResult(Contract):
@@ -978,18 +985,74 @@ class TranscriptionRequest(Contract):
     attachment_id: str
     language: str | None = None
     diarization: bool = False
+    local_only: bool = False
+    word_timestamps: bool = False
+    idempotency_key: str | None = Field(default=None, min_length=1, max_length=128)
+    terminology: dict[str, str] = Field(default_factory=dict, max_length=200)
+
+    @field_validator("terminology")
+    @classmethod
+    def bound_terminology(cls, value):
+        if any(not key or len(key) > 200 or len(replacement) > 200 for key, replacement in value.items()):
+            raise ValueError("术语不能为空，每个术语与替换文本最多 200 字符")
+        return value
+
+
+class TranscriptSegment(Contract):
+    segment_id: str
+    start_time: float = Field(ge=0)
+    end_time: float = Field(ge=0)
+    text: str
+    speaker: str | None = None
+    language: str | None = None
+
+    @model_validator(mode="after")
+    def valid_interval(self):
+        import math
+        if not math.isfinite(self.start_time) or not math.isfinite(self.end_time) or self.end_time < self.start_time:
+            raise ValueError("invalid segment time range")
+        return self
 
 
 class TranscriptionJob(Contract):
     job_id: str
     attachment_id: str
-    status: Literal["queued", "processing", "completed", "failed"]
+    status: Literal["queued", "processing", "running", "completed", "failed", "cancelled"]
     text: str | None = None
     error_code: str | None = None
     error_message: str | None = None
     created_at: datetime
     source: Literal["api", "local", "sidecar"] | None = None
     fallback_reason: str | None = None
+    segments: list[TranscriptSegment] = Field(default_factory=list)
+    original_text: str | None = None
+    original_segments: list[TranscriptSegment] = Field(default_factory=list)
+    speaker_names: dict[str, str] = Field(default_factory=dict)
+    warnings: list[str] = Field(default_factory=list)
+    progress: float | None = Field(default=None, ge=0, le=1)
+    revision: int = 1
+    started_at: datetime | None = None
+    updated_at: datetime | None = None
+    completed_at: datetime | None = None
+    language: str | None = None
+    local_only: bool = False
+    previous_job_id: str | None = None
+    model_snapshot: dict[str, Any] = Field(default_factory=dict)
+    corrections: list[dict[str, str]] = Field(default_factory=list)
+
+
+class TranscriptEditRequest(Contract):
+    revision: int = Field(ge=1)
+    text: str = Field(max_length=1_000_000)
+    segments: list[TranscriptSegment] = Field(default_factory=list, max_length=10000)
+    speaker_names: dict[str, str] = Field(default_factory=dict, max_length=200)
+
+
+class TranscriptNoteRequest(Contract):
+    title: str = Field(min_length=1, max_length=200)
+    folder: str | None = None
+    include_timestamps: bool = True
+    include_speakers: bool = True
 
 
 class IndexStatus(Contract):

backend/app/database/migrations.py

@@ -96,6 +96,30 @@ MIGRATIONS: list[str] = [
     CREATE INDEX IF NOT EXISTS idx_agent_events_type
         ON agent_events(run_id, event, sequence);
     """,
+    # v4: durable media jobs, replayable events and revisions.
+    """
+    CREATE TABLE media_jobs (
+        job_id TEXT PRIMARY KEY, status TEXT NOT NULL, job_json TEXT NOT NULL,
+        request_json TEXT NOT NULL, created_at TEXT NOT NULL, updated_at TEXT NOT NULL,
+        idempotency_key TEXT UNIQUE, fingerprint TEXT NOT NULL
+    );
+    CREATE INDEX media_jobs_created ON media_jobs(created_at DESC);
+    CREATE TABLE media_events (
+        job_id TEXT NOT NULL REFERENCES media_jobs(job_id) ON DELETE CASCADE,
+        sequence INTEGER NOT NULL, event TEXT NOT NULL, data_json TEXT NOT NULL,
+        timestamp TEXT NOT NULL, PRIMARY KEY(job_id, sequence)
+    );
+    CREATE TABLE media_revisions (
+        job_id TEXT NOT NULL REFERENCES media_jobs(job_id) ON DELETE CASCADE,
+        revision INTEGER NOT NULL, job_json TEXT NOT NULL,
+        PRIMARY KEY(job_id, revision)
+    );
+    CREATE TABLE media_notes (
+        job_id TEXT NOT NULL REFERENCES media_jobs(job_id), revision INTEGER NOT NULL,
+        options_hash TEXT NOT NULL, note_id TEXT NOT NULL REFERENCES notes(note_id) ON DELETE CASCADE,
+        PRIMARY KEY(job_id, revision, options_hash)
+    );
+    """,
 ]
 
 

backend/app/local_model_routes.py

@@ -0,0 +1,38 @@
+from fastapi import APIRouter
+from app.local_models import manager
+from app.local_models.runtime import RuntimeConfig, configuration, configure, interpreter, runtime
+
+router = APIRouter(prefix="/api/local-models", tags=["Local models"])
+
+
+@router.get("")
+async def list_models():
+    return {**manager.describe(), "runtime_installed": interpreter().is_file(), "config": configuration(),
+            "active_models": list(runtime.active.values()), "queued_requests": len(runtime.waiters),
+            "last_inference": runtime.diagnostics[-1] if runtime.diagnostics else None}
+
+
+@router.put("/config")
+async def update_config(request: RuntimeConfig):
+    return configure(request)
+
+
+@router.post("/{key}/download", status_code=202)
+async def download(key: str):
+    return await manager.download(key)
+
+
+@router.post("/{key}/cancel")
+async def cancel(key: str):
+    return await manager.cancel_download(key)
+
+
+@router.delete("/{key}")
+async def delete(key: str):
+    return await manager.delete(key)
+
+
+@router.get("/diagnostics")
+async def diagnostics():
+    return {"items": runtime.diagnostics, "config": configuration(), "scope": "current_process",
+            "contains": "model_revision_device_timing_resources_only"}

backend/app/local_models/__init__.py

@@ -0,0 +1 @@
+"""Optional local inference; importing this package does not load model libraries."""

backend/app/local_models/catalog.py

@@ -0,0 +1,31 @@
+"""Reviewed model identities. Runtime never resolves a moving model revision."""
+from dataclasses import asdict, dataclass
+
+
+@dataclass(frozen=True)
+class ModelSpec:
+    key: str
+    name: str
+    capability: str
+    repository: str
+    revision: str
+    license: str
+    source: str = "huggingface"
+    dimensions: int | None = None
+
+    def public(self):
+        return asdict(self)
+
+
+CATALOG = {
+    spec.key: spec for spec in [
+        ModelSpec("bekko", "Bekko Embedding v1 A8M", "embedding", "hotchpotch/bekko-embedding-v1-a8m",
+                  "c721113d59a1d91b447450324f51c4b3332c924a", "MIT", dimensions=384),
+        ModelSpec("granite", "Granite Embedding 97M Multilingual r2", "embedding", "ibm-granite/granite-embedding-97m-multilingual-r2",
+                  "835ad14087e140460703cf0fae09f97d469d65c2", "Apache-2.0", dimensions=384),
+        ModelSpec("qwen3-asr", "Qwen3 ASR 0.6B", "transcription", "Qwen/Qwen3-ASR-0.6B",
+                  "5eb144179a02acc5e5ba31e748d22b0cf3e303b0", "Apache-2.0"),
+        ModelSpec("eres2netv2", "ERes2NetV2 中文声纹", "speaker_matching", "iic/speech_eres2netv2_sv_zh-cn_16k-common",
+                  "3317286545c587ae682dbc166831d9448780eebb", "Apache-2.0", source="modelscope", dimensions=192),
+    ]
+}

backend/app/local_models/manager.py

@@ -0,0 +1,178 @@
+"""Explicit resumable downloads; inference itself never fetches weights."""
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import json
+import shutil
+from pathlib import Path
+from urllib.parse import quote
+
+import httpx
+
+from app.config import get_settings
+from app.errors import ApiError
+from app.local_models.catalog import CATALOG
+
+_downloads: dict[tuple[str, str], asyncio.Task] = {}
+
+
+def model_path(key: str) -> Path:
+    if key not in CATALOG:
+        raise ApiError(404, "MODEL_NOT_FOUND", "Unknown local model.")
+    return get_settings().data_dir / "models" / key / CATALOG[key].revision
+
+
+def state_path(key):
+    return model_path(key) / "install-state.json"
+
+
+def read_state(key):
+    try:
+        state = json.loads(state_path(key).read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        state = {"status": "not_installed", "downloaded_bytes": 0, "total_bytes": None}
+    if state["status"] == "downloading" and task_key(key) not in _downloads:
+        state.update(status="interrupted", error_code="DOWNLOAD_INTERRUPTED")
+    return state
+
+
+def write_state(key, state):
+    path = state_path(key)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    temporary = path.with_suffix(".tmp")
+    temporary.write_text(json.dumps(state), encoding="utf-8")
+    temporary.replace(path)
+
+
+def task_key(key):
+    return str(model_path(key)), key
+
+
+def describe():
+    return {"items": [{**spec.public(), **read_state(key)} for key, spec in CATALOG.items()]}
+
+
+async def download(key):
+    model_path(key)
+    if task_key(key) not in _downloads and read_state(key)["status"] != "installed":
+        write_state(key, {"status": "downloading", "downloaded_bytes": 0, "total_bytes": None})
+        task = asyncio.create_task(_download(key))
+        _downloads[task_key(key)] = task
+        task.add_done_callback(lambda done: _downloads.pop(task_key(key), None))
+    return read_state(key)
+
+
+async def cancel_download(key):
+    task = _downloads.get(task_key(key))
+    if task:
+        task.cancel()
+        await asyncio.gather(task, return_exceptions=True)
+    state = read_state(key)
+    if state["status"] == "downloading":
+        state["status"] = "interrupted"
+        write_state(key, state)
+    return state
+
+
+async def delete(key):
+    from app.local_models.runtime import runtime
+    if runtime.in_use(key):
+        raise ApiError(409, "MODEL_IN_USE", "Model is serving an active request.")
+    await cancel_download(key)
+    path = model_path(key).resolve()
+    root = (get_settings().data_dir / "models").resolve()
+    if not path.is_relative_to(root) or path == root:
+        raise ApiError(400, "INVALID_MODEL_PATH", "Model path escapes storage.")
+    if path.exists():
+        shutil.rmtree(path)
+    return read_state(key)
+
+
+async def _manifest(client, spec):
+    if spec.source == "huggingface":
+        response = await client.get(f"https://huggingface.co/api/models/{spec.repository}/revision/{spec.revision}?blobs=true")
+        response.raise_for_status()
+        files = []
+        for item in response.json()["siblings"]:
+            name = item["rfilename"]
+            if name.startswith(("onnx/", "openvino/", ".")) or not name.endswith((".json", ".txt", ".safetensors", ".md")):
+                continue
+            lfs = item.get("lfs") or {}
+            files.append({"path": name, "size": item["size"], "hash": lfs.get("sha256") or item["blobId"],
+                          "algorithm": "sha256" if lfs else "git-blob",
+                          "url": f"https://huggingface.co/{spec.repository}/resolve/{spec.revision}/{quote(name)}"})
+        return files
+    response = await client.get(f"https://modelscope.cn/api/v1/models/{spec.repository}/repo/files",
+                                params={"Revision": spec.revision, "Recursive": "true"})
+    response.raise_for_status()
+    return [{"path": f["Path"], "size": f["Size"], "hash": f["Sha256"], "algorithm": "sha256",
+             "url": f"https://modelscope.cn/api/v1/models/{spec.repository}/repo?Revision={spec.revision}&FilePath={quote(f['Path'])}"}
+            for f in response.json()["Data"]["Files"]
+            if f["Path"] in {"configuration.json", "pretrained_eres2netv2.ckpt", "README.md"}]
+
+
+def valid_file(path, entry):
+    if not path.is_file() or path.stat().st_size != entry["size"]:
+        return False
+    digest = hashlib.sha256() if entry["algorithm"] == "sha256" else hashlib.sha1()
+    if entry["algorithm"] == "git-blob":
+        digest.update(f"blob {entry['size']}\0".encode())
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest() == entry["hash"]
+
+
+async def _download(key):
+    spec, root = CATALOG[key], model_path(key).resolve()
+    state = {"status": "downloading", "downloaded_bytes": 0, "total_bytes": None}
+    try:
+        async with httpx.AsyncClient(timeout=60, follow_redirects=True) as client:
+            manifest = await _manifest(client, spec)
+            if not manifest or not any(f["path"].endswith((".safetensors", ".ckpt")) for f in manifest):
+                raise ValueError("Missing weights in model manifest")
+            state["total_bytes"] = sum(f["size"] for f in manifest)
+            root.mkdir(parents=True, exist_ok=True)
+            if shutil.disk_usage(root).free < state["total_bytes"] + 100 * 1024 * 1024:
+                raise ApiError(507, "MODEL_DISK_FULL", "Insufficient free disk space.")
+            complete = 0
+            for entry in manifest:
+                path = (root / entry["path"]).resolve()
+                if not path.is_relative_to(root):
+                    raise ValueError("Invalid model manifest path")
+                path.parent.mkdir(parents=True, exist_ok=True)
+                if await asyncio.to_thread(valid_file, path, entry):
+                    complete += entry["size"]
+                    continue
+                partial = path.with_suffix(path.suffix + ".partial")
+                offset = partial.stat().st_size if partial.exists() else 0
+                if offset >= entry["size"]:
+                    partial.unlink()
+                    offset = 0
+                async with client.stream("GET", entry["url"], headers={"Range": f"bytes={offset}-"} if offset else {}) as response:
+                    response.raise_for_status()
+                    if offset and response.status_code != 206:
+                        offset = 0
+                    if response.status_code == 206 and not response.headers.get("content-range", "").startswith(f"bytes {offset}-"):
+                        raise ValueError("Invalid download range")
+                    with partial.open("ab" if offset else "wb") as stream:
+                        async for chunk in response.aiter_bytes(1024 * 1024):
+                            offset += len(chunk)
+                            if offset > entry["size"]:
+                                raise ValueError("Download exceeds manifest size")
+                            stream.write(chunk)
+                            state["downloaded_bytes"] = complete + offset
+                            write_state(key, state)
+                if not await asyncio.to_thread(valid_file, partial, entry):
+                    partial.unlink(missing_ok=True)
+                    raise ApiError(422, "MODEL_CHECKSUM_FAILED", "Model file checksum did not match.")
+                partial.replace(path)
+                complete += entry["size"]
+            (root / "verified-manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+            state.update(status="installed", downloaded_bytes=complete)
+    except asyncio.CancelledError:
+        state.update(status="interrupted", error_code="DOWNLOAD_CANCELLED")
+    except Exception as exc:
+        state.update(status="failed", error_code=exc.code if isinstance(exc, ApiError) else "MODEL_DOWNLOAD_FAILED")
+    write_state(key, state)

backend/app/local_models/runtime.py

@@ -0,0 +1,198 @@
+"""Bounded, cancellable model subprocesses with CPU as the default device."""
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+from contextlib import closing
+from contextvars import ContextVar
+from pathlib import Path
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+from app.config import BACKEND_DIR
+from app.database.db import connect
+from app.errors import ApiError
+from app.local_models.catalog import CATALOG
+from app.local_models.manager import model_path, read_state
+from app.providers.base import ProviderError
+
+
+class RuntimeConfig(BaseModel):
+    device: Literal["cpu", "cuda"] = "cpu"
+    cpu_threads: int = Field(default=2, ge=1, le=32)
+    memory_limit_mb: int = Field(default=8192, ge=1024, le=131072)
+    gpu_memory_limit_mb: int = Field(default=4096, ge=512, le=65536)
+    timeout_seconds: int = Field(default=1800, ge=30, le=14400)
+    embedding_model: Literal["bekko", "granite"] = "bekko"
+    version: int = Field(default=1, ge=1)
+
+
+runtime_context = ContextVar("runtime_config", default=None)
+runtime_progress = ContextVar("runtime_progress", default=None)
+
+
+def configuration():
+    if runtime_context.get() is not None:
+        return runtime_context.get()
+    with closing(connect()) as conn:
+        conn.execute("CREATE TABLE IF NOT EXISTS local_runtime_config (id INTEGER PRIMARY KEY CHECK(id=1), config_json TEXT NOT NULL)")
+        row = conn.execute("SELECT config_json FROM local_runtime_config WHERE id=1").fetchone()
+    return RuntimeConfig.model_validate_json(row[0]) if row else RuntimeConfig()
+
+
+def configure(request):
+    from app.database.db import transaction
+    configuration()
+    with closing(connect()) as conn, transaction(conn):
+        row = conn.execute("SELECT config_json FROM local_runtime_config WHERE id=1").fetchone()
+        previous = RuntimeConfig.model_validate_json(row[0]) if row else RuntimeConfig()
+        if request.version != previous.version:
+            raise ApiError(409, "VERSION_CONFLICT", "Local runtime settings changed; reload first.")
+        request = request.model_copy(update={"version": request.version + 1})
+        conn.execute("INSERT OR REPLACE INTO local_runtime_config VALUES (1,?)", (request.model_dump_json(),))
+    return request
+
+
+def interpreter():
+    return Path(os.getenv("APP_MODEL_PYTHON", str(BACKEND_DIR / ".venv-models" / ("Scripts/python.exe" if os.name == "nt" else "bin/python"))))
+
+
+class Runtime:
+    def __init__(self):
+        self.active = {}
+        self.active_files = {}
+        self.waiters = []
+        self.counter = 0
+        self.diagnostics = []
+
+    def in_use(self, key):
+        return key in self.active.values()
+
+    def media_in_use(self, path):
+        target = str(Path(path).resolve())
+        return any(target in paths for paths in self.active_files.values())
+
+    async def infer(self, key, operation, payload, *, priority=10):
+        if read_state(key)["status"] != "installed":
+            raise ProviderError("LOCAL_MODEL_NOT_INSTALLED", "请先在模型配置中下载本地模型。")
+        if not interpreter().is_file():
+            raise ProviderError("LOCAL_RUNTIME_NOT_INSTALLED", "请先运行本地模型 CPU/CUDA 安装脚本。")
+        config = configuration()
+        self.counter += 1
+        ticket = (priority, self.counter)
+        self.waiters.append(ticket)
+        process = None
+        attempt = None
+        try:
+            # One resident model at a time prevents overlapping CPU/GPU allocations.
+            while self.active or ticket != min(self.waiters):
+                await asyncio.sleep(0.05)
+            self.waiters.remove(ticket)
+            self.active[ticket] = key
+            self.active_files[ticket] = {str(Path(payload[name]).resolve()) for name in ("source", "reference") if payload.get(name)}
+            # Deletion may have occurred while this request was queued.
+            if read_state(key)["status"] != "installed":
+                raise ProviderError("LOCAL_MODEL_NOT_INSTALLED", "模型文件已被删除。")
+            from app.services.usage_service import UsageAttempt
+            attempt = UsageAttempt("local-models", CATALOG[key].repository, "local", operation, source="local")
+            env = {**os.environ, "HF_HUB_OFFLINE": "1", "TRANSFORMERS_OFFLINE": "1",
+                   "HF_HUB_DISABLE_TELEMETRY": "1", "OMP_NUM_THREADS": str(config.cpu_threads),
+                   "PYTHONIOENCODING": "utf-8"}
+            process = await asyncio.create_subprocess_exec(str(interpreter()), str(Path(__file__).with_name("worker.py")),
+                stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.DEVNULL,
+                env=env, limit=16 * 1024 * 1024, **({"creationflags": 0x08000000} if os.name == "nt" else {}))
+            request = {"key": key, "operation": operation, "model_path": str(model_path(key).resolve()),
+                       "config": config.model_dump(), "payload": payload}
+            async def receive():
+                process.stdin.write(json.dumps(request).encode())
+                await process.stdin.drain()
+                process.stdin.close()
+                final = None
+                while line := await process.stdout.readline():
+                    if len(line) > 16 * 1024 * 1024:
+                        raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "本地模型输出超限。")
+                    message = json.loads(line)
+                    if "progress" in message:
+                        callback = runtime_progress.get()
+                        if callback:
+                            callback(message)
+                    else:
+                        final = message
+                await process.wait()
+                return final
+            try:
+                result = await asyncio.wait_for(receive(), config.timeout_seconds)
+            except TimeoutError as exc:
+                raise ProviderError("LOCAL_MODEL_TIMEOUT", "本地模型处理超时。") from exc
+            if process.returncode != 0:
+                raise ProviderError("LOCAL_MODEL_PROCESS_FAILED", "本地模型进程退出，请检查依赖与资源预算。")
+            if not isinstance(result, dict):
+                raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "本地模型进程未返回有效结果。")
+            if "error_code" in result:
+                raise ProviderError(result["error_code"], result.get("message", "本地推理失败。"))
+            attempt.observe(result)
+            attempt.completed = True
+            self.diagnostics.append({"model": CATALOG[key].repository, "revision": CATALOG[key].revision,
+                                     **result.get("diagnostics", {})})
+            self.diagnostics = self.diagnostics[-100:]
+            return result["result"]
+        finally:
+            if ticket in self.waiters:
+                self.waiters.remove(ticket)
+            if process is not None and process.returncode is None:
+                process.kill()
+                await process.wait()
+            self.active.pop(ticket, None)
+            self.active_files.pop(ticket, None)
+            if attempt:
+                attempt.persist()
+
+
+runtime = Runtime()
+
+
+class LocalEmbedding:
+    dim = 384
+
+    @property
+    def model_id(self):
+        spec = CATALOG[configuration().embedding_model]
+        return f"{spec.repository}@{spec.revision}"
+
+    @property
+    def version(self):
+        return CATALOG[configuration().embedding_model].revision
+
+    @property
+    def available(self):
+        return read_state(configuration().embedding_model)["status"] == "installed" and interpreter().is_file()
+
+    async def embed_documents(self, texts):
+        return await runtime.infer(configuration().embedding_model, "embedding", {"texts": texts}, priority=0)
+
+    async def embed_query(self, query):
+        return (await self.embed_documents([query]))[0]
+
+
+class LocalSpeech:
+    @property
+    def available(self):
+        return self.available_for("transcription")
+
+    def available_for(self, capability):
+        key = "qwen3-asr" if capability == "transcription" else "eres2netv2"
+        return read_state(key)["status"] == "installed" and interpreter().is_file()
+
+    async def transcribe(self, source, language):
+        from app.providers.routing import RoutedTranscript
+        from app.contracts import TranscriptSegment
+        result = await runtime.infer("qwen3-asr", "transcription", {"source": str(source.resolve()), "language": language})
+        return RoutedTranscript(text=result["text"], source="local",
+                                segments=[TranscriptSegment(**s) for s in result["segments"]])
+
+    async def match(self, source, reference):
+        result = await runtime.infer("eres2netv2", "speaker_matching",
+            {"source": str(source.resolve()), "reference": str(reference.resolve())}, priority=0)
+        return result["score"]

backend/app/local_models/worker.py

@@ -0,0 +1,175 @@
+"""One offline inference process. Heavy libraries stay out of the API process."""
+from __future__ import annotations
+
+import contextlib
+import json
+import os
+import sys
+import threading
+import time
+
+
+def decode(path, *, limit_seconds=3600):
+    import av
+    import numpy as np
+    frames = []
+    samples = 0
+    with av.open(path, options={"protocol_whitelist": "file,pipe"}) as container:
+        if not container.streams.audio:
+            raise ValueError("Media has no audio track")
+        resampler = av.AudioResampler(format="fltp", layout="mono", rate=16000)
+        for frame in container.decode(audio=0):
+            for output in resampler.resample(frame):
+                audio = output.to_ndarray().reshape(-1)
+                samples += len(audio)
+                if samples > limit_seconds * 16000:
+                    raise ValueError("Audio exceeds one hour")
+                frames.append(audio)
+        for output in resampler.resample(None):
+            frames.append(output.to_ndarray().reshape(-1))
+    if not frames:
+        raise ValueError("Audio is empty")
+    audio = np.concatenate(frames).astype(np.float32)
+    if not np.isfinite(audio).all() or len(audio) < 1600:
+        raise ValueError("Invalid or too short audio")
+    return audio
+
+
+def speech_regions(audio):
+    """Energy-based segmentation, not word alignment; retain original sample offsets."""
+    import numpy as np
+    window = 480
+    energies = [float(np.sqrt(np.mean(audio[i:i + window] ** 2))) for i in range(0, len(audio), window)]
+    threshold = max(0.002, float(np.percentile(energies, 20)) * 2)
+    active = [i for i, energy in enumerate(energies) if energy >= threshold]
+    if not active:
+        return []
+    regions, start, previous = [], active[0], active[0]
+    for index in active[1:]:
+        if index - previous > 20 or (index - start) * window >= 20 * 16000:
+            regions.append((max(0, start * window - 2400), min(len(audio), (previous + 1) * window + 2400)))
+            start = index
+        previous = index
+    regions.append((max(0, start * window - 2400), min(len(audio), (previous + 1) * window + 2400)))
+    return regions
+
+
+def speaker_model(path, device):
+    import torch
+    from modelscope.models.audio.sv.ERes2NetV2 import ERes2NetV2
+    from pathlib import Path
+    model = ERes2NetV2(baseWidth=26, scale=2, expansion=2, embed_dim=192)
+    weights = torch.load(Path(path) / "pretrained_eres2netv2.ckpt", map_location="cpu", weights_only=True)
+    model.load_state_dict(weights, strict=True)
+    return model.to(device).eval()
+
+
+def voice_embedding(model, audio, device):
+    import torch
+    import torchaudio.compliance.kaldi as kaldi
+    if len(audio) < 16000:
+        raise ValueError("Speaker comparison needs at least one second of audio")
+    features = kaldi.fbank(torch.from_numpy(audio).unsqueeze(0), num_mel_bins=80, sample_frequency=16000)
+    features -= features.mean(dim=0, keepdim=True)
+    with torch.inference_mode():
+        vector = model(features.unsqueeze(0).to(device)).flatten()
+        return torch.nn.functional.normalize(vector, dim=0)
+
+
+def run(request):
+    import torch
+    import psutil
+    config, payload = request["config"], request["payload"]
+    torch.set_num_threads(config["cpu_threads"])
+    requested = config["device"]
+    device = "cuda:0" if requested == "cuda" and torch.cuda.is_available() else "cpu"
+    if device != "cpu":
+        total = torch.cuda.get_device_properties(0).total_memory
+        torch.cuda.set_per_process_memory_fraction(min(1.0, config["gpu_memory_limit_mb"] * 1024 ** 2 / total))
+    process = psutil.Process()
+    peak = [0]
+    stop = threading.Event()
+
+    def monitor():
+        while not stop.wait(0.2):
+            used = process.memory_info().rss
+            peak[0] = max(peak[0], used)
+            if used > config["memory_limit_mb"] * 1024 ** 2:
+                os._exit(75)
+
+    threading.Thread(target=monitor, daemon=True).start()
+    started = time.monotonic()
+    path, operation = request["model_path"], request["operation"]
+    try:
+        usage = {}
+        if operation == "embedding":
+            from sentence_transformers import SentenceTransformer
+            model = SentenceTransformer(path, device=device, local_files_only=True, trust_remote_code=False,
+                                        model_kwargs={"attn_implementation": "sdpa"})
+            loaded = time.monotonic()
+            result = model.encode(payload["texts"], batch_size=4, normalize_embeddings=True, show_progress_bar=False).tolist()
+            # Count the tokenizer's actual encoded input, not characters or words.
+            usage = {"input_tokens": int(model.tokenize(payload["texts"])["attention_mask"].sum())}
+        elif operation == "transcription":
+            from qwen_asr import Qwen3ASRModel
+            model = Qwen3ASRModel.from_pretrained(path, dtype=torch.float32 if device == "cpu" else torch.float16,
+                device_map=device, attn_implementation="sdpa", max_inference_batch_size=1, max_new_tokens=512)
+            loaded = time.monotonic()
+            audio = decode(payload["source"])
+            regions = speech_regions(audio)
+            language = {"zh": "Chinese", "en": "English", "ja": "Japanese", "yue": "Cantonese"}.get(payload.get("language"), payload.get("language"))
+            segments = []
+            for start, end in regions:
+                output = model.transcribe(audio=(audio[start:end], 16000), language=language)[0]
+                if output.text.strip():
+                    segments.append({"segment_id": f"segment_{len(segments) + 1}", "start_time": start / 16000,
+                                     "end_time": end / 16000, "text": output.text, "language": output.language})
+                    sys.__stdout__.write(json.dumps({"progress": end / len(audio), "segment": segments[-1]}, ensure_ascii=False) + "\n")
+                    sys.__stdout__.flush()
+            result = {"text": "\n".join(s["text"] for s in segments), "segments": segments}
+        elif operation == "speaker_matching":
+            model = speaker_model(path, device)
+            loaded = time.monotonic()
+            first = voice_embedding(model, decode(payload["source"]), device)
+            second = voice_embedding(model, decode(payload["reference"]), device)
+            # Similarity, not a calibrated identity probability.
+            result = {"score": max(0.0, min(1.0, float(torch.dot(first, second))))}
+        elif operation == "diarization":
+            model = speaker_model(path, device)
+            loaded = time.monotonic()
+            audio = decode(payload["source"])
+            centroids, speakers = [], []
+            for segment in payload["segments"]:
+                sample = audio[int(segment["start_time"] * 16000):int(segment["end_time"] * 16000)]
+                if len(sample) < 16000:
+                    speakers.append(None)
+                    continue
+                vector = voice_embedding(model, sample, device)
+                similarities = [float(torch.dot(vector, c)) for c in centroids]
+                best = max(range(len(similarities)), key=similarities.__getitem__) if similarities else None
+                if best is None or similarities[best] < 0.36:
+                    best = len(centroids)
+                    centroids.append(vector)
+                speakers.append(f"speaker_{best + 1}")
+            result = {"speakers": speakers}
+        else:
+            raise ValueError("Unknown inference operation")
+        return {"result": result, "usage": usage, "diagnostics": {"requested_device": requested, "actual_device": device,
+                "fallback_reason": "CUDA_UNAVAILABLE" if requested == "cuda" and device == "cpu" else None,
+                "load_seconds": loaded - started, "inference_seconds": time.monotonic() - loaded,
+                "peak_memory_bytes": max(peak[0], process.memory_info().rss), "operation": operation}}
+    finally:
+        stop.set()
+
+
+if __name__ == "__main__":
+    request = json.loads(sys.stdin.buffer.read())
+    # Third-party progress/logging must never corrupt the protocol or leak into API errors.
+    with contextlib.redirect_stdout(sys.stderr):
+        try:
+            response = run(request)
+        except (ImportError, ModuleNotFoundError):
+            response = {"error_code": "LOCAL_RUNTIME_DEPENDENCY_MISSING", "message": "本地模型运行依赖不完整，请重新运行安装脚本。"}
+        except Exception:
+            response = {"error_code": "LOCAL_INFERENCE_FAILED", "message": "本地推理失败，请检查媒体格式、模型和设备配置。"}
+    sys.stdout.buffer.write((json.dumps(response, ensure_ascii=False, allow_nan=False) + "\n").encode("utf-8"))

backend/app/main.py

@@ -9,6 +9,10 @@ from app.config import get_settings
 from app.container import container
 from app.errors import ApiError, api_error_handler, http_error_handler, validation_error_handler
 from app.routes import router as api_router
+from app.media_routes import router as media_router
+from app.local_model_routes import router as local_model_router
+from app.usage_routes import router as usage_router
+from app.provider_preview_routes import router as provider_preview_router
 from app.schemas import HealthResponse, ServiceStatusResponse
 
 settings = get_settings()
@@ -16,10 +20,17 @@ settings = get_settings()
 
 @asynccontextmanager
 async def lifespan(_: FastAPI):
-    yield
-    # 第三方 MCP Server 必须跟随 AI Core 退出，不能遗留孤儿进程。
-    container.plugins.shutdown()
-    container.mcp_servers.shutdown()
+    from app.services import transcription_service
+    transcription_service.recover_interrupted()
+    try:
+        yield
+    finally:
+        await transcription_service.shutdown()
+        from app.local_models import manager
+        for _, key in list(manager._downloads):
+            await manager.cancel_download(key)
+        container.plugins.shutdown()
+        container.mcp_servers.shutdown()
 
 
 app = FastAPI(
@@ -41,6 +52,10 @@ app.add_exception_handler(ApiError, api_error_handler)
 app.add_exception_handler(RequestValidationError, validation_error_handler)
 app.add_exception_handler(StarletteHttpException, http_error_handler)
 app.include_router(api_router)
+app.include_router(media_router)
+app.include_router(local_model_router)
+app.include_router(usage_router)
+app.include_router(provider_preview_router)
 
 
 @app.get("/health", response_model=HealthResponse, tags=["System"])

backend/app/media_routes.py

@@ -0,0 +1,162 @@
+"""Media storage and durable transcription controls."""
+from __future__ import annotations
+
+import asyncio
+import json
+from contextlib import closing
+from pathlib import Path
+from uuid import uuid4
+
+from fastapi import APIRouter, Header, Query, Request
+from fastapi.responses import FileResponse, StreamingResponse
+
+from app.contracts import TranscriptEditRequest, TranscriptNoteRequest, TranscriptionJob
+from app.database.db import connect, transaction
+from app.errors import ApiError
+from app.services import transcription_service as jobs
+from app.services.attachment_service import attachment_path
+
+router = APIRouter(prefix="/api/media", tags=["Media"])
+MAX_UPLOAD_BYTES = 25 * 1024 * 1024
+MEDIA_SUFFIXES = {".wav", ".mp3", ".flac", ".ogg", ".m4a", ".mp4", ".webm", ".txt", ".md"}
+
+
+@router.post("/attachments", status_code=201)
+async def upload_attachment(request: Request, filename: str = Query(min_length=1, max_length=255)):
+    suffix = Path(filename).suffix.lower()
+    if suffix not in MEDIA_SUFFIXES:
+        raise ApiError(422, "UNSUPPORTED_MEDIA", "Unsupported attachment extension.")
+    attachment_id = f"media_{uuid4().hex}{suffix}"
+    destination = attachment_path(attachment_id)
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    temporary = destination.with_suffix(destination.suffix + ".upload")
+    size = 0
+    try:
+        with temporary.open("xb") as stream:
+            async for chunk in request.stream():
+                size += len(chunk)
+                if size > MAX_UPLOAD_BYTES:
+                    raise ApiError(413, "ATTACHMENT_TOO_LARGE", "Attachment exceeds 25 MiB.")
+                stream.write(chunk)
+        if not size:
+            raise ApiError(422, "EMPTY_ATTACHMENT", "Attachment is empty.")
+        temporary.replace(destination)
+    finally:
+        temporary.unlink(missing_ok=True)
+    return {"attachment_id": attachment_id, "filename": Path(filename).name, "size": size}
+
+
+@router.get("/attachments/{attachment_id}")
+async def download_attachment(attachment_id: str):
+    path = attachment_path(attachment_id)
+    if not path.is_file():
+        raise ApiError(404, "ATTACHMENT_NOT_FOUND", "Attachment was not found.")
+    return FileResponse(path, headers={"X-Content-Type-Options": "nosniff"})
+
+
+@router.get("/transcriptions")
+async def list_jobs(status: str | None = None, limit: int = Query(50, ge=1, le=200), offset: int = Query(0, ge=0)):
+    if status is not None and status not in jobs.TERMINAL | {"queued", "running", "processing"}:
+        raise ApiError(422, "INVALID_STATUS", "Unknown transcription status.")
+    return jobs.list_transcriptions(status, limit, offset)
+
+
+@router.post("/transcriptions/{job_id}/cancel", response_model=TranscriptionJob)
+async def cancel_job(job_id: str):
+    return await jobs.cancel(job_id)
+
+
+@router.post("/transcriptions/{job_id}/retry", response_model=TranscriptionJob, status_code=202)
+async def retry_job(job_id: str):
+    return await jobs.retry(job_id)
+
+
+@router.patch("/transcriptions/{job_id}", response_model=TranscriptionJob)
+async def edit_job(job_id: str, request: TranscriptEditRequest):
+    return jobs.edit(job_id, request)
+
+
+@router.get("/transcriptions/{job_id}/revisions")
+async def revisions(job_id: str):
+    current = jobs.require_job(job_id)
+    with closing(connect()) as conn:
+        rows = conn.execute("SELECT job_json FROM media_revisions WHERE job_id=? ORDER BY revision", (job_id,)).fetchall()
+    return {"items": [TranscriptionJob.model_validate_json(row[0]) for row in rows] + [current]}
+
+
+@router.get("/transcriptions/{job_id}/events")
+async def stream_events(job_id: str, request: Request, after: int = Query(-1, ge=-1),
+                        last_event_id: str | None = Header(None)):
+    jobs.require_job(job_id)
+    if last_event_id is not None:
+        try:
+            after = max(after, int(last_event_id))
+        except ValueError as exc:
+            raise ApiError(422, "INVALID_EVENT_CURSOR", "Last-Event-ID must be an integer.") from exc
+
+    async def stream():
+        cursor = after
+        idle = 0
+        while not await request.is_disconnected():
+            batch = jobs.events(job_id, cursor)
+            for event in batch:
+                cursor = event["sequence"]
+                yield f"id: {cursor}\nevent: {event['event']}\ndata: {json.dumps(event, ensure_ascii=False)}\n\n"
+            if len(batch) == 200:
+                continue
+            if jobs.require_job(job_id).status in jobs.TERMINAL:
+                # Re-read once: completion may have been committed after this batch was read.
+                if jobs.events(job_id, cursor):
+                    continue
+                return
+            idle += 1
+            if idle % 30 == 0:
+                yield ": keepalive\n\n"
+            await asyncio.sleep(0.5)
+
+    return StreamingResponse(stream(), media_type="text/event-stream",
+                             headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"})
+
+
+@router.post("/transcriptions/{job_id}/notes", status_code=201)
+async def create_note(job_id: str, request: TranscriptNoteRequest):
+    from app.services.media_notes import create_transcript_note
+    return await create_transcript_note(job_id, request)
+
+
+@router.get("/attachments/{attachment_id}/cleanup-impact")
+async def cleanup_impact(attachment_id: str):
+    attachment_path(attachment_id)
+    with closing(connect()) as conn:
+        records = conn.execute("SELECT job_json FROM media_jobs").fetchall()
+        affected = [TranscriptionJob.model_validate_json(row[0]) for row in records]
+        affected = [job for job in affected if job.attachment_id == attachment_id]
+        note_ids = []
+        for job in affected:
+            note_ids.extend(row[0] for row in conn.execute("SELECT note_id FROM media_notes WHERE job_id=?", (job.job_id,)))
+    return {"job_ids": [job.job_id for job in affected], "retained_note_ids": sorted(set(note_ids)),
+            "message": "清理原附件、转写正文、修订和术语记录；已保存笔记保留，音频链接将失效。"}
+
+
+@router.delete("/attachments/{attachment_id}")
+async def cleanup_attachment(attachment_id: str):
+    from app.local_models.runtime import runtime
+    impact = await cleanup_impact(attachment_id)
+    affected = [jobs.require_job(job_id) for job_id in impact["job_ids"]]
+    if runtime.media_in_use(attachment_path(attachment_id)) or any(job.status not in jobs.TERMINAL for job in affected):
+        raise ApiError(409, "MEDIA_IN_USE", "Wait for media processing to finish before cleanup.")
+    for path in (attachment_path(attachment_id), attachment_path(f"{attachment_id}.txt")):
+        path.unlink(missing_ok=True)
+    with closing(connect()) as conn, transaction(conn):
+        for job in affected:
+            job.text = job.original_text = None
+            job.segments = []; job.original_segments = []; job.speaker_names = {}; job.corrections = []
+            job.model_snapshot = {}
+            job.status = "cancelled"; job.error_code = "MEDIA_PURGED"; job.error_message = "附件与转写内容已清理。"
+            job.updated_at = jobs.now()
+            conn.execute("UPDATE media_jobs SET job_json=?,status=?,request_json='{}' WHERE job_id=?",
+                         (job.model_dump_json(), job.status, job.job_id))
+            conn.execute("DELETE FROM media_revisions WHERE job_id=?", (job.job_id,))
+            conn.execute("DELETE FROM media_events WHERE job_id=?", (job.job_id,))
+            jobs._event(conn, job, "Purged")
+    return impact

backend/app/provider_preview_routes.py

@@ -0,0 +1,43 @@
+from fastapi import APIRouter
+from pydantic import BaseModel
+from app.contracts import ProviderCreateRequest, ProviderConfig, ModelRequest, Message, MessageRole
+from app.providers.factory import ProviderFactory
+from app.request_overrides import apply_overrides
+
+router = APIRouter(prefix="/api/providers", tags=["Providers"])
+
+
+class PreviewRequest(BaseModel):
+    provider: ProviderCreateRequest
+    stream: bool = True
+    capability: str = "chat"
+
+
+@router.post("/request-preview")
+async def preview(request: PreviewRequest):
+    class NoCredentials:
+        def resolve(self, key):
+            return None
+    config = ProviderConfig(provider_id="preview", **request.provider.model_dump())
+    if request.capability != "chat":
+        from app.errors import ApiError
+        if request.capability not in {"embedding", "transcription", "speaker_matching"}:
+            raise ApiError(422, "INVALID_CAPABILITY", "Unknown capability.")
+        payload = {"model": config.default_model or "<模型 ID>"}
+        payload["input" if request.capability == "embedding" else "file"] = "<运行时输入，不包含正文或文件>"
+        if request.capability == "speaker_matching":
+            payload["reference_file"] = "<声纹参考附件>"
+    else:
+        from app.providers.factory import UnsupportedProviderError
+        from app.errors import ApiError
+        try:
+            adapter = ProviderFactory(NoCredentials()).build(config)
+        except UnsupportedProviderError as exc:
+            raise ApiError(422, "PROVIDER_TYPE_UNSUPPORTED", "该协议不支持请求预览。") from exc
+        model_request = ModelRequest(provider_id="preview", model=config.default_model or "<模型 ID>",
+            messages=[Message(role=MessageRole.user, content="<运行时消息，已隐藏>")])
+        build = getattr(adapter, "_payload", None) or adapter._chat_payload
+        payload = build(model_request, stream=request.stream)
+    return {"body": apply_overrides(payload, config.request_overrides, request.capability,
+                                    stream=request.stream if request.capability == "chat" else False),
+            "contains_credentials": False, "execution": "preview_only"}

backend/app/providers/factory.py

@@ -16,6 +16,30 @@ class ProviderFactory:
         self.credentials = ProviderCredentialResolver(credentials)
 
     def build(self, config: ProviderConfig) -> ModelProvider:
+        adapter = self._build(config)
+        adapter.provider_config = config.model_copy(deep=True)
+        from app.services.usage_service import usage_context
+        from contextlib import aclosing
+        from uuid import uuid4
+        complete, stream = adapter.complete, adapter.stream
+        async def complete_with_trace(request):
+            token = usage_context.set({"request_id": uuid4().hex, "run_id": request.metadata.get("run_id")})
+            try:
+                return await complete(request)
+            finally:
+                usage_context.reset(token)
+        async def stream_with_trace(request):
+            token = usage_context.set({"request_id": uuid4().hex, "run_id": request.metadata.get("run_id")})
+            try:
+                async with aclosing(stream(request)) as events:
+                    async for event in events:
+                        yield event
+            finally:
+                usage_context.reset(token)
+        adapter.complete, adapter.stream = complete_with_trace, stream_with_trace
+        return adapter
+
+    def _build(self, config: ProviderConfig) -> ModelProvider:
         if config.provider_type == ProviderType.openai_responses:
             from app.providers.openai_responses import OpenAIResponsesProvider
             return OpenAIResponsesProvider(

backend/app/providers/http_base.py

@@ -253,6 +253,18 @@ class HTTPProviderMixin:
     stream_path = "/chat/completions"
     stream_format = "sse"
 
+    def _custom_payload(self, payload):
+        from app.request_overrides import apply_overrides
+        config = getattr(self, "provider_config", None)
+        return apply_overrides(payload, config.request_overrides, "chat", stream=bool(payload.get("stream"))) if config else payload
+
+    def _usage_attempt(self, payload):
+        from app.services.usage_service import UsageAttempt
+        config = getattr(self, "provider_config", None)
+        protocol = config.provider_type.value if config else "openai_compatible"
+        return UsageAttempt(config.provider_id if config else "unregistered", str(payload.get("model", "")), protocol,
+                            source="local" if protocol == "ollama" else "api")
+
     def _headers(self) -> dict[str, str]:
         return {"Content-Type": "application/json"}
 
@@ -268,11 +280,18 @@ class HTTPProviderMixin:
 
     async def _request(self, method: str, path: str, **kwargs) -> dict:
         headers = self._headers()
+        attempt = None
+        if isinstance(kwargs.get("json"), dict) and path == self.stream_path:
+            kwargs["json"] = self._custom_payload(kwargs["json"])
+            attempt = self._usage_attempt(kwargs["json"])
         try:
             async with httpx.AsyncClient(timeout=self.timeout_seconds, transport=self.transport) as client:
                 response = await client.request(method, f"{self.base_url}{path}", headers=headers, **kwargs)
                 response.raise_for_status()
                 data = object_value(response.json())
+                if attempt:
+                    attempt.observe(data)
+                    attempt.completed = True
                 check_error(data)
                 return data
         except httpx.TimeoutException as exc:
@@ -283,8 +302,13 @@ class HTTPProviderMixin:
             raise ProviderError("PROVIDER_UNAVAILABLE", "Provider is unavailable.") from exc
         except (ValueError, TypeError) as exc:
             raise invalid_response() from exc
+        finally:
+            if attempt:
+                attempt.persist()
 
     async def _stream_json(self, payload: dict[str, object]) -> AsyncIterator[dict]:
+        payload = self._custom_payload(payload)
+        attempt = self._usage_attempt(payload)
         headers = self._headers()
         headers["Accept"] = "text/event-stream" if self.stream_format == "sse" else "application/x-ndjson"
         try:
@@ -295,12 +319,14 @@ class HTTPProviderMixin:
                     if self.stream_format == "sse":
                         async with aclosing(sse_objects(response)) as objects:
                             async for data in objects:
+                                attempt.observe(data)
                                 yield data
                     else:
                         async for line in response.aiter_lines():
                             if line.strip():
                                 data = object_value(json.loads(line))
                                 check_error(data)
+                                attempt.observe(data)
                                 yield data
         except httpx.TimeoutException as exc:
             raise ProviderError("PROVIDER_TIMEOUT", "Provider request timed out.") from exc
@@ -310,3 +336,5 @@ class HTTPProviderMixin:
             raise ProviderError("PROVIDER_UNAVAILABLE", "Provider is unavailable.") from exc
         except (ValueError, TypeError) as exc:
             raise invalid_response() from exc
+        finally:
+            attempt.persist()

backend/app/providers/routing.py

@@ -1,14 +1,14 @@
 """Capability routing: validated remote results, then an explicit local backend.
 
-Phase E supplies HTTP adapters and injectable local contracts. Hash embeddings are
-still a development placeholder; speech models are installed in phase F.
+Production injects installed CPU/CUDA backends. Deterministic embeddings remain
+available only for explicitly injected tests and protocol fixtures.
 """
 from __future__ import annotations
 
 import hashlib
 import json
 import math
-from dataclasses import dataclass
+from dataclasses import dataclass, field, replace
 from pathlib import Path
 from typing import Protocol
 
@@ -55,6 +55,7 @@ class RoutedTranscript:
     text: str
     source: str
     fallback_reason: str | None = None
+    segments: list = field(default_factory=list)
 
 
 def invalid_response() -> ProviderError:
@@ -87,6 +88,19 @@ class ModelRoutingService:
         conn.execute("CREATE TABLE IF NOT EXISTS model_routing (id INTEGER PRIMARY KEY CHECK(id=1), config_json TEXT NOT NULL)")
         return conn
 
+    def snapshot(self):
+        from copy import copy
+        from app.providers.registry import RegisteredProvider
+        frozen = copy(self)
+        config = self.configuration().model_copy(deep=True)
+        providers = ProviderRegistry()
+        for item in self.providers.list_configs():
+            original = self.providers.get_any(item.provider_id)
+            providers._providers[item.provider_id] = RegisteredProvider(item, original.adapter)
+        frozen.providers = providers
+        frozen.configuration = lambda: config
+        return frozen
+
     def configuration(self) -> ModelRoutingConfig:
         conn = self._connection()
         try:
@@ -98,11 +112,16 @@ class ModelRoutingService:
             conn.close()
 
     def describe(self) -> ModelRoutingResponse:
+        is_hash = isinstance(self.local_embedding, HashEmbeddingProvider)
+        embedding_available = getattr(self.local_embedding, "available", True)
+        def speech_available(capability):
+            check = getattr(self.local_speech, "available_for", None)
+            return check(capability) if check else self.local_speech.available
         return ModelRoutingResponse(config=self.configuration(), local_backends=[
-            LocalBackendStatus(capability="embedding", status="placeholder" if isinstance(self.local_embedding, HashEmbeddingProvider) else "ready",
-                               message="当前为 hash-v1 确定性占位向量，真实本地语义模型尚未集成。" if isinstance(self.local_embedding, HashEmbeddingProvider) else "本地 Embedding 模型已就绪。"),
-            *[LocalBackendStatus(capability=capability, status="ready" if self.local_speech.available else "not_installed",
-                                 message="本地模型已就绪。" if self.local_speech.available else "阶段 F 接入本地模型；当前保留回退接口。")
+            LocalBackendStatus(capability="embedding", status="placeholder" if is_hash else ("ready" if embedding_available else "not_installed"),
+                               message="测试占位向量。" if is_hash else ("本地 Embedding 文件和运行环境已安装。" if embedding_available else "请安装本地模型运行环境并下载 Embedding 权重。")),
+            *[LocalBackendStatus(capability=capability, status="ready" if speech_available(capability) else "not_installed",
+                                 message="本地模型文件和运行环境已安装。" if speech_available(capability) else "请安装运行环境并下载对应本地模型。")
               for capability in ("transcription", "speaker_matching")],
         ])
 
@@ -150,8 +169,16 @@ class ModelRoutingService:
         url = (provider.base_url or "https://api.openai.com/v1").rstrip("/") + binding.endpoint
         return url, {"Authorization": f"Bearer {key}"} if key else {}
 
-    async def _request(self, binding: ModelBinding, *, remote: tuple[str, dict[str, str]] | None = None, **kwargs) -> tuple[dict, str]:
+    async def _request(self, binding: ModelBinding, *, remote: tuple[str, dict[str, str]] | None = None, provider_config=None, **kwargs) -> tuple[dict, str]:
         url, headers = remote or self._remote(binding)
+        from app.request_overrides import apply_overrides
+        from app.services.usage_service import UsageAttempt
+        capability = "embedding" if "json" in kwargs else ("speaker_matching" if "reference_file" in kwargs.get("files", {}) else "transcription")
+        provider = provider_config or self.providers.get(binding.provider_id).config
+        field = "json" if capability == "embedding" else "data"
+        payload = apply_overrides(kwargs.get(field, {}), provider.request_overrides, capability)
+        kwargs[field] = payload if field == "json" else {key: json.dumps(value) if isinstance(value, (dict, list, bool)) or value is None else value for key, value in payload.items()}
+        attempt = UsageAttempt(binding.provider_id, binding.model, provider.provider_type.value, capability)
         try:
             async with httpx.AsyncClient(timeout=30, transport=self.transport) as client:
                 async with client.stream("POST", url, headers=headers, **kwargs) as response:
@@ -162,6 +189,8 @@ class ModelRoutingService:
                         if len(body) > MAX_RESPONSE_BYTES:
                             raise invalid_response()
                     data = json.loads(body)
+                    attempt.observe(data)
+                    attempt.completed = True
         except httpx.TimeoutException as exc:
             raise ProviderError("PROVIDER_TIMEOUT", "Model API timed out.") from exc
         except httpx.HTTPStatusError as exc:
@@ -171,6 +200,8 @@ class ModelRoutingService:
             raise ProviderError("PROVIDER_UNAVAILABLE", "Model API is unavailable.") from exc
         except (ValueError, UnicodeError) as exc:
             raise invalid_response() from exc
+        finally:
+            attempt.persist()
         if not isinstance(data, dict) or data.get("error"):
             raise invalid_response()
         return data, url
@@ -187,12 +218,13 @@ class ModelRoutingService:
                 dimension = binding.dimensions
                 # Freeze the origin across batches, even if the user edits the provider.
                 remote = self._remote(binding)
+                provider_config = self.providers.get(binding.provider_id).config.model_copy(deep=True)
                 for start in range(0, len(texts), 32):
                     batch = texts[start:start + 32]
                     payload = {"model": binding.model, "input": batch, "encoding_format": "float"}
                     if binding.dimensions is not None:
                         payload["dimensions"] = binding.dimensions
-                    data, url = await self._request(binding, remote=remote, json=payload)
+                    data, url = await self._request(binding, remote=remote, provider_config=provider_config, json=payload)
                     items = data.get("data")
                     if not isinstance(items, list) or len(items) != len(batch):
                         raise invalid_response()
@@ -213,12 +245,20 @@ class ModelRoutingService:
                             raise invalid_response()
                         indexed[index] = [value / norm for value in vector]
                     vectors.extend(indexed[index] for index in range(len(batch)))
-                identity = json.dumps([url, binding.model, dimension], separators=(",", ":"))
+                identity_parts = [url, binding.model, dimension]
+                extensions = [rule.model_dump() for rule in provider_config.request_overrides
+                              if rule.capability == "embedding" and rule.model in (None, binding.model)]
+                if extensions:
+                    identity_parts.append(extensions)
+                identity = json.dumps(identity_parts, separators=(",", ":"))
                 return EmbeddingResult(vectors=vectors, source="api", dimensions=dimension,
                                        model_id="api-" + hashlib.sha256(identity.encode()).hexdigest())
             except ProviderError as exc:
                 reason = exc.code
-        vectors = await self.local_embedding.embed_documents(texts)
+        try:
+            vectors = await self.local_embedding.embed_documents(texts)
+        except ProviderError as exc:
+            raise ApiError(503, exc.code, exc.message, {"fallback_reason": reason}) from exc
         return EmbeddingResult(vectors=vectors, source="local", model_id=self.local_embedding.model_id,
                                dimensions=self.local_embedding.dim, fallback_reason=reason)
 
@@ -234,8 +274,8 @@ class ModelRoutingService:
             raise ApiError(413, "ATTACHMENT_TOO_LARGE", "Audio attachment must be between 1 byte and 25 MiB.")
         return handle
 
-    async def transcribe(self, source: Path, language: str | None) -> RoutedTranscript:
-        binding = self.configuration().transcription
+    async def transcribe(self, source: Path, language: str | None, *, local_only: bool = False) -> RoutedTranscript:
+        binding = None if local_only else self.configuration().transcription
         if binding is None:
             with self._media_file(source):
                 pass
@@ -251,19 +291,41 @@ class ModelRoutingService:
                 text = data.get("text")
                 if not isinstance(text, str) or not text.strip():
                     raise invalid_response()
-                return RoutedTranscript(text=text, source="api")
+                segments = []
+                raw_segments = data.get("segments", [])
+                if not isinstance(raw_segments, list) or len(raw_segments) > 10000:
+                    raise invalid_response()
+                from app.contracts import TranscriptSegment
+                for index, raw in enumerate(raw_segments):
+                    if not isinstance(raw, dict):
+                        raise invalid_response()
+                    start, end = raw.get("start", raw.get("start_time")), raw.get("end", raw.get("end_time"))
+                    if not finite_number(start) or not finite_number(end) or not isinstance(raw.get("text"), str):
+                        raise invalid_response()
+                    try:
+                        segments.append(TranscriptSegment(segment_id=f"segment_{index + 1}", start_time=start,
+                            end_time=end, text=raw["text"], speaker=raw.get("speaker")))
+                    except ValueError as exc:
+                        raise invalid_response() from exc
+                if segments != sorted(segments, key=lambda segment: segment.start_time):
+                    raise invalid_response()
+                return RoutedTranscript(text=text, source="api", segments=segments)
             except ProviderError as exc:
                 reason = exc.code
         try:
             text = await self.local_speech.transcribe(source, language)
+            if isinstance(text, RoutedTranscript):
+                if not text.text.strip():
+                    raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "Local transcription was empty.")
+                return replace(text, source="local", fallback_reason=reason)
             if not isinstance(text, str) or not text.strip():
                 raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "Local transcription was empty.")
             return RoutedTranscript(text=text, source="local", fallback_reason=reason)
         except ProviderError as exc:
             raise ApiError(503, exc.code, exc.message, {"fallback_reason": reason}) from exc
 
-    async def match_speakers(self, source: Path, reference: Path) -> SpeakerMatchResult:
-        binding = self.configuration().speaker_matching
+    async def match_speakers(self, source: Path, reference: Path, *, local_only: bool = False) -> SpeakerMatchResult:
+        binding = None if local_only else self.configuration().speaker_matching
         if binding is None:
             with self._media_file(source), self._media_file(reference):
                 pass

backend/app/request_overrides.py

@@ -0,0 +1,68 @@
+"""Declarative request-body extensions with explicit host-owned field conflicts."""
+import copy
+import json
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
+
+PROTECTED = {"model", "messages", "input", "system", "instructions", "tools", "tool_choice", "parallel_tool_calls",
+             "functions", "function_call", "file", "audio", "reference_file", "stream", "previous_response_id",
+             "conversation", "background", "store"}
+SECRETS = {"api_key", "apikey", "authorization", "headers", "url", "base_url", "access_token", "secret", "password"}
+
+
+class RequestOverride(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    capability: Literal["chat", "embedding", "transcription", "speaker_matching"] = "chat"
+    model: str | None = Field(default=None, max_length=200)
+    stream: bool | None = None
+    body: dict = Field(default_factory=dict)
+
+    @model_validator(mode="after")
+    def valid_mode(self):
+        if self.capability != "chat" and self.stream is True:
+            raise ValueError("当前 Embedding 与媒体接口不使用流式请求")
+        return self
+
+    @field_validator("body")
+    @classmethod
+    def validate_body(cls, value):
+        if len(json.dumps(value, allow_nan=False).encode()) > 32768:
+            raise ValueError("自定义请求 JSON 不得超过 32 KiB")
+        conflicts = PROTECTED.intersection(value)
+        if conflicts:
+            raise ValueError("运行请求管理字段不可覆盖：" + ", ".join(sorted(conflicts)))
+        def check(item, depth=0):
+            if depth > 12:
+                raise ValueError("JSON 嵌套不得超过 12 层")
+            if isinstance(item, dict):
+                if any(str(k).lower().replace("-", "_") in SECRETS for k in item):
+                    raise ValueError("密钥、Header 和 URL 请使用独立配置，不得放入请求 JSON")
+                for child in item.values():
+                    check(child, depth + 1)
+            elif isinstance(item, list):
+                for child in item:
+                    check(child, depth + 1)
+        check(value)
+        if "stream_options" in value:
+            options = value["stream_options"]
+            if not isinstance(options, dict) or ("include_usage" in options and type(options["include_usage"]) is not bool):
+                raise ValueError("stream_options 必须是对象，include_usage 必须是布尔值")
+        return value
+
+
+def deep_merge(base, extension):
+    result = copy.deepcopy(base)
+    for key, value in extension.items():
+        result[key] = deep_merge(result[key], value) if isinstance(value, dict) and isinstance(result.get(key), dict) else copy.deepcopy(value)
+    return result
+
+
+def apply_overrides(payload, rules, capability, *, stream=False):
+    selected = [rule for rule in rules if rule.capability == capability and rule.model in (None, payload.get("model"))
+                and (rule.stream is None or rule.stream == stream)]
+    # General defaults precede model overrides; explicit stream conditions are most specific.
+    selected.sort(key=lambda rule: (rule.model is not None, rule.stream is not None))
+    for rule in selected:
+        payload = deep_merge(payload, rule.body)
+    return payload

backend/app/retrieval/embedding.py

@@ -1,7 +1,6 @@
 """Embedding 统一接口与轻量实现。
 
-真实默认是本地 BGE-M3 类模型，但第一阶段先跑通链路，这里用确定性的特征哈希向量代替。
-后续接入真实模型时实现同样的 EmbeddingProvider 接口替换即可，上层检索逻辑不变。
+生产环境使用 local_models 的真实模型。特征哈希实现仅供测试显式注入。
 """
 
 from __future__ import annotations

backend/app/retrieval/engine.py

@@ -20,6 +20,7 @@ from app.contracts import (
 )
 from app.repository import BlockHit
 from app.retrieval.embedding import EmbeddingProvider, HashEmbeddingProvider
+from app.local_models.runtime import LocalEmbedding
 from app.retrieval.hybrid import normalize_scores, rrf_fuse
 from app.retrieval.reranker import LexicalReranker, RankedCandidate, RerankerProvider
 from app.retrieval import routed_vectors
@@ -88,8 +89,13 @@ class RetrievalEngine:
                 and self.embedding is self._routed_defaults[0]
                 and self.vector_store is self._routed_defaults[1]
             ):
-                vec_hits = await routed_vectors.search_remote(request.query, top_k=recall)
+                vec_hits = await routed_vectors.search_remote(request.query, top_k=recall, accept_local=isinstance(self.embedding, LocalEmbedding))
             if vec_hits is None:
+                if isinstance(self.embedding, LocalEmbedding):
+                    if request.mode == SearchMode.hybrid:
+                        return self._search_fts(request)
+                    from app.errors import ApiError
+                    raise ApiError(409, "SEMANTIC_INDEX_UNAVAILABLE", "语义索引未就绪。请配置 Embedding 或下载本地模型后重建索引。")
                 query_vec = await self.embedding.embed_query(request.query)
                 vec_hits = await self.vector_store.search(query_vec, top_k=recall)
                 record_embedding(source="local", model_id=self.embedding.model_id,
@@ -287,5 +293,5 @@ def _utc(dt: datetime) -> datetime:
 
 # 默认引擎实例：轻量实现跑通链路，后续可替换真实模型实现
 engine = RetrievalEngine(
-    HashEmbeddingProvider(), LexicalReranker(), SqliteVecStore(), route_embeddings=True,
+    LocalEmbedding(), LexicalReranker(), SqliteVecStore(), route_embeddings=True,
 )

backend/app/retrieval/routed_vectors.py

@@ -42,6 +42,7 @@ class RemoteEmbeddings:
     space_id: str
     dimensions: int
     vectors: list[list[float]]
+    source: str = "api"
 
 
 def get_model_routing() -> EmbeddingRuntime | None:
@@ -67,7 +68,7 @@ def _unit_vector(vector: list[float], dimensions: int) -> list[float]:
     return [value / norm for value in scaled]
 
 
-async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
+async def embed_remote(texts: list[str], *, accept_local=False) -> RemoteEmbeddings | None:
     """Return validated API vectors, or None to use the caller's local baseline.
 
     Do not use the runtime's local result: the caller may have injected its own
@@ -80,7 +81,7 @@ async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
         if runtime is None:
             return None
         result = await runtime.embed(texts)
-        if result.source != "api":
+        if result.source != "api" and not accept_local:
             record_embedding(fallback_reason=result.fallback_reason)
             return None
         if not isinstance(result.model_id, str) or not result.model_id or result.model_id == "hash-v1":
@@ -93,6 +94,7 @@ async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
             space_id=result.model_id,
             dimensions=result.dimensions,
             vectors=[_unit_vector(vector, result.dimensions) for vector in result.vectors],
+            source=result.source,
         )
     except Exception as exc:
         # Avoid logging provider exceptions containing credentials or note text.
@@ -152,13 +154,13 @@ def store_remote(
         logger.warning("Remote vector storage unavailable (%s); local index retained", type(exc).__name__)
 
 
-async def search_remote(query: str, *, top_k: int) -> list[VectorHit] | None:
+async def search_remote(query: str, *, top_k: int, accept_local=False) -> list[VectorHit] | None:
     """None means fallback, including any missing/invalid current-block vector.
 
     Read coverage and vectors together so concurrent note updates cannot produce
     an apparently complete subset. Never fill missing remote hits with local hits.
     """
-    batch = await embed_remote([query])
+    batch = await embed_remote([query], accept_local=accept_local)
     if batch is None:
         return None
     record_embedding(attempted_space={"model_id": batch.space_id, "dimensions": batch.dimensions})
@@ -190,7 +192,7 @@ async def search_remote(query: str, *, top_k: int) -> list[VectorHit] | None:
                         yield VectorHit(id=row["block_id"], score=max(0.0, min(1.0, score)))
 
                 result = heapq.nlargest(top_k, hits(), key=lambda hit: hit.score)
-                record_embedding(source="api", model_id=batch.space_id,
+                record_embedding(source=batch.source, model_id=batch.space_id,
                                  dimensions=batch.dimensions, fallback_reason=None)
                 return result
         finally:

backend/app/routes.py

@@ -915,6 +915,7 @@ async def create_provider(request: ProviderCreateRequest) -> ProviderConfig:
         default_model=request.default_model,
         credential_id=request.credential_id,
         enabled=request.enabled,
+        request_overrides=request.request_overrides,
         capabilities=container.provider_factory.capabilities(request.provider_type),
     )
     try:
@@ -943,8 +944,12 @@ async def update_provider(
             409, "BUILTIN_PROVIDER_IMMUTABLE", "Mock provider cannot be modified."
         )
     fields = request.model_fields_set
+    if request.version is not None and request.version != current.version:
+        raise ApiError(409, "PROVIDER_VERSION_CONFLICT", "提供商配置已变更，请重新加载后保存。")
     if ("provider_type" in fields and request.provider_type is None) or ("name" in fields and request.name is None) or (
         "enabled" in fields and request.enabled is None
+    ) or (
+        "request_overrides" in fields and request.request_overrides is None
     ):
         raise ApiError(
             422,
@@ -952,6 +957,7 @@ async def update_provider(
             "provider_type, name and enabled cannot be null when explicitly provided.",
         )
     updates = {name: getattr(request, name) for name in fields}
+    updates["version"] = current.version + 1
     if "credential_id" in fields:
         validate_public_credential_id(request.credential_id)
     config = ProviderConfig.model_validate(
@@ -1100,6 +1106,7 @@ async def create_embeddings(request: EmbeddingRequest) -> EmbeddingResult:
 async def match_speakers(request: SpeakerMatchRequest) -> SpeakerMatchResult:
     return await container.model_routing.match_speakers(
         attachment_path(request.attachment_id), attachment_path(request.reference_attachment_id),
+        local_only=request.local_only,
     )
 
 
@@ -1111,7 +1118,7 @@ async def match_speakers(request: SpeakerMatchRequest) -> SpeakerMatchResult:
 )
 async def create_transcription(request: TranscriptionRequest) -> TranscriptionJob:
     return await transcription_service.create_transcription(
-        request.attachment_id, request.language, diarization=request.diarization
+        **request.model_dump(), wait=False
     )
 
 

backend/app/services/index_service.py

@@ -97,6 +97,7 @@ async def rebuild(request: IndexRebuildRequest) -> IndexJob:
                 task_note_links = dict(conn.execute(
                     "SELECT task_id, note_id FROM tasks WHERE note_id IS NOT NULL"
                 ).fetchall())
+                media_links = conn.execute("SELECT job_id,revision,options_hash,note_id FROM media_notes").fetchall()
                 repository.clear_all(conn=conn)
                 await vector_store.clear(conn=conn)
                 for parsed, prepared in prepared_notes:
@@ -107,6 +108,9 @@ async def rebuild(request: IndexRebuildRequest) -> IndexJob:
                         "AND EXISTS (SELECT 1 FROM notes WHERE note_id = ?)",
                         (note_id, task_id, note_id),
                     )
+                for link in media_links:
+                    conn.execute("INSERT OR IGNORE INTO media_notes SELECT ?,?,?,? WHERE EXISTS (SELECT 1 FROM notes WHERE note_id=?)",
+                                 (*link, link["note_id"]))
         finally:
             conn.close()
     except BaseException as exc:

backend/app/services/media_notes.py

@@ -0,0 +1,55 @@
+"""Idempotent transcript export without overwriting an edited note."""
+import asyncio
+import hashlib
+from contextlib import closing
+
+from app.config import get_settings
+from app.database.db import connect, transaction
+from app.errors import ApiError
+from app.services import note_service
+from app.services.transcription_service import require_job
+
+_locks = {}
+
+
+async def create_transcript_note(job_id, options):
+    identity = (str(get_settings().db_path), job_id)
+    lock = _locks.setdefault(identity, asyncio.Lock())
+    async with lock:
+        job = require_job(job_id)
+        if job.status != "completed":
+            raise ApiError(409, "TRANSCRIPT_NOT_READY", "Only completed transcripts can become notes.")
+        options_hash = hashlib.sha256(options.model_dump_json().encode()).hexdigest()
+        with closing(connect()) as conn:
+            row = conn.execute("SELECT note_id FROM media_notes WHERE job_id=? AND revision=? AND options_hash=?",
+                               (job_id, job.revision, options_hash)).fetchone()
+        if row:
+            return await note_service.get_note(row[0])
+        marker = f"<!-- transcription:{job_id}:{job.revision}:{options_hash} -->"
+        title = f"{options.title} · {job_id[-8:]}-r{job.revision}-{options_hash[:6]}"
+        lines = [marker, f"# {options.title}", "", f"[源音频](/#/media?job={job_id})", ""]
+        if job.segments:
+            for segment in job.segments:
+                prefix = []
+                if options.include_timestamps:
+                    seconds = segment.start_time
+                    label = f"{int(seconds // 60):02}:{int(seconds % 60):02}"
+                    prefix.append(f"[{label}](/#/media?job={job_id}&time={seconds})")
+                if options.include_speakers and segment.speaker:
+                    prefix.append(job.speaker_names.get(segment.speaker, segment.speaker))
+                lines.append(" ".join([*prefix, segment.text]))
+                lines.append("")
+        else:
+            lines.append(job.text or "")
+        try:
+            note = await note_service.create_note(title=title, markdown="\n".join(lines), folder=options.folder, tags=["转写"])
+        except ApiError as exc:
+            if exc.code != "RESOURCE_CONFLICT" or "note_id" not in exc.details:
+                raise
+            # Recover a crash between successful note creation and linking the job.
+            note = await note_service.get_note(exc.details["note_id"])
+            if note is None or marker not in note.markdown:
+                raise
+        with closing(connect()) as conn, transaction(conn):
+            conn.execute("INSERT OR IGNORE INTO media_notes VALUES (?,?,?,?)", (job_id, job.revision, options_hash, note.note_id))
+        return note

backend/app/services/note_service.py

@@ -17,7 +17,7 @@ from app.contracts import Note, NoteBlock, NoteSummary
 from app.database.db import connect, transaction
 from app.errors import ApiError
 from app.knowledge.parser import ParsedNote, parse_note
-from app.retrieval.embedding import HashEmbeddingProvider
+from app.local_models.runtime import LocalEmbedding
 from app.retrieval import routed_vectors
 from app.retrieval.vectorstore import SqliteVecStore, VectorRecord
 from app.services.coordination import serialized_vault_mutation
@@ -28,8 +28,8 @@ from app.services.vault_paths import (
     safe_note_filename,
 )
 
-# 轻量实现实例（无状态，可直接复用）；接入真实模型后替换为对应 Provider
-embedding = HashEmbeddingProvider()
+# 真实模型接口不在 API 进程加载权重；测试可显式替换该实例。
+embedding = LocalEmbedding()
 vector_store = SqliteVecStore()
 
 
@@ -80,6 +80,10 @@ PreparedIndex = tuple[list[list[float]], routed_vectors.RemoteEmbeddings | None]
 async def prepare_note_index(parsed: ParsedNote) -> PreparedIndex:
     """Compute vectors before opening a write transaction (including API I/O)."""
     texts = [block.content for block in parsed.blocks]
+    if isinstance(embedding, LocalEmbedding):
+        # One routed invocation: API first, validated local fallback. No hash vectors.
+        remote = await routed_vectors.embed_remote(texts, accept_local=True)
+        return [], remote
     vectors = await embedding.embed_documents(texts)
     remote = await routed_vectors.embed_remote(texts)
     return vectors, remote
@@ -127,7 +131,8 @@ async def index_note(
             await vector_store.upsert(records, conn=conn)
             routed_vectors.store_remote(conn, [block.block_id for block in parsed.blocks], remote)
             repository.set_index_meta(
-                {"embedding_model": embedding.model_id, "embedding_dim": str(embedding.dim)},
+                {"embedding_model": remote.space_id if remote and isinstance(embedding, LocalEmbedding) else embedding.model_id,
+                 "embedding_dim": str(remote.dimensions if remote and isinstance(embedding, LocalEmbedding) else embedding.dim)},
                 conn=conn,
             )
     finally:

backend/app/services/transcription_service.py

@@ -1,65 +1,243 @@
-"""转写作业：API 优先，本地模型回退；保留已有 Host 文本入口。"""
-
+"""Persistent media jobs and replayable events; HTTP enqueues, tools await."""
 from __future__ import annotations
-
-from collections import OrderedDict
+import asyncio
+import hashlib
+import json
+from contextlib import closing
 from datetime import datetime, timezone
 from uuid import uuid4
-
-from app.contracts import TranscriptionJob
+from app.config import get_settings
+from app.contracts import TranscriptionJob, TranscriptionRequest, TranscriptEditRequest
+from app.database.db import connect, transaction
 from app.errors import ApiError
 from app.services.attachment_service import attachment_path
 
-_jobs: OrderedDict[str, TranscriptionJob] = OrderedDict()
-MAX_JOBS = 100
+TERMINAL = {"completed", "failed", "cancelled"}
+_tasks: dict[tuple[str, str], asyncio.Task] = {}
 
+def now():
+    return datetime.now(timezone.utc)
 
-async def create_transcription(attachment_id: str, language: str | None = None, *, diarization: bool = False) -> TranscriptionJob:
-    from app.container import container
-
-    source = attachment_path(attachment_id)
-    job = TranscriptionJob(
-        job_id=f"transcription_{uuid4().hex}",
-        attachment_id=attachment_id,
-        status="processing",
-        created_at=datetime.now(timezone.utc),
-    )
-    try:
-        if diarization:
-            # Speaker verification and diarization are different capabilities.
-            raise ApiError(501, "DIARIZATION_NOT_IMPLEMENTED", "说话人分离将在阶段 F 接入，当前不能忽略 diarization 请求。")
-        transcript = source if source.suffix.lower() in {".txt", ".md"} else attachment_path(f"{attachment_id}.txt")
-        # A saved transcript remains an explicit import path, never faked ASR.
-        if transcript.is_file() and (source == transcript or container.model_routing.configuration().transcription is None):
-            with transcript.open("rb") as handle:
-                content = handle.read(1024 * 1024 + 1)
-            if len(content) > 1024 * 1024:
-                raise ApiError(413, "TRANSCRIPT_TOO_LARGE", "Transcript exceeds 1 MiB.")
-            job.text = content.decode("utf-8")
-            if not job.text.strip():
-                raise ApiError(422, "TRANSCRIPT_EMPTY", "Transcript is empty.")
-            job.source = "sidecar"
-        else:
-            result = await container.model_routing.transcribe(source, language)
-            job.text = result.text
-            job.source = result.source
-            job.fallback_reason = result.fallback_reason
-        job.status = "completed"
-    except ApiError as exc:
-        job.status = "failed"
-        job.error_code = exc.code
-        job.error_message = exc.message
-        job.fallback_reason = exc.details.get("fallback_reason")
-    except (OSError, UnicodeError):
-        job.status = "failed"
-        job.error_code = "TRANSCRIPT_UNREADABLE"
-        job.error_message = "Transcript could not be read."
-    _jobs[job.job_id] = job
-    while len(_jobs) > MAX_JOBS:
-        _jobs.popitem(last=False)
-    return job.model_copy(deep=True)
-
+def task_key(job_id):
+    return str(get_settings().db_path), job_id
 
 def get_transcription(job_id: str) -> TranscriptionJob | None:
-    job = _jobs.get(job_id)
-    return job.model_copy(deep=True) if job else None
+    with closing(connect()) as conn:
+        row = conn.execute("SELECT job_json FROM media_jobs WHERE job_id=?", (job_id,)).fetchone()
+    return TranscriptionJob.model_validate_json(row[0]) if row else None
+
+def require_job(job_id):
+    job = get_transcription(job_id)
+    if job is None:
+        raise ApiError(404, "RESOURCE_NOT_FOUND", "Transcription job not found.")
+    return job
+
+def _event(conn, job, event, data=None):
+    sequence = conn.execute("SELECT COALESCE(MAX(sequence),-1)+1 FROM media_events WHERE job_id=?", (job.job_id,)).fetchone()[0]
+    conn.execute("INSERT INTO media_events VALUES (?,?,?,?,?)", (job.job_id, sequence, event,
+        json.dumps(data or {"status": job.status, "progress": job.progress}), now().isoformat()))
+
+def save(job, event):
+    job.updated_at = now()
+    with closing(connect()) as conn, transaction(conn):
+        conn.execute("UPDATE media_jobs SET status=?,job_json=?,updated_at=? WHERE job_id=?",
+            (job.status, job.model_dump_json(), job.updated_at.isoformat(), job.job_id))
+        _event(conn, job, event)
+
+def list_transcriptions(status=None, limit=50, offset=0):
+    where, args = (" WHERE status=?", [status]) if status else ("", [])
+    with closing(connect()) as conn:
+        total = conn.execute("SELECT COUNT(*) FROM media_jobs" + where, args).fetchone()[0]
+        rows = conn.execute("SELECT job_json FROM media_jobs" + where + " ORDER BY created_at DESC LIMIT ? OFFSET ?", [*args, limit, offset]).fetchall()
+    return {"items": [TranscriptionJob.model_validate_json(row[0]) for row in rows], "page": {"total": total, "limit": limit, "offset": offset}}
+
+def events(job_id, after=-1):
+    require_job(job_id)
+    with closing(connect()) as conn:
+        rows = conn.execute("SELECT * FROM media_events WHERE job_id=? AND sequence>? ORDER BY sequence LIMIT 200", (job_id, after)).fetchall()
+    return [{"job_id": job_id, "sequence": r["sequence"], "event": r["event"], "data": json.loads(r["data_json"]), "timestamp": r["timestamp"]} for r in rows]
+
+def recover_interrupted():
+    with closing(connect()) as conn:
+        rows = conn.execute("SELECT job_json FROM media_jobs WHERE status IN ('queued','running','processing')").fetchall()
+    for row in rows:
+        job = TranscriptionJob.model_validate_json(row[0])
+        if task_key(job.job_id) not in _tasks:
+            job.status, job.error_code = "failed", "TRANSCRIPTION_INTERRUPTED"
+            job.error_message = "AI Core stopped before completion. Retry to start a new attempt."
+            job.completed_at = now()
+            save(job, "Failed")
+
+async def shutdown():
+    tasks = [t for k, t in list(_tasks.items()) if k[0] == str(get_settings().db_path)]
+    for task in tasks:
+        task.cancel()
+    await asyncio.gather(*tasks, return_exceptions=True)
+
+async def create_transcription(attachment_id, language=None, *, diarization=False, local_only=False,
+        word_timestamps=False, idempotency_key=None, terminology=None, wait=True, previous_job_id=None):
+    request = TranscriptionRequest(attachment_id=attachment_id, language=language, diarization=diarization,
+        local_only=local_only, word_timestamps=word_timestamps, idempotency_key=idempotency_key, terminology=terminology or {})
+    source = attachment_path(attachment_id)
+    actual = source if source.is_file() else attachment_path(f"{attachment_id}.txt")
+    if not actual.is_file():
+        raise ApiError(404, "ATTACHMENT_NOT_FOUND", "Attachment was not found.")
+    if not 0 < actual.stat().st_size <= 25 * 1024 * 1024:
+        raise ApiError(413, "ATTACHMENT_TOO_LARGE", "Attachment must be between 1 byte and 25 MiB.")
+    digest = await asyncio.to_thread(lambda: hashlib.sha256(actual.read_bytes()).hexdigest())
+    from app.container import container
+    from app.local_models.runtime import configuration
+    from app.local_models.catalog import CATALOG
+    routing = container.model_routing.snapshot()
+    route = routing.configuration()
+    binding = None if local_only else route.transcription
+    snapshot = {"local_runtime": configuration().model_dump(), "models": {k:v.revision for k,v in CATALOG.items()},
+                "transcription": binding.model_dump() if binding else None}
+    if binding:
+        provider = routing.providers.get_any(binding.provider_id).config
+        snapshot["provider"] = provider.model_dump(exclude={"credential_id"})
+    fingerprint = hashlib.sha256((digest + request.model_dump_json(exclude={"idempotency_key"}) + json.dumps(snapshot, sort_keys=True)).encode()).hexdigest()
+    job = TranscriptionJob(job_id=f"transcription_{uuid4().hex}", attachment_id=attachment_id, status="queued",
+        created_at=now(), updated_at=now(), language=language, local_only=local_only, previous_job_id=previous_job_id, model_snapshot=snapshot)
+    existing = None
+    with closing(connect()) as conn, transaction(conn):
+        if idempotency_key:
+            existing = conn.execute("SELECT job_json,fingerprint FROM media_jobs WHERE idempotency_key=?", (idempotency_key,)).fetchone()
+        if existing:
+            if existing["fingerprint"] != fingerprint:
+                raise ApiError(409, "IDEMPOTENCY_CONFLICT", "This key was used for different input.")
+            job = TranscriptionJob.model_validate_json(existing["job_json"])
+        else:
+            conn.execute("INSERT INTO media_jobs VALUES (?,?,?,?,?,?,?,?)", (job.job_id, job.status,
+                job.model_dump_json(), request.model_dump_json(), job.created_at.isoformat(), job.updated_at.isoformat(), idempotency_key, fingerprint))
+            _event(conn, job, "Queued")
+    key = task_key(job.job_id)
+    if not existing:
+        task = asyncio.create_task(_execute(job.job_id, request, routing))
+        _tasks[key] = task
+        task.add_done_callback(lambda finished: _tasks.pop(key, None))
+    if wait and key in _tasks:
+        try:
+            await _tasks[key]
+        except asyncio.CancelledError:
+            await cancel(job.job_id)
+            raise
+        return require_job(job.job_id)
+    return job
+
+async def _execute(job_id, request, routing=None):
+    from app.container import container
+    job = require_job(job_id)
+    if job.status in TERMINAL:
+        return
+    from app.local_models.runtime import runtime_context, runtime_progress, RuntimeConfig
+    from app.contracts import TranscriptSegment
+    token = runtime_context.set(RuntimeConfig.model_validate(job.model_snapshot.get("local_runtime", {})))
+    def progress(message):
+        job.progress = max(0.0, min(0.99, message["progress"]))
+        job.segments.append(TranscriptSegment.model_validate(message["segment"]))
+        save(job, "SegmentReady")
+    progress_token = runtime_progress.set(progress)
+    job.status, job.started_at = "running", now()
+    save(job, "TranscriptionStarted")
+    cancelled = False
+    try:
+        source = attachment_path(job.attachment_id)
+        transcript = source if source.suffix.lower() in {".txt", ".md"} else attachment_path(f"{job.attachment_id}.txt")
+        if transcript.is_file() and (source == transcript or not source.exists()):
+            def read_transcript():
+                with transcript.open("rb") as stream:
+                    return stream.read(1024 * 1024 + 1)
+            content = await asyncio.to_thread(read_transcript)
+            if len(content) > 1024 * 1024:
+                raise ApiError(413, "TRANSCRIPT_TOO_LARGE", "Transcript exceeds 1 MiB.")
+            job.text, job.source = content.decode("utf-8"), "sidecar"
+        else:
+            result = await (routing or container.model_routing).transcribe(source, request.language, local_only=request.local_only)
+            job.text, job.source, job.fallback_reason = result.text, result.source, result.fallback_reason
+            job.segments = getattr(result, "segments", []) or []
+        if not job.text or not job.text.strip():
+            raise ApiError(422, "TRANSCRIPT_EMPTY", "Transcript is empty.")
+        if request.diarization:
+            if job.segments:
+                from app.local_models.runtime import runtime
+                from app.providers.base import ProviderError
+                try:
+                    result = await runtime.infer("eres2netv2", "diarization", {"source": str(source.resolve()),
+                        "segments": [s.model_dump() for s in job.segments]})
+                    for segment, speaker in zip(job.segments, result["speakers"], strict=True):
+                        segment.speaker = speaker
+                    job.warnings.append("DIARIZATION_SEGMENT_LEVEL")
+                except ProviderError:
+                    job.warnings.append("DIARIZATION_UNAVAILABLE")
+            else:
+                job.warnings.append("DIARIZATION_UNAVAILABLE")
+        if request.word_timestamps:
+            job.warnings.append("WORD_TIMESTAMPS_UNAVAILABLE")
+        job.original_text, job.original_segments = job.text, [s.model_copy(deep=True) for s in job.segments]
+        for original, replacement in request.terminology.items():
+            if original and original != replacement and original in job.text:
+                job.text = job.text.replace(original, replacement)
+                for segment in job.segments:
+                    segment.text = segment.text.replace(original, replacement)
+                job.corrections.append({"original": original, "replacement": replacement, "source": "terminology_postprocessing"})
+        job.status, job.progress = "completed", 1
+    except asyncio.CancelledError:
+        cancelled = True
+        job.status, job.error_code = "cancelled", "TRANSCRIPTION_CANCELLED"
+    except ApiError as exc:
+        job.status, job.error_code, job.error_message = "failed", exc.code, exc.message
+        job.fallback_reason = exc.details.get("fallback_reason")
+    except Exception:
+        job.status, job.error_code, job.error_message = "failed", "TRANSCRIPTION_FAILED", "Transcription could not be completed."
+    job.completed_at = now()
+    save(job, {"completed": "Completed", "cancelled": "Cancelled", "failed": "Failed"}[job.status])
+    runtime_context.reset(token)
+    runtime_progress.reset(progress_token)
+    if cancelled:
+        raise asyncio.CancelledError
+
+async def cancel(job_id):
+    job = require_job(job_id)
+    if job.status in TERMINAL:
+        return job
+    task = _tasks.get(task_key(job_id))
+    if task:
+        task.cancel()
+        await asyncio.gather(task, return_exceptions=True)
+    job = require_job(job_id)
+    if job.status not in TERMINAL:
+        job.status, job.error_code, job.completed_at = "cancelled", "TRANSCRIPTION_CANCELLED", now()
+        save(job, "Cancelled")
+    return job
+
+async def retry(job_id):
+    if require_job(job_id).error_code == "MEDIA_PURGED":
+        raise ApiError(409, "MEDIA_PURGED", "Purged jobs cannot be retried.")
+    if require_job(job_id).status not in {"failed", "cancelled"}:
+        raise ApiError(409, "TRANSCRIPTION_NOT_RETRYABLE", "Only failed or cancelled jobs can be retried.")
+    with closing(connect()) as conn:
+        raw = conn.execute("SELECT request_json FROM media_jobs WHERE job_id=?", (job_id,)).fetchone()[0]
+    request = TranscriptionRequest.model_validate_json(raw)
+    return await create_transcription(**request.model_dump(exclude={"idempotency_key"}), wait=False, previous_job_id=job_id)
+
+def edit(job_id, request: TranscriptEditRequest):
+    with closing(connect()) as conn, transaction(conn):
+        row = conn.execute("SELECT job_json FROM media_jobs WHERE job_id=?", (job_id,)).fetchone()
+        if not row:
+            raise ApiError(404, "RESOURCE_NOT_FOUND", "Transcription job not found.")
+        job = TranscriptionJob.model_validate_json(row[0])
+        if job.status != "completed":
+            raise ApiError(409, "TRANSCRIPT_NOT_READY", "Only completed transcripts can be edited.")
+        if job.revision != request.revision:
+            raise ApiError(409, "VERSION_CONFLICT", "Transcript has changed; reload before saving.")
+        ids = [s.segment_id for s in request.segments]
+        if len(ids) != len(set(ids)) or request.segments != sorted(request.segments, key=lambda s: s.start_time):
+            raise ApiError(422, "INVALID_SEGMENTS", "Segments must have unique IDs and ordered timestamps.")
+        conn.execute("INSERT INTO media_revisions VALUES (?,?,?)", (job_id, job.revision, job.model_dump_json()))
+        job.text, job.segments, job.speaker_names = request.text, request.segments, request.speaker_names
+        job.revision += 1
+        job.updated_at = now()
+        conn.execute("UPDATE media_jobs SET job_json=?,updated_at=? WHERE job_id=?", (job.model_dump_json(), job.updated_at.isoformat(), job_id))
+        _event(conn, job, "Revised", {"revision": job.revision})
+    return job

backend/app/services/usage_service.py

@@ -0,0 +1,132 @@
+"""Application-observed usage per actual HTTP attempt; never an account bill."""
+from __future__ import annotations
+
+import json
+import logging
+from contextlib import closing
+from contextvars import ContextVar
+from datetime import datetime, timezone
+from uuid import uuid4
+
+from app.database.db import connect
+
+METRICS = ("input_tokens", "output_tokens", "total_tokens", "cache_hit_tokens", "cache_miss_tokens", "cache_write_tokens", "reasoning_tokens")
+logger = logging.getLogger(__name__)
+usage_context = ContextVar("usage_context", default=None)
+
+
+def connection():
+    conn = connect()
+    conn.execute("""CREATE TABLE IF NOT EXISTS model_usage (
+        attempt_id TEXT PRIMARY KEY, provider_id TEXT NOT NULL, model TEXT NOT NULL,
+        capability TEXT NOT NULL, source TEXT NOT NULL, started_at TEXT NOT NULL,
+        completed INTEGER NOT NULL, counters_json TEXT NOT NULL, raw_json TEXT NOT NULL)""")
+    conn.execute("CREATE INDEX IF NOT EXISTS usage_time_provider ON model_usage(started_at,provider_id,model)")
+    columns = {row[1] for row in conn.execute("PRAGMA table_info(model_usage)")}
+    for column in ("request_id", "run_id"):
+        if column not in columns:
+            conn.execute(f"ALTER TABLE model_usage ADD COLUMN {column} TEXT")
+    return conn
+
+
+def numeric_leaves(value, prefix=""):
+    """Keep known numerical counters only; vendor usage objects may contain arbitrary text."""
+    result = {}
+    if not isinstance(value, dict):
+        return result
+    allowed = {"prompt_tokens", "completion_tokens", "input_tokens", "output_tokens", "total_tokens", "cached_tokens",
+               "cache_read_input_tokens", "cache_creation_input_tokens", "prompt_cache_hit_tokens", "prompt_cache_miss_tokens",
+               "reasoning_tokens", "prompt_eval_count", "eval_count"}
+    for key, item in value.items():
+        path = f"{prefix}.{key}" if prefix else key
+        if key in allowed and type(item) is int and 0 <= item <= 2 ** 53:
+            result[path] = item
+        elif key in {"prompt_tokens_details", "completion_tokens_details", "input_tokens_details", "output_tokens_details"}:
+            result.update(numeric_leaves(item, path))
+    return result
+
+
+class UsageAttempt:
+    def __init__(self, provider_id, model, protocol, capability="chat", source="api"):
+        self.attempt_id = uuid4().hex
+        self.provider_id, self.model, self.protocol = provider_id, model, protocol
+        self.capability, self.source = capability, source
+        self.started_at = datetime.now(timezone.utc).isoformat()
+        self.raw = {}
+        self.completed = False
+        context = usage_context.get() or {}
+        self.request_id = context.get("request_id") or uuid4().hex
+        self.run_id = context.get("run_id")
+
+    def observe(self, data):
+        if not isinstance(data, dict):
+            return
+        values = [data.get("usage"), (data.get("message") or {}).get("usage") if isinstance(data.get("message"), dict) else None,
+                  (data.get("response") or {}).get("usage") if isinstance(data.get("response"), dict) else None]
+        if self.protocol == "ollama":
+            values.append(data)
+        for value in values:
+            for key, count in numeric_leaves(value).items():
+                self.raw[key] = max(self.raw.get(key, 0), count)
+        if data.get("type") in {"[DONE]", "response.completed", "message_stop"} or data.get("done") is True:
+            self.completed = True
+
+    def counters(self):
+        raw = self.raw
+        def first(*names):
+            return next((raw[name] for name in names if name in raw), None)
+        inputs = first("input_tokens", "prompt_tokens", "prompt_eval_count")
+        outputs = first("output_tokens", "completion_tokens", "eval_count")
+        hit = first("cache_read_input_tokens", "prompt_cache_hit_tokens", "input_tokens_details.cached_tokens", "prompt_tokens_details.cached_tokens")
+        write = first("cache_creation_input_tokens")
+        miss = first("prompt_cache_miss_tokens")
+        if self.protocol == "anthropic_messages":
+            miss = inputs
+            inputs = inputs + hit + write if inputs is not None and hit is not None and write is not None else None
+        elif miss is None and inputs is not None and hit is not None and 0 <= hit <= inputs:
+            miss = inputs - hit
+        if hit is not None and inputs is not None and hit > inputs:
+            hit, miss = None, None
+        return dict(input_tokens=inputs, output_tokens=outputs,
+                    total_tokens=inputs + outputs if inputs is not None and outputs is not None else first("total_tokens"),
+                    cache_hit_tokens=hit, cache_miss_tokens=miss, cache_write_tokens=write,
+                    reasoning_tokens=first("output_tokens_details.reasoning_tokens", "completion_tokens_details.reasoning_tokens"))
+
+    def persist(self):
+        try:
+            with closing(connection()) as conn:
+                conn.execute("INSERT OR REPLACE INTO model_usage VALUES (?,?,?,?,?,?,?,?,?,?,?)", (
+                    self.attempt_id, self.provider_id, self.model, self.capability, self.source, self.started_at,
+                    int(self.completed), json.dumps(self.counters()), json.dumps(self.raw), self.request_id, self.run_id))
+        except Exception:
+            logger.warning("Usage persistence failed; model response remains available")
+
+
+def aggregate(start, end, provider_id=None, model=None, source=None):
+    query = "SELECT counters_json,completed FROM model_usage WHERE started_at>=? AND started_at<?"
+    args = [start.astimezone(timezone.utc).isoformat(), end.astimezone(timezone.utc).isoformat()]
+    for column, value in (("provider_id", provider_id), ("model", model), ("source", source)):
+        if value:
+            query += f" AND {column}=?"
+            args.append(value)
+    with closing(connection()) as conn:
+        rows = conn.execute(query, args).fetchall()
+        options = conn.execute("SELECT DISTINCT provider_id,model,source FROM model_usage ORDER BY provider_id,model").fetchall()
+    totals = {key: None for key in METRICS}
+    coverage = {key: 0 for key in METRICS}
+    hits, eligible_input, cache_requests = 0, 0, 0
+    for row in rows:
+        counts = json.loads(row[0])
+        for key in METRICS:
+            if counts.get(key) is not None:
+                totals[key] = (totals[key] or 0) + counts[key]
+                coverage[key] += 1
+        if counts.get("cache_hit_tokens") is not None and counts.get("cache_miss_tokens") is not None:
+            hits += counts["cache_hit_tokens"]
+            eligible_input += counts["input_tokens"] if counts.get("input_tokens") is not None else counts["cache_hit_tokens"] + counts["cache_miss_tokens"]
+            cache_requests += 1
+    return {"totals": totals, "coverage": coverage, "request_count": len(rows),
+            "complete_requests": sum(row[1] for row in rows), "cache_covered_requests": cache_requests,
+            "cache_hit_rate": hits / eligible_input if eligible_input else None,
+            "options": [dict(row) for row in options], "start": start, "end": end,
+            "scope": "application_observed_usage"}

backend/app/usage_routes.py

@@ -0,0 +1,19 @@
+from datetime import datetime, timedelta, timezone
+from fastapi import APIRouter, Query
+from app.errors import ApiError
+from app.services.usage_service import aggregate
+
+router = APIRouter(prefix="/api/usage", tags=["Usage"])
+
+
+@router.get("")
+async def usage(start: datetime | None = None, end: datetime | None = None,
+                provider_id: str | None = Query(None, max_length=200), model: str | None = Query(None, max_length=200),
+                source: str | None = None):
+    end = end or datetime.now(timezone.utc)
+    start = start or end - timedelta(days=7)
+    if not start.tzinfo or not end.tzinfo or end <= start:
+        raise ApiError(422, "INVALID_TIME_RANGE", "Provide timezone-aware start/end with end after start.")
+    if source not in {None, "local", "api"}:
+        raise ApiError(422, "INVALID_USAGE_SOURCE", "Unknown usage source.")
+    return aggregate(start, end, provider_id, model, source)