Initial commit

domain/ai/__init__.py

@@ -0,0 +1,67 @@
+from .inference import (
+    MissingModelError,
+    chat_completion,
+    chat_completion_stream,
+    describe_image_base64,
+    get_text_embedding,
+    provider_service,
+    rerank_texts,
+)
+from .service import (
+    AIProviderService,
+    FILE_COLLECTION_NAME,
+    VECTOR_COLLECTION_NAME,
+    DEFAULT_VECTOR_DIMENSION,
+    VectorDBConfigManager,
+    VectorDBService,
+)
+from .types import (
+    ABILITIES,
+    AIDefaultsUpdate,
+    AIModelCreate,
+    AIModelUpdate,
+    AIProviderCreate,
+    AIProviderUpdate,
+    VectorDBConfigPayload,
+    normalize_capabilities,
+)
+from .vector_providers import (
+    BaseVectorProvider,
+    MilvusLiteProvider,
+    MilvusServerProvider,
+    QdrantProvider,
+    get_provider_class,
+    get_provider_entry,
+    list_providers,
+)
+
+__all__ = [
+    "MissingModelError",
+    "chat_completion",
+    "chat_completion_stream",
+    "describe_image_base64",
+    "get_text_embedding",
+    "provider_service",
+    "rerank_texts",
+    "AIProviderService",
+    "VectorDBService",
+    "VectorDBConfigManager",
+    "DEFAULT_VECTOR_DIMENSION",
+    "VECTOR_COLLECTION_NAME",
+    "FILE_COLLECTION_NAME",
+    "BaseVectorProvider",
+    "MilvusLiteProvider",
+    "MilvusServerProvider",
+    "QdrantProvider",
+    "list_providers",
+    "get_provider_entry",
+    "get_provider_class",
+    "ABILITIES",
+    "normalize_capabilities",
+    "AIDefaultsUpdate",
+    "AIModelCreate",
+    "AIModelUpdate",
+    "AIProviderCreate",
+    "AIProviderUpdate",
+    "VectorDBConfigPayload",
+]

domain/ai/api.py

@@ -0,0 +1,304 @@
+from typing import Annotated, Dict, Optional
+
+import httpx
+from fastapi import APIRouter, Depends, HTTPException, Path, Request
+
+from api.response import success
+from domain.audit import AuditAction, audit
+from domain.auth import User, get_current_active_user
+from .service import AIProviderService, VectorDBConfigManager, VectorDBService
+from .types import (
+    AIDefaultsUpdate,
+    AIModelCreate,
+    AIModelUpdate,
+    AIProviderCreate,
+    AIProviderUpdate,
+    VectorDBConfigPayload,
+)
+from .vector_providers import get_provider_class, get_provider_entry, list_providers
+
+router_ai = APIRouter(prefix="/api/ai", tags=["ai"])
+router_vector_db = APIRouter(prefix="/api/vector-db", tags=["vector-db"])
+
+
+@audit(action=AuditAction.READ, description="获取 AI 提供商列表")
+@router_ai.get("/providers")
+async def list_providers_endpoint(
+    request: Request,
+    current_user: Annotated[User, Depends(get_current_active_user)]
+):
+    providers = await AIProviderService.list_providers()
+    return success({"providers": providers})
+
+
+@audit(
+    action=AuditAction.CREATE,
+    description="创建 AI 提供商",
+    body_fields=["name", "identifier", "provider_type", "api_format", "base_url", "logo_url"],
+    redact_fields=["api_key"],
+)
+@router_ai.post("/providers")
+async def create_provider(
+    request: Request,
+    payload: AIProviderCreate,
+    current_user: Annotated[User, Depends(get_current_active_user)]
+):
+    provider = await AIProviderService.create_provider(payload.dict())
+    return success(provider)
+
+
+@audit(action=AuditAction.READ, description="获取 AI 提供商详情")
+@router_ai.get("/providers/{provider_id}")
+async def get_provider(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    provider = await AIProviderService.get_provider(provider_id, with_models=True)
+    return success(provider)
+
+
+@audit(
+    action=AuditAction.UPDATE,
+    description="更新 AI 提供商",
+    body_fields=["name", "provider_type", "api_format", "base_url", "logo_url", "api_key"],
+    redact_fields=["api_key"],
+)
+@router_ai.put("/providers/{provider_id}")
+async def update_provider(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    payload: AIProviderUpdate,
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    data = {k: v for k, v in payload.dict().items() if v is not None}
+    if not data:
+        raise HTTPException(status_code=400, detail="No fields to update")
+    provider = await AIProviderService.update_provider(provider_id, data)
+    return success(provider)
+
+
+@audit(action=AuditAction.DELETE, description="删除 AI 提供商")
+@router_ai.delete("/providers/{provider_id}")
+async def delete_provider(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    await AIProviderService.delete_provider(provider_id)
+    return success({"id": provider_id})
+
+
+@audit(action=AuditAction.UPDATE, description="同步模型列表")
+@router_ai.post("/providers/{provider_id}/sync-models")
+async def sync_models(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    try:
+        result = await AIProviderService.sync_models(provider_id)
+    except (httpx.RequestError, httpx.HTTPStatusError) as exc:
+        raise HTTPException(status_code=502, detail=f"Failed to synchronize models: {exc}") from exc
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    return success(result)
+
+
+@audit(action=AuditAction.READ, description="获取远程模型列表")
+@router_ai.get("/providers/{provider_id}/remote-models")
+async def fetch_remote_models(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    try:
+        models = await AIProviderService.fetch_remote_models(provider_id)
+    except (httpx.RequestError, httpx.HTTPStatusError) as exc:
+        raise HTTPException(status_code=502, detail=f"Failed to pull models: {exc}") from exc
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    return success({"models": models})
+
+
+@audit(action=AuditAction.READ, description="获取模型列表")
+@router_ai.get("/providers/{provider_id}/models")
+async def list_models(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    models = await AIProviderService.list_models(provider_id)
+    return success({"models": models})
+
+
+@audit(
+    action=AuditAction.CREATE,
+    description="创建模型",
+    body_fields=["name", "display_name", "capabilities", "context_window", "embedding_dimensions"],
+)
+@router_ai.post("/providers/{provider_id}/models")
+async def create_model(
+    request: Request,
+    provider_id: Annotated[int, Path(..., gt=0)],
+    payload: AIModelCreate,
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    model = await AIProviderService.create_model(provider_id, payload.dict())
+    return success(model)
+
+
+@audit(
+    action=AuditAction.UPDATE,
+    description="更新模型",
+    body_fields=["display_name", "description", "capabilities", "context_window", "embedding_dimensions"],
+)
+@router_ai.put("/models/{model_id}")
+async def update_model(
+    request: Request,
+    model_id: Annotated[int, Path(..., gt=0)],
+    payload: AIModelUpdate,
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    data = {k: v for k, v in payload.dict().items() if v is not None}
+    if not data:
+        raise HTTPException(status_code=400, detail="No fields to update")
+    model = await AIProviderService.update_model(model_id, data)
+    return success(model)
+
+
+@audit(action=AuditAction.DELETE, description="删除模型")
+@router_ai.delete("/models/{model_id}")
+async def delete_model(
+    request: Request,
+    model_id: Annotated[int, Path(..., gt=0)],
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    await AIProviderService.delete_model(model_id)
+    return success({"id": model_id})
+
+
+def _get_embedding_dimension(entry: Optional[Dict]) -> Optional[int]:
+    if not entry:
+        return None
+    value = entry.get("embedding_dimensions")
+    return int(value) if value is not None else None
+
+
+@audit(action=AuditAction.READ, description="获取默认模型")
+@router_ai.get("/defaults")
+async def get_defaults(
+    request: Request,
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    defaults = await AIProviderService.get_default_models()
+    return success(defaults)
+
+
+@audit(
+    action=AuditAction.UPDATE,
+    description="更新默认模型",
+    body_fields=["chat", "vision", "embedding", "rerank", "voice", "tools"],
+)
+@router_ai.put("/defaults")
+async def update_defaults(
+    request: Request,
+    payload: AIDefaultsUpdate,
+    current_user: Annotated[User, Depends(get_current_active_user)],
+):
+    previous = await AIProviderService.get_default_models()
+    try:
+        updated = await AIProviderService.set_default_models(payload.as_mapping())
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    prev_dim = _get_embedding_dimension(previous.get("embedding"))
+    next_dim = _get_embedding_dimension(updated.get("embedding"))
+
+    if prev_dim and next_dim and prev_dim != next_dim:
+        try:
+            await VectorDBService().clear_all_data()
+        except Exception as exc:  # noqa: BLE001
+            raise HTTPException(status_code=500, detail=f"Failed to clear vector database: {exc}") from exc
+
+    return success(updated)
+
+
+@audit(action=AuditAction.UPDATE, description="清空向量数据库")
+@router_vector_db.post("/clear-all", summary="清空向量数据库")
+async def clear_vector_db(request: Request, user: User = Depends(get_current_active_user)):
+    try:
+        service = VectorDBService()
+        await service.clear_all_data()
+        return success(msg="向量数据库已清空")
+    except Exception as e:  # noqa: BLE001
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@audit(action=AuditAction.READ, description="获取向量数据库统计")
+@router_vector_db.get("/stats", summary="获取向量数据库统计")
+async def get_vector_db_stats(request: Request, user: User = Depends(get_current_active_user)):
+    try:
+        service = VectorDBService()
+        data = await service.get_all_stats()
+        return success(data=data)
+    except Exception as e:  # noqa: BLE001
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@audit(action=AuditAction.READ, description="获取向量数据库提供者列表")
+@router_vector_db.get("/providers", summary="列出可用向量数据库提供者")
+async def list_vector_providers(request: Request):
+    return success(list_providers())
+
+
+@audit(action=AuditAction.READ, description="获取向量数据库配置")
+@router_vector_db.get("/config", summary="获取当前向量数据库配置")
+async def get_vector_db_config(request: Request, user: User = Depends(get_current_active_user)):
+    service = VectorDBService()
+    data = await service.current_provider()
+    return success(data)
+
+
+@audit(action=AuditAction.UPDATE, description="更新向量数据库配置", body_fields=["type"])
+@router_vector_db.post("/config", summary="更新向量数据库配置")
+async def update_vector_db_config(
+    request: Request, payload: VectorDBConfigPayload, user: User = Depends(get_current_active_user)
+):
+    entry = get_provider_entry(payload.type)
+    if not entry:
+        raise HTTPException(
+            status_code=400, detail=f"未知的向量数据库类型: {payload.type}")
+    if not entry.get("enabled", True):
+        raise HTTPException(status_code=400, detail="该向量数据库类型暂不可用")
+
+    provider_cls = get_provider_class(payload.type)
+    if not provider_cls:
+        raise HTTPException(
+            status_code=400, detail=f"未找到类型 {payload.type} 对应的实现")
+
+    test_provider = provider_cls(payload.config)
+    try:
+        await test_provider.initialize()
+    except Exception as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    finally:
+        client = getattr(test_provider, "client", None)
+        close_fn = getattr(client, "close", None)
+        if callable(close_fn):
+            try:
+                close_fn()
+            except Exception:
+                pass
+
+    await VectorDBConfigManager.save_config(payload.type, payload.config)
+    service = VectorDBService()
+    await service.reload()
+    config_data = await service.current_provider()
+    stats = await service.get_all_stats()
+    return success({"config": config_data, "stats": stats})
+
+
+__all__ = ["router_ai", "router_vector_db"]

domain/ai/service.py

@@ -0,0 +1,501 @@
+import asyncio
+import json
+from collections.abc import Iterable
+from typing import Any, Dict, List, Optional, Tuple
+
+import httpx
+from tortoise.exceptions import DoesNotExist
+from tortoise.transactions import in_transaction
+
+from domain.config import ConfigService
+from models.database import AIDefaultModel, AIModel, AIProvider
+
+from .types import ABILITIES, normalize_capabilities
+from .vector_providers import (
+    BaseVectorProvider,
+    get_provider_class,
+    get_provider_entry,
+    list_providers,
+)
+
+DEFAULT_VECTOR_DIMENSION = 4096
+VECTOR_COLLECTION_NAME = "vector_collection"
+FILE_COLLECTION_NAME = "file_collection"
+
+OPENAI_EMBEDDING_DIMS = {
+    "text-embedding-3-large": 3072,
+    "text-embedding-3-small": 1536,
+    "text-embedding-ada-002": 1536,
+}
+
+
+class VectorDBConfigManager:
+    TYPE_KEY = "VECTOR_DB_TYPE"
+    CONFIG_KEY = "VECTOR_DB_CONFIG"
+    DEFAULT_TYPE = "milvus_lite"
+
+    @classmethod
+    async def load_config(cls) -> Tuple[str, Dict[str, Any]]:
+        raw_type = await ConfigService.get(cls.TYPE_KEY, cls.DEFAULT_TYPE)
+        provider_type = str(raw_type or cls.DEFAULT_TYPE)
+
+        raw_config = await ConfigService.get(cls.CONFIG_KEY)
+        config_dict: Dict[str, Any] = {}
+        if isinstance(raw_config, str) and raw_config:
+            try:
+                config_dict = json.loads(raw_config)
+            except json.JSONDecodeError:
+                config_dict = {}
+        elif isinstance(raw_config, dict):
+            config_dict = raw_config
+        return provider_type, config_dict
+
+    @classmethod
+    async def save_config(cls, provider_type: str, config: Dict[str, Any]) -> None:
+        await ConfigService.set(cls.TYPE_KEY, provider_type)
+        await ConfigService.set(cls.CONFIG_KEY, json.dumps(config or {}))
+
+    @classmethod
+    async def get_type(cls) -> str:
+        provider_type, _ = await cls.load_config()
+        return provider_type
+
+    @classmethod
+    async def get_config(cls) -> Dict[str, Any]:
+        _, config = await cls.load_config()
+        return config
+
+
+def _normalize_embedding_dim(value: Any) -> Optional[int]:
+    if value is None:
+        return None
+    try:
+        casted = int(value)
+    except (TypeError, ValueError):
+        return None
+    return casted if casted > 0 else None
+
+
+def _apply_embedding_dim_to_metadata(
+    data: Dict[str, Any],
+    embedding_dim: Optional[int],
+    base_metadata: Optional[Dict[str, Any]] = None,
+) -> Dict[str, Any]:
+    source = base_metadata if isinstance(base_metadata, dict) else {}
+    metadata: Dict[str, Any] = dict(source)
+    override = data.get("metadata")
+    if isinstance(override, dict) and override:
+        metadata.update(override)
+    if embedding_dim is None:
+        metadata.pop("embedding_dimensions", None)
+    else:
+        metadata["embedding_dimensions"] = embedding_dim
+    data["metadata"] = metadata or None
+    return data
+
+
+def infer_openai_capabilities(model_id: str) -> Tuple[List[str], Optional[int]]:
+    lower = model_id.lower()
+    caps = set()
+
+    if any(keyword in lower for keyword in ["gpt", "chat", "turbo", "o1", "sonnet", "haiku", "thinking"]):
+        caps.update({"chat", "tools"})
+
+    if any(keyword in lower for keyword in ["vision", "gpt-4o", "gpt-4.1", "o1", "vision-preview", "omni"]):
+        caps.add("vision")
+
+    if any(keyword in lower for keyword in ["embed", "embedding"]):
+        caps.add("embedding")
+
+    if "rerank" in lower or "re-rank" in lower:
+        caps.add("rerank")
+
+    if any(keyword in lower for keyword in ["tts", "speech", "audio"]):
+        caps.add("voice")
+
+    embedding_dim = OPENAI_EMBEDDING_DIMS.get(model_id)
+    return normalize_capabilities(caps), embedding_dim
+
+
+def infer_gemini_capabilities(methods: Iterable[str]) -> List[str]:
+    caps = set()
+    for method in methods:
+        m = method.lower()
+        if m in {"generatecontent", "counttokens"}:
+            caps.update({"chat", "tools", "vision"})
+        if m == "embedcontent":
+            caps.add("embedding")
+        if m in {"generatespeech", "audiogeneration"}:
+            caps.add("voice")
+        if m == "rerank":
+            caps.add("rerank")
+    return normalize_capabilities(caps)
+
+
+def serialize_provider(provider: AIProvider) -> Dict[str, Any]:
+    return {
+        "id": provider.id,
+        "name": provider.name,
+        "identifier": provider.identifier,
+        "provider_type": provider.provider_type,
+        "api_format": provider.api_format,
+        "base_url": provider.base_url,
+        "has_api_key": bool(provider.api_key),
+        "logo_url": provider.logo_url,
+        "extra_config": provider.extra_config or {},
+        "created_at": provider.created_at,
+        "updated_at": provider.updated_at,
+    }
+
+
+def model_to_dict(model: AIModel, provider: Optional[AIProvider] = None) -> Dict[str, Any]:
+    provider_obj = provider or getattr(model, "provider", None)
+    provider_data = serialize_provider(provider_obj) if provider_obj else None
+    return {
+        "id": model.id,
+        "provider_id": model.provider_id,
+        "name": model.name,
+        "display_name": model.display_name,
+        "description": model.description,
+        "capabilities": normalize_capabilities(model.capabilities),
+        "context_window": model.context_window,
+        "embedding_dimensions": model.embedding_dimensions,
+        "metadata": model.metadata or {},
+        "created_at": model.created_at,
+        "updated_at": model.updated_at,
+        "provider": provider_data,
+    }
+
+
+def provider_to_dict(provider: AIProvider, models: Optional[List[AIModel]] = None) -> Dict[str, Any]:
+    data = serialize_provider(provider)
+    if models is not None:
+        data["models"] = [model_to_dict(m, provider=provider) for m in models]
+    return data
+
+
+class AIProviderService:
+    @classmethod
+    async def list_providers(cls) -> List[Dict[str, Any]]:
+        providers = await AIProvider.all().order_by("id").prefetch_related("models")
+        return [provider_to_dict(p, models=list(p.models)) for p in providers]
+
+    @classmethod
+    async def get_provider(cls, provider_id: int, with_models: bool = False) -> Dict[str, Any]:
+        if with_models:
+            provider = await AIProvider.get(id=provider_id)
+            models = await provider.models.all()
+            return provider_to_dict(provider, models=models)
+        provider = await AIProvider.get(id=provider_id)
+        return provider_to_dict(provider)
+
+    @classmethod
+    async def create_provider(cls, payload: Dict[str, Any]) -> Dict[str, Any]:
+        data = payload.copy()
+        data.setdefault("extra_config", {})
+        provider = await AIProvider.create(**data)
+        return provider_to_dict(provider)
+
+    @classmethod
+    async def update_provider(cls, provider_id: int, payload: Dict[str, Any]) -> Dict[str, Any]:
+        provider = await AIProvider.get(id=provider_id)
+        for field, value in payload.items():
+            setattr(provider, field, value)
+        await provider.save()
+        return provider_to_dict(provider)
+
+    @classmethod
+    async def delete_provider(cls, provider_id: int) -> None:
+        await AIProvider.filter(id=provider_id).delete()
+
+    @classmethod
+    async def list_models(cls, provider_id: int) -> List[Dict[str, Any]]:
+        models = await AIModel.filter(provider_id=provider_id).order_by("id").prefetch_related("provider")
+        return [model_to_dict(m) for m in models]
+
+    @classmethod
+    async def create_model(cls, provider_id: int, payload: Dict[str, Any]) -> Dict[str, Any]:
+        data = payload.copy()
+        data["provider_id"] = provider_id
+        data["capabilities"] = normalize_capabilities(data.get("capabilities"))
+        embedding_dim = _normalize_embedding_dim(data.pop("embedding_dimensions", None))
+        data = _apply_embedding_dim_to_metadata(data, embedding_dim)
+        model = await AIModel.create(**data)
+        await model.fetch_related("provider")
+        return model_to_dict(model)
+
+    @classmethod
+    async def update_model(cls, model_id: int, payload: Dict[str, Any]) -> Dict[str, Any]:
+        model = await AIModel.get(id=model_id)
+        data = payload.copy()
+        if "capabilities" in data:
+            data["capabilities"] = normalize_capabilities(data.get("capabilities"))
+        embedding_dim = None
+        if "embedding_dimensions" in data:
+            embedding_dim = _normalize_embedding_dim(data.pop("embedding_dimensions", None))
+            _apply_embedding_dim_to_metadata(data, embedding_dim, base_metadata=model.metadata)
+        for field, value in data.items():
+            setattr(model, field, value)
+        if embedding_dim is not None or ("embedding_dimensions" in payload and embedding_dim is None):
+            model.embedding_dimensions = embedding_dim
+        await model.save()
+        await model.fetch_related("provider")
+        return model_to_dict(model)
+
+    @classmethod
+    async def delete_model(cls, model_id: int) -> None:
+        await AIModel.filter(id=model_id).delete()
+
+    @classmethod
+    async def fetch_remote_models(cls, provider_id: int) -> List[Dict[str, Any]]:
+        provider = await AIProvider.get(id=provider_id)
+        return await cls._get_remote_models(provider)
+
+    @classmethod
+    async def _get_remote_models(cls, provider: AIProvider) -> List[Dict[str, Any]]:
+        if not provider.base_url:
+            raise ValueError("Provider base_url is required for syncing models")
+
+        fmt = (provider.api_format or "").lower()
+        if fmt not in {"openai", "gemini"}:
+            raise ValueError(f"Unsupported api_format '{provider.api_format}' for syncing models")
+
+        if fmt == "openai":
+            return await cls._fetch_openai_models(provider)
+        return await cls._fetch_gemini_models(provider)
+
+    @classmethod
+    async def sync_models(cls, provider_id: int) -> Dict[str, int]:
+        provider = await AIProvider.get(id=provider_id)
+        remote_models = await cls._get_remote_models(provider)
+
+        created = 0
+        updated = 0
+        for entry in remote_models:
+            defaults = entry.copy()
+            model_id = defaults.pop("name")
+            defaults["capabilities"] = normalize_capabilities(defaults.get("capabilities"))
+            embedding_dim = _normalize_embedding_dim(defaults.pop("embedding_dimensions", None))
+            defaults = _apply_embedding_dim_to_metadata(defaults, embedding_dim)
+            obj, is_created = await AIModel.get_or_create(
+                provider_id=provider.id,
+                name=model_id,
+                defaults=defaults,
+            )
+            if is_created:
+                created += 1
+                continue
+            for field, value in defaults.items():
+                setattr(obj, field, value)
+            if embedding_dim is not None or ("embedding_dimensions" in entry and embedding_dim is None):
+                obj.embedding_dimensions = embedding_dim
+            await obj.save()
+            updated += 1
+
+        return {"created": created, "updated": updated}
+
+    @classmethod
+    async def get_default_models(cls) -> Dict[str, Optional[Dict[str, Any]]]:
+        defaults = await AIDefaultModel.all().prefetch_related("model__provider")
+        result: Dict[str, Optional[Dict[str, Any]]] = {ability: None for ability in ABILITIES}
+        for item in defaults:
+            result[item.ability] = model_to_dict(item.model, provider=item.model.provider)  # type: ignore[attr-defined]
+        return result
+
+    @classmethod
+    async def set_default_models(cls, mapping: Dict[str, Optional[int]]) -> Dict[str, Optional[Dict[str, Any]]]:
+        normalized = {ability: mapping.get(ability) for ability in ABILITIES}
+        async with in_transaction() as connection:
+            for ability, model_id in normalized.items():
+                record = await AIDefaultModel.get_or_none(ability=ability)
+                if model_id:
+                    try:
+                        model = await AIModel.get(id=model_id)
+                    except DoesNotExist:
+                        raise ValueError(f"Model {model_id} not found")
+                    if record:
+                        record.model_id = model_id
+                        await record.save(using_db=connection)
+                    else:
+                        await AIDefaultModel.create(ability=ability, model_id=model_id)
+                elif record:
+                    await record.delete(using_db=connection)
+        return await cls.get_default_models()
+
+    @classmethod
+    async def get_default_model(cls, ability: str) -> Optional[AIModel]:
+        ability_key = ability.lower()
+        if ability_key not in ABILITIES:
+            return None
+        record = await AIDefaultModel.get_or_none(ability=ability_key)
+        if not record:
+            return None
+        model = await AIModel.get_or_none(id=record.model_id)
+        if model:
+            await model.fetch_related("provider")
+        return model
+
+    @classmethod
+    async def _fetch_openai_models(cls, provider: AIProvider) -> List[Dict[str, Any]]:
+        base_url = provider.base_url.rstrip("/")
+        url = f"{base_url}/models"
+        headers = {}
+        if provider.api_key:
+            headers["Authorization"] = f"Bearer {provider.api_key}"
+
+        async with httpx.AsyncClient(timeout=30.0) as client:
+            response = await client.get(url, headers=headers)
+            response.raise_for_status()
+            payload = response.json()
+
+        data = payload.get("data", [])
+        entries: List[Dict[str, Any]] = []
+        for item in data:
+            model_id = item.get("id")
+            if not model_id:
+                continue
+            capabilities, embedding_dim = infer_openai_capabilities(model_id)
+            entries.append({
+                "name": model_id,
+                "display_name": item.get("display_name"),
+                "description": item.get("description"),
+                "capabilities": capabilities,
+                "context_window": item.get("context_window"),
+                "embedding_dimensions": embedding_dim,
+                "metadata": item,
+            })
+        return entries
+
+    @classmethod
+    async def _fetch_gemini_models(cls, provider: AIProvider) -> List[Dict[str, Any]]:
+        base_url = provider.base_url.rstrip("/")
+        suffix = "/models"
+        if provider.api_key:
+            suffix += f"?key={provider.api_key}"
+        url = f"{base_url}{suffix}"
+
+        async with httpx.AsyncClient(timeout=30.0) as client:
+            response = await client.get(url)
+            response.raise_for_status()
+            payload = response.json()
+
+        data = payload.get("models", [])
+        entries: List[Dict[str, Any]] = []
+        for item in data:
+            model_id = item.get("name")
+            if not model_id:
+                continue
+            methods = item.get("supportedGenerationMethods") or []
+            capabilities = infer_gemini_capabilities(methods)
+            entries.append({
+                "name": model_id,
+                "display_name": item.get("displayName"),
+                "description": item.get("description"),
+                "capabilities": capabilities,
+                "context_window": item.get("inputTokenLimit"),
+                "embedding_dimensions": item.get("embeddingDimensions"),
+                "metadata": item,
+            })
+        return entries
+
+
+class VectorDBService:
+    _instance: Optional["VectorDBService"] = None
+
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self):
+        if not hasattr(self, "_provider"):
+            self._provider: Optional[BaseVectorProvider] = None
+            self._provider_type: Optional[str] = None
+            self._provider_config: Dict[str, Any] | None = None
+            self._lock = asyncio.Lock()
+
+    async def _ensure_provider(self) -> BaseVectorProvider:
+        if self._provider is None:
+            await self.reload()
+        assert self._provider is not None
+        return self._provider
+
+    async def reload(self) -> BaseVectorProvider:
+        async with self._lock:
+            provider_type, provider_config = await VectorDBConfigManager.load_config()
+            normalized_config = dict(provider_config or {})
+            if (
+                self._provider
+                and self._provider_type == provider_type
+                and self._provider_config == normalized_config
+            ):
+                return self._provider
+
+            entry = get_provider_entry(provider_type)
+            if not entry:
+                raise RuntimeError(f"Unknown vector database provider: {provider_type}")
+            if not entry.get("enabled", True):
+                raise RuntimeError(f"Vector database provider '{provider_type}' is disabled")
+
+            provider_cls = get_provider_class(provider_type)
+            if not provider_cls:
+                raise RuntimeError(f"Provider class not found for '{provider_type}'")
+
+            provider = provider_cls(provider_config)
+            await provider.initialize()
+
+            self._provider = provider
+            self._provider_type = provider_type
+            self._provider_config = normalized_config
+            return provider
+
+    async def ensure_collection(self, collection_name: str, vector: bool = True, dim: int = DEFAULT_VECTOR_DIMENSION) -> None:
+        provider = await self._ensure_provider()
+        provider.ensure_collection(collection_name, vector, dim)
+
+    async def upsert_vector(self, collection_name: str, data: Dict[str, Any]) -> None:
+        provider = await self._ensure_provider()
+        provider.upsert_vector(collection_name, data)
+
+    async def delete_vector(self, collection_name: str, path: str) -> None:
+        provider = await self._ensure_provider()
+        provider.delete_vector(collection_name, path)
+
+    async def search_vectors(self, collection_name: str, query_embedding, top_k: int = 5):
+        provider = await self._ensure_provider()
+        return provider.search_vectors(collection_name, query_embedding, top_k)
+
+    async def search_by_path(self, collection_name: str, query_path: str, top_k: int = 20):
+        provider = await self._ensure_provider()
+        return provider.search_by_path(collection_name, query_path, top_k)
+
+    async def get_all_stats(self) -> Dict[str, Any]:
+        provider = await self._ensure_provider()
+        return provider.get_all_stats()
+
+    async def clear_all_data(self) -> None:
+        provider = await self._ensure_provider()
+        provider.clear_all_data()
+
+    async def current_provider(self) -> Dict[str, Any]:
+        provider_type, provider_config = await VectorDBConfigManager.load_config()
+        entry = get_provider_entry(provider_type) or {}
+        return {
+            "type": provider_type,
+            "config": provider_config,
+            "label": entry.get("label"),
+            "enabled": entry.get("enabled", True),
+        }
+
+
+__all__ = [
+    "AIProviderService",
+    "VectorDBService",
+    "VectorDBConfigManager",
+    "DEFAULT_VECTOR_DIMENSION",
+    "list_providers",
+    "get_provider_entry",
+    "get_provider_class",
+    "normalize_capabilities",
+    "ABILITIES",
+]

domain/ai/types.py

@@ -0,0 +1,121 @@
+from typing import Any, Dict, Iterable, List, Optional
+
+from pydantic import BaseModel, Field, field_validator
+
+ABILITIES = ["chat", "vision", "embedding", "rerank", "voice", "tools"]
+
+
+def normalize_capabilities(items: Optional[Iterable[str]]) -> List[str]:
+    if not items:
+        return []
+    normalized: List[str] = []
+    for cap in items:
+        key = str(cap).strip().lower()
+        if key in ABILITIES and key not in normalized:
+            normalized.append(key)
+    return normalized
+
+
+class AIProviderBase(BaseModel):
+    name: str
+    identifier: str = Field(..., pattern=r"^[a-z0-9_\-\.]+$")
+    provider_type: Optional[str] = None
+    api_format: str
+    base_url: Optional[str] = None
+    api_key: Optional[str] = None
+    logo_url: Optional[str] = None
+    extra_config: Optional[dict] = None
+
+    @field_validator("api_format")
+    @classmethod
+    def normalize_format(cls, value: str) -> str:
+        fmt = value.lower()
+        if fmt not in {"openai", "gemini", "anthropic", "ollama"}:
+            raise ValueError("api_format must be 'openai', 'gemini', 'anthropic', or 'ollama'")
+        return fmt
+
+
+class AIProviderCreate(AIProviderBase):
+    pass
+
+
+class AIProviderUpdate(BaseModel):
+    name: Optional[str] = None
+    provider_type: Optional[str] = None
+    api_format: Optional[str] = None
+    base_url: Optional[str] = None
+    api_key: Optional[str] = None
+    logo_url: Optional[str] = None
+    extra_config: Optional[dict] = None
+
+    @field_validator("api_format")
+    @classmethod
+    def normalize_format(cls, value: Optional[str]) -> Optional[str]:
+        if value is None:
+            return value
+        fmt = value.lower()
+        if fmt not in {"openai", "gemini", "anthropic", "ollama"}:
+            raise ValueError("api_format must be 'openai', 'gemini', 'anthropic', or 'ollama'")
+        return fmt
+
+
+class AIModelBase(BaseModel):
+    name: str
+    display_name: Optional[str] = None
+    description: Optional[str] = None
+    capabilities: Optional[List[str]] = None
+    context_window: Optional[int] = None
+    embedding_dimensions: Optional[int] = None
+    metadata: Optional[dict] = None
+
+    @field_validator("capabilities")
+    @classmethod
+    def validate_capabilities(cls, items: Optional[List[str]]) -> Optional[List[str]]:
+        if items is None:
+            return None
+        normalized = normalize_capabilities(items)
+        invalid = set(items) - set(normalized)
+        if invalid:
+            raise ValueError(f"Unsupported capabilities: {', '.join(invalid)}")
+        return normalized
+
+
+class AIModelCreate(AIModelBase):
+    pass
+
+
+class AIModelUpdate(BaseModel):
+    display_name: Optional[str] = None
+    description: Optional[str] = None
+    capabilities: Optional[List[str]] = None
+    context_window: Optional[int] = None
+    embedding_dimensions: Optional[int] = None
+    metadata: Optional[dict] = None
+
+    @field_validator("capabilities")
+    @classmethod
+    def validate_capabilities(cls, items: Optional[List[str]]) -> Optional[List[str]]:
+        if items is None:
+            return None
+        normalized = normalize_capabilities(items)
+        invalid = set(items) - set(normalized)
+        if invalid:
+            raise ValueError(f"Unsupported capabilities: {', '.join(invalid)}")
+        return normalized
+
+
+class AIDefaultsUpdate(BaseModel):
+    chat: Optional[int] = None
+    vision: Optional[int] = None
+    embedding: Optional[int] = None
+    rerank: Optional[int] = None
+    voice: Optional[int] = None
+    tools: Optional[int] = None
+
+    def as_mapping(self) -> Dict[str, Optional[int]]:
+        return {ability: getattr(self, ability) for ability in ABILITIES}
+
+
+class VectorDBConfigPayload(BaseModel):
+    type: str = Field(..., description="向量数据库提供者类型")
+    config: Dict[str, Any] = Field(default_factory=dict, description="提供者配置参数")

domain/ai/vector_providers/__init__.py

@@ -0,0 +1,65 @@
+from typing import Dict, List, Type
+
+from .base import BaseVectorProvider
+from .milvus_lite import MilvusLiteProvider
+from .milvus_server import MilvusServerProvider
+from .qdrant import QdrantProvider
+
+_PROVIDER_REGISTRY: Dict[str, Dict[str, object]] = {
+    MilvusLiteProvider.type: {
+        "class": MilvusLiteProvider,
+        "label": MilvusLiteProvider.label,
+        "description": MilvusLiteProvider.description,
+        "enabled": MilvusLiteProvider.enabled,
+        "config_schema": MilvusLiteProvider.config_schema,
+    },
+    MilvusServerProvider.type: {
+        "class": MilvusServerProvider,
+        "label": MilvusServerProvider.label,
+        "description": MilvusServerProvider.description,
+        "enabled": MilvusServerProvider.enabled,
+        "config_schema": MilvusServerProvider.config_schema,
+    },
+    QdrantProvider.type: {
+        "class": QdrantProvider,
+        "label": QdrantProvider.label,
+        "description": QdrantProvider.description,
+        "enabled": QdrantProvider.enabled,
+        "config_schema": QdrantProvider.config_schema,
+    },
+}
+
+
+def list_providers() -> List[Dict[str, object]]:
+    return [
+        {
+            "type": type_key,
+            "label": meta["label"],
+            "description": meta.get("description"),
+            "enabled": meta.get("enabled", True),
+            "config_schema": meta.get("config_schema", []),
+        }
+        for type_key, meta in _PROVIDER_REGISTRY.items()
+    ]
+
+
+def get_provider_entry(provider_type: str) -> Dict[str, object] | None:
+    return _PROVIDER_REGISTRY.get(provider_type)
+
+
+def get_provider_class(provider_type: str) -> Type[BaseVectorProvider] | None:
+    entry = get_provider_entry(provider_type)
+    if not entry:
+        return None
+    return entry.get("class")  # type: ignore[return-value]
+
+
+__all__ = [
+    "BaseVectorProvider",
+    "MilvusLiteProvider",
+    "MilvusServerProvider",
+    "QdrantProvider",
+    "list_providers",
+    "get_provider_entry",
+    "get_provider_class",
+]

domain/ai/vector_providers/base.py

@@ -0,0 +1,39 @@
+from typing import Any, Dict, List
+
+
+class BaseVectorProvider:
+    """向量数据库提供者基础类，所有实际实现需继承该类"""
+
+    type: str = ""
+    label: str = ""
+    description: str | None = None
+    enabled: bool = True
+    config_schema: List[Dict[str, Any]] = []
+
+    def __init__(self, config: Dict[str, Any] | None = None):
+        self.config = config or {}
+
+    async def initialize(self) -> None:
+        """执行初始化逻辑，例如建立连接"""
+        raise NotImplementedError
+
+    def ensure_collection(self, collection_name: str, vector: bool, dim: int) -> None:
+        raise NotImplementedError
+
+    def upsert_vector(self, collection_name: str, data: Dict[str, Any]) -> None:
+        raise NotImplementedError
+
+    def delete_vector(self, collection_name: str, path: str) -> None:
+        raise NotImplementedError
+
+    def search_vectors(self, collection_name: str, query_embedding, top_k: int):
+        raise NotImplementedError
+
+    def search_by_path(self, collection_name: str, query_path: str, top_k: int):
+        raise NotImplementedError
+
+    def get_all_stats(self) -> Dict[str, Any]:
+        raise NotImplementedError
+
+    def clear_all_data(self) -> None:
+        raise NotImplementedError

domain/ai/vector_providers/milvus_lite.py

@@ -0,0 +1,276 @@
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from pymilvus import CollectionSchema, DataType, FieldSchema, MilvusClient
+
+from .base import BaseVectorProvider
+
+
+class MilvusLiteProvider(BaseVectorProvider):
+    type = "milvus_lite"
+    label = "Milvus Lite"
+    description = "Embedded Milvus Lite (local file storage)."
+    enabled = True
+    config_schema: List[Dict[str, Any]] = [
+        {
+            "key": "db_path",
+            "label": "Database file path",
+            "type": "text",
+            "default": "data/db/milvus.db",
+            "required": False,
+        }
+    ]
+
+    def __init__(self, config: Dict[str, Any] | None = None):
+        super().__init__(config)
+        self.db_path = Path(self.config.get("db_path") or "data/db/milvus.db")
+        self.client: MilvusClient | None = None
+
+    async def initialize(self) -> None:
+        try:
+            self.client = MilvusClient(str(self.db_path))
+        except Exception as exc:  # pragma: no cover - depends on local environment
+            raise RuntimeError(f"Failed to open Milvus Lite at {self.db_path}: {exc}") from exc
+
+    def _get_client(self) -> MilvusClient:
+        if not self.client:
+            raise RuntimeError("Milvus Lite client is not initialized")
+        return self.client
+
+    @staticmethod
+    def _extract_hit_payload(hit: Any) -> tuple[Any, Any, Dict[str, Any]]:
+        hit_id = getattr(hit, "id", None)
+        distance = getattr(hit, "distance", None)
+        payload: Dict[str, Any] = {}
+
+        raw: Dict[str, Any] | None = None
+        if hasattr(hit, "entity"):
+            raw_entity = getattr(hit, "entity")
+            if hasattr(raw_entity, "to_dict"):
+                raw = dict(raw_entity.to_dict())
+            else:
+                raw = dict(raw_entity)
+        elif isinstance(hit, dict):
+            raw = dict(hit)
+
+        if raw:
+            hit_id = hit_id or raw.get("id")
+            distance = distance if distance is not None else raw.get("distance")
+            inner = raw.get("entity")
+            if isinstance(inner, dict):
+                payload = dict(inner)
+            else:
+                payload = {k: v for k, v in raw.items() if k not in {"id", "distance", "entity"}}
+
+        payload.setdefault("path", payload.get("source_path"))
+        payload.setdefault("source_path", payload.get("path"))
+        return hit_id, distance, payload
+
+    @staticmethod
+    def _to_int(value: Any) -> int:
+        try:
+            return int(value)
+        except (TypeError, ValueError):
+            return 0
+
+    def ensure_collection(self, collection_name: str, vector: bool, dim: int) -> None:
+        client = self._get_client()
+        if client.has_collection(collection_name):
+            return
+        common_fields = [
+            FieldSchema(name="path", dtype=DataType.VARCHAR, max_length=512, is_primary=True, auto_id=False),
+            FieldSchema(name="source_path", dtype=DataType.VARCHAR, max_length=512, is_primary=False, auto_id=False),
+        ]
+
+        if vector:
+            vector_dim = dim if isinstance(dim, int) and dim > 0 else 0
+            if vector_dim <= 0:
+                vector_dim = 4096
+            fields = [
+                *common_fields,
+                FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=vector_dim),
+            ]
+            schema = CollectionSchema(fields, description="Vector collection", enable_dynamic_field=True)
+            client.create_collection(collection_name, schema=schema)
+            index_params = MilvusClient.prepare_index_params()
+            index_params.add_index(
+                field_name="embedding",
+                index_type="IVF_FLAT",
+                index_name="vector_index",
+                metric_type="COSINE",
+                params={"nlist": 64},
+            )
+            client.create_index(collection_name, index_params=index_params)
+        else:
+            schema = CollectionSchema(common_fields, description="Simple file index", enable_dynamic_field=True)
+            client.create_collection(collection_name, schema=schema)
+
+    def upsert_vector(self, collection_name: str, data: Dict[str, Any]) -> None:
+        payload = dict(data)
+        payload.setdefault("source_path", payload.get("path"))
+        payload.setdefault("vector_id", payload.get("path"))
+        self._get_client().upsert(collection_name, data=[payload])
+
+    def delete_vector(self, collection_name: str, path: str) -> None:
+        client = self._get_client()
+        escaped = path.replace('"', '\\"')
+        client.delete(collection_name, filter=f'source_path == "{escaped}"')
+
+    def search_vectors(self, collection_name: str, query_embedding, top_k: int):
+        search_params = {"metric_type": "COSINE"}
+        output_fields = [
+            "path",
+            "source_path",
+            "chunk_id",
+            "mime",
+            "text",
+            "start_offset",
+            "end_offset",
+            "type",
+            "name",
+        ]
+        raw_results = self._get_client().search(
+            collection_name,
+            data=[query_embedding],
+            anns_field="embedding",
+            search_params=search_params,
+            limit=top_k,
+            output_fields=output_fields,
+        )
+        formatted: List[List[Dict[str, Any]]] = []
+        for hits in raw_results:
+            bucket: List[Dict[str, Any]] = []
+            for hit in hits:
+                hit_id, distance, entity = self._extract_hit_payload(hit)
+                bucket.append({
+                    "id": hit_id,
+                    "distance": distance,
+                    "entity": entity,
+                })
+            formatted.append(bucket)
+        return formatted
+
+    def search_by_path(self, collection_name: str, query_path: str, top_k: int):
+        if query_path:
+            escaped = query_path.replace('"', '\\"')
+            filter_expr = f'source_path like \"%{escaped}%\"'
+        else:
+            filter_expr = "source_path like '%%'"
+        results = self._get_client().query(
+            collection_name,
+            filter=filter_expr,
+            limit=top_k,
+            output_fields=[
+                "path",
+                "source_path",
+                "chunk_id",
+                "mime",
+                "text",
+                "start_offset",
+                "end_offset",
+                "type",
+                "name",
+            ],
+        )
+        formatted = []
+        for row in results:
+            entity = dict(row)
+            entity.setdefault("path", entity.get("source_path"))
+            formatted.append({
+                "id": entity.get("path"),
+                "distance": 1.0,
+                "entity": entity,
+            })
+        return [formatted]
+
+    def get_all_stats(self) -> Dict[str, Any]:
+        client = self._get_client()
+        try:
+            collection_names = client.list_collections()
+        except Exception as exc:
+            raise RuntimeError(f"Failed to list collections: {exc}") from exc
+
+        collections: List[Dict[str, Any]] = []
+        total_vectors = 0
+        total_estimated_memory = 0
+
+        for name in collection_names:
+            try:
+                stats = client.get_collection_stats(name) or {}
+            except Exception:
+                stats = {}
+            row_count = self._to_int(stats.get("row_count"))
+            total_vectors += row_count
+
+            dimension: Optional[int] = None
+            is_vector_collection = False
+            try:
+                description = client.describe_collection(name)
+            except Exception:
+                description = None
+
+            if description:
+                for field in description.get("fields", []):
+                    if field.get("type") == DataType.FLOAT_VECTOR:
+                        params = field.get("params") or {}
+                        dimension = self._to_int(params.get("dim")) or 4096
+                        is_vector_collection = True
+                        break
+
+            estimated_memory = 0
+            if is_vector_collection and dimension:
+                estimated_memory = row_count * dimension * 4
+                total_estimated_memory += estimated_memory
+
+            indexes: List[Dict[str, Any]] = []
+            try:
+                index_names = client.list_indexes(name) or []
+            except Exception:
+                index_names = []
+
+            for index_name in index_names:
+                try:
+                    detail = client.describe_index(name) or {}
+                except Exception:
+                    detail = {}
+                indexes.append(
+                    {
+                        "index_name": index_name,
+                        "index_type": detail.get("index_type"),
+                        "metric_type": detail.get("metric_type"),
+                        "indexed_rows": self._to_int(detail.get("indexed_rows")),
+                        "pending_index_rows": self._to_int(detail.get("pending_index_rows")),
+                        "state": detail.get("state"),
+                    }
+                )
+
+            collections.append(
+                {
+                    "name": name,
+                    "row_count": row_count,
+                    "dimension": dimension if is_vector_collection else None,
+                    "estimated_memory_bytes": estimated_memory,
+                    "is_vector_collection": is_vector_collection,
+                    "indexes": indexes,
+                }
+            )
+
+        db_file_size = None
+        try:
+            if self.db_path.exists():
+                db_file_size = self.db_path.stat().st_size
+        except OSError:
+            db_file_size = None
+
+        return {
+            "collections": collections,
+            "collection_count": len(collections),
+            "total_vectors": total_vectors,
+            "estimated_total_memory_bytes": total_estimated_memory,
+            "db_file_size_bytes": db_file_size,
+        }
+
+    def clear_all_data(self) -> None:
+        client = self._get_client()
+        for collection_name in client.list_collections():
+            client.drop_collection(collection_name)

domain/ai/vector_providers/milvus_server.py

@@ -0,0 +1,276 @@
+from typing import Any, Dict, List, Optional
+
+from pymilvus import CollectionSchema, DataType, FieldSchema, MilvusClient
+
+from .base import BaseVectorProvider
+
+
+class MilvusServerProvider(BaseVectorProvider):
+    type = "milvus_server"
+    label = "Milvus Server"
+    description = "Remote Milvus instance accessed via URI."
+    enabled = True
+    config_schema: List[Dict[str, Any]] = [
+        {
+            "key": "uri",
+            "label": "Server URI",
+            "type": "text",
+            "required": True,
+            "placeholder": "http://localhost:19530",
+        },
+        {
+            "key": "token",
+            "label": "Token",
+            "type": "password",
+            "required": False,
+            "placeholder": "user:password",
+        },
+    ]
+
+    def __init__(self, config: Dict[str, Any] | None = None):
+        super().__init__(config)
+        self.client: MilvusClient | None = None
+
+    async def initialize(self) -> None:
+        uri = self.config.get("uri")
+        if not uri:
+            raise RuntimeError("Milvus Server URI is required")
+        try:
+            self.client = MilvusClient(uri=uri, token=self.config.get("token"))
+        except Exception as exc:  # pragma: no cover - depends on remote availability
+            raise RuntimeError(f"Failed to connect to Milvus Server {uri}: {exc}") from exc
+
+    def _get_client(self) -> MilvusClient:
+        if not self.client:
+            raise RuntimeError("Milvus Server client is not initialized")
+        return self.client
+
+    @staticmethod
+    def _extract_hit_payload(hit: Any) -> tuple[Any, Any, Dict[str, Any]]:
+        hit_id = getattr(hit, "id", None)
+        distance = getattr(hit, "distance", None)
+        payload: Dict[str, Any] = {}
+
+        raw: Dict[str, Any] | None = None
+        if hasattr(hit, "entity"):
+            raw_entity = getattr(hit, "entity")
+            if hasattr(raw_entity, "to_dict"):
+                raw = dict(raw_entity.to_dict())
+            else:
+                raw = dict(raw_entity)
+        elif isinstance(hit, dict):
+            raw = dict(hit)
+
+        if raw:
+            hit_id = hit_id or raw.get("id")
+            distance = distance if distance is not None else raw.get("distance")
+            inner = raw.get("entity")
+            if isinstance(inner, dict):
+                payload = dict(inner)
+            else:
+                payload = {k: v for k, v in raw.items() if k not in {"id", "distance", "entity"}}
+
+        payload.setdefault("path", payload.get("source_path"))
+        payload.setdefault("source_path", payload.get("path"))
+        return hit_id, distance, payload
+
+    @staticmethod
+    def _to_int(value: Any) -> int:
+        try:
+            return int(value)
+        except (TypeError, ValueError):
+            return 0
+
+    def ensure_collection(self, collection_name: str, vector: bool, dim: int) -> None:
+        client = self._get_client()
+        if client.has_collection(collection_name):
+            return
+        common_fields = [
+            FieldSchema(name="path", dtype=DataType.VARCHAR, max_length=512, is_primary=True, auto_id=False),
+            FieldSchema(name="source_path", dtype=DataType.VARCHAR, max_length=512, is_primary=False, auto_id=False),
+        ]
+        if vector:
+            vector_dim = dim if isinstance(dim, int) and dim > 0 else 0
+            if vector_dim <= 0:
+                vector_dim = 4096
+            fields = [
+                *common_fields,
+                FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=vector_dim),
+            ]
+            schema = CollectionSchema(fields, description="Vector collection", enable_dynamic_field=True)
+            client.create_collection(collection_name, schema=schema)
+            index_params = MilvusClient.prepare_index_params()
+            index_params.add_index(
+                field_name="embedding",
+                index_type="IVF_FLAT",
+                index_name="vector_index",
+                metric_type="COSINE",
+                params={"nlist": 64},
+            )
+            client.create_index(collection_name, index_params=index_params)
+        else:
+            schema = CollectionSchema(common_fields, description="Simple file index", enable_dynamic_field=True)
+            client.create_collection(collection_name, schema=schema)
+
+    def upsert_vector(self, collection_name: str, data: Dict[str, Any]) -> None:
+        payload = dict(data)
+        payload.setdefault("source_path", payload.get("path"))
+        payload.setdefault("vector_id", payload.get("path"))
+        self._get_client().upsert(collection_name, data=[payload])
+
+    def delete_vector(self, collection_name: str, path: str) -> None:
+        client = self._get_client()
+        escaped = path.replace('"', '\\"')
+        client.delete(collection_name, filter=f'source_path == "{escaped}"')
+
+    def search_vectors(self, collection_name: str, query_embedding, top_k: int):
+        search_params = {"metric_type": "COSINE"}
+        output_fields = [
+            "path",
+            "source_path",
+            "chunk_id",
+            "mime",
+            "text",
+            "start_offset",
+            "end_offset",
+            "type",
+            "name",
+        ]
+        raw_results = self._get_client().search(
+            collection_name,
+            data=[query_embedding],
+            anns_field="embedding",
+            search_params=search_params,
+            limit=top_k,
+            output_fields=output_fields,
+        )
+        formatted: List[List[Dict[str, Any]]] = []
+        for hits in raw_results:
+            bucket: List[Dict[str, Any]] = []
+            for hit in hits:
+                hit_id, distance, entity = self._extract_hit_payload(hit)
+                bucket.append({
+                    "id": hit_id,
+                    "distance": distance,
+                    "entity": entity,
+                })
+            formatted.append(bucket)
+        return formatted
+
+    def search_by_path(self, collection_name: str, query_path: str, top_k: int):
+        if query_path:
+            escaped = query_path.replace('"', '\\"')
+            filter_expr = f'source_path like \"%{escaped}%\"'
+        else:
+            filter_expr = "source_path like '%%'"
+        results = self._get_client().query(
+            collection_name,
+            filter=filter_expr,
+            limit=top_k,
+            output_fields=[
+                "path",
+                "source_path",
+                "chunk_id",
+                "mime",
+                "text",
+                "start_offset",
+                "end_offset",
+                "type",
+                "name",
+            ],
+        )
+        formatted = []
+        for row in results:
+            entity = dict(row)
+            entity.setdefault("path", entity.get("source_path"))
+            formatted.append({
+                "id": entity.get("path"),
+                "distance": 1.0,
+                "entity": entity,
+            })
+        return [formatted]
+
+    def get_all_stats(self) -> Dict[str, Any]:
+        client = self._get_client()
+        try:
+            collection_names = client.list_collections()
+        except Exception as exc:
+            raise RuntimeError(f"Failed to list collections: {exc}") from exc
+
+        collections: List[Dict[str, Any]] = []
+        total_vectors = 0
+        total_estimated_memory = 0
+
+        for name in collection_names:
+            try:
+                stats = client.get_collection_stats(name) or {}
+            except Exception:
+                stats = {}
+            row_count = self._to_int(stats.get("row_count"))
+            total_vectors += row_count
+
+            dimension: Optional[int] = None
+            is_vector_collection = False
+            try:
+                description = client.describe_collection(name)
+            except Exception:
+                description = None
+
+            if description:
+                for field in description.get("fields", []):
+                    if field.get("type") == DataType.FLOAT_VECTOR:
+                        params = field.get("params") or {}
+                        dimension = self._to_int(params.get("dim")) or 4096
+                        is_vector_collection = True
+                        break
+
+            estimated_memory = 0
+            if is_vector_collection and dimension:
+                estimated_memory = row_count * dimension * 4
+                total_estimated_memory += estimated_memory
+
+            indexes: List[Dict[str, Any]] = []
+            try:
+                index_names = client.list_indexes(name) or []
+            except Exception:
+                index_names = []
+
+            for index_name in index_names:
+                try:
+                    detail = client.describe_index(name) or {}
+                except Exception:
+                    detail = {}
+                indexes.append(
+                    {
+                        "index_name": index_name,
+                        "index_type": detail.get("index_type"),
+                        "metric_type": detail.get("metric_type"),
+                        "indexed_rows": self._to_int(detail.get("indexed_rows")),
+                        "pending_index_rows": self._to_int(detail.get("pending_index_rows")),
+                        "state": detail.get("state"),
+                    }
+                )
+
+            collections.append(
+                {
+                    "name": name,
+                    "row_count": row_count,
+                    "dimension": dimension if is_vector_collection else None,
+                    "estimated_memory_bytes": estimated_memory,
+                    "is_vector_collection": is_vector_collection,
+                    "indexes": indexes,
+                }
+            )
+
+        return {
+            "collections": collections,
+            "collection_count": len(collections),
+            "total_vectors": total_vectors,
+            "estimated_total_memory_bytes": total_estimated_memory,
+            "db_file_size_bytes": None,
+        }
+
+    def clear_all_data(self) -> None:
+        client = self._get_client()
+        for collection_name in client.list_collections():
+            client.drop_collection(collection_name)

domain/ai/vector_providers/qdrant.py

@@ -0,0 +1,271 @@
+from typing import Any, Dict, List, Optional, Sequence
+from uuid import NAMESPACE_URL, uuid5
+
+from qdrant_client import QdrantClient
+from qdrant_client.http import models as qmodels
+
+from .base import BaseVectorProvider
+
+
+class QdrantProvider(BaseVectorProvider):
+    type = "qdrant"
+    label = "Qdrant"
+    description = "Qdrant vector database (HTTP API)."
+    enabled = True
+    config_schema: List[Dict[str, Any]] = [
+        {
+            "key": "url",
+            "label": "Server URL",
+            "type": "text",
+            "required": True,
+            "placeholder": "http://localhost:6333",
+        },
+        {
+            "key": "api_key",
+            "label": "API Key",
+            "type": "password",
+            "required": False,
+        },
+    ]
+
+    def __init__(self, config: Dict[str, Any] | None = None):
+        super().__init__(config)
+        self.client: Optional[QdrantClient] = None
+
+    async def initialize(self) -> None:
+        url = (self.config.get("url") or "").strip()
+        if not url:
+            raise RuntimeError("Qdrant URL is required")
+
+        api_key = (self.config.get("api_key") or None) or None
+        try:
+            client = QdrantClient(url=url, api_key=api_key)
+            client.get_collections()
+            self.client = client
+        except Exception as exc:  # pragma: no cover - 依赖外部服务
+            raise RuntimeError(f"Failed to connect to Qdrant at {url}: {exc}") from exc
+
+    def _get_client(self) -> QdrantClient:
+        if not self.client:
+            raise RuntimeError("Qdrant client is not initialized")
+        return self.client
+
+    @staticmethod
+    def _vector_params(vector: bool, dim: int) -> qmodels.VectorParams:
+        size = dim if vector and isinstance(dim, int) and dim > 0 else 1
+        return qmodels.VectorParams(size=size, distance=qmodels.Distance.COSINE)
+
+    def _ensure_payload_indexes(self, client: QdrantClient, collection_name: str) -> None:
+        for field in ("path", "source_path"):
+            try:
+                client.create_payload_index(
+                    collection_name=collection_name,
+                    field_name=field,
+                    field_schema="keyword",
+                )
+            except Exception as exc:  # pragma: no cover - 依赖外部服务
+                message = str(exc).lower()
+                if "already exists" in message or "index exists" in message:
+                    continue
+                raise
+
+    def ensure_collection(self, collection_name: str, vector: bool, dim: int) -> None:
+        client = self._get_client()
+        try:
+            exists = client.collection_exists(collection_name)
+        except Exception as exc:  # pragma: no cover - 依赖外部服务
+            raise RuntimeError(f"Failed to check Qdrant collection '{collection_name}': {exc}") from exc
+
+        if exists:
+            try:
+                self._ensure_payload_indexes(client, collection_name)
+            except Exception:
+                pass
+            return
+
+        vectors_config = self._vector_params(vector, dim)
+        try:
+            client.create_collection(collection_name=collection_name, vectors_config=vectors_config)
+        except Exception as exc:  # pragma: no cover
+            if "already exists" in str(exc).lower():
+                try:
+                    self._ensure_payload_indexes(client, collection_name)
+                except Exception:
+                    pass
+                return
+            raise RuntimeError(f"Failed to create Qdrant collection '{collection_name}': {exc}") from exc
+
+        try:
+            self._ensure_payload_indexes(client, collection_name)
+        except Exception:
+            pass
+
+    @staticmethod
+    def _point_id(uid: str) -> str:
+        return str(uuid5(NAMESPACE_URL, uid))
+
+    def _prepare_point(self, data: Dict[str, Any]) -> qmodels.PointStruct:
+        uid = data.get("path")
+        if not uid:
+            raise ValueError("Qdrant upsert requires 'path' in data")
+
+        embedding = data.get("embedding")
+        if embedding is None:
+            vector = [0.0]
+        else:
+            vector = [float(x) for x in embedding]
+
+        payload = {k: v for k, v in data.items() if k != "embedding"}
+        payload.setdefault("vector_id", uid)
+        source_path = payload.get("source_path") or payload.get("path")
+        payload["path"] = source_path
+        return qmodels.PointStruct(id=self._point_id(str(uid)), vector=vector, payload=payload)
+
+    def upsert_vector(self, collection_name: str, data: Dict[str, Any]) -> None:
+        client = self._get_client()
+        point = self._prepare_point(data)
+        client.upsert(collection_name=collection_name, wait=True, points=[point])
+
+    def delete_vector(self, collection_name: str, path: str) -> None:
+        client = self._get_client()
+        condition = qmodels.FieldCondition(
+            key="path",
+            match=qmodels.MatchValue(value=path),
+        )
+        flt = qmodels.Filter(must=[condition])
+        selector = qmodels.FilterSelector(filter=flt)
+        client.delete(collection_name=collection_name, points_selector=selector, wait=True)
+
+    def _format_search_results(self, points: Sequence[qmodels.ScoredPoint]):
+        return [
+            {
+                "id": point.id,
+                "distance": point.score,
+                "entity": point.payload or {},
+            }
+            for point in points
+        ]
+
+    def search_vectors(self, collection_name: str, query_embedding, top_k: int):
+        client = self._get_client()
+        vector = [float(x) for x in query_embedding]
+        points = client.search(
+            collection_name=collection_name,
+            query_vector=vector,
+            limit=top_k,
+            with_payload=True,
+        )
+        return [self._format_search_results(points)]
+
+    def search_by_path(self, collection_name: str, query_path: str, top_k: int):
+        client = self._get_client()
+        results: List[Dict[str, Any]] = []
+        offset: Optional[str | int] = None
+        remaining = max(top_k, 1)
+
+        while len(results) < top_k:
+            batch_size = min(max(remaining * 2, 10), 200)
+            records, next_offset = client.scroll(
+                collection_name=collection_name,
+                limit=batch_size,
+                offset=offset,
+                with_payload=True,
+            )
+            if not records:
+                break
+
+            for record in records:
+                payload = record.payload or {}
+                path = payload.get("path")
+                if query_path and path and query_path not in path:
+                    continue
+                results.append({"id": record.id, "distance": 1.0, "entity": payload})
+                if len(results) >= top_k:
+                    break
+
+            if next_offset is None or len(results) >= top_k:
+                break
+            offset = next_offset
+            remaining = top_k - len(results)
+
+        return [results]
+
+    def _extract_vector_config(self, vectors) -> Optional[qmodels.VectorParams]:
+        if isinstance(vectors, qmodels.VectorParams):
+            return vectors
+        if isinstance(vectors, dict):
+            for value in vectors.values():
+                if isinstance(value, qmodels.VectorParams):
+                    return value
+        return None
+
+    def get_all_stats(self) -> Dict[str, Any]:
+        client = self._get_client()
+        try:
+            response = client.get_collections()
+        except Exception as exc:  # pragma: no cover
+            raise RuntimeError(f"Failed to list Qdrant collections: {exc}") from exc
+
+        collections: List[Dict[str, Any]] = []
+        total_vectors = 0
+        total_estimated_memory = 0
+
+        for description in response.collections or []:
+            name = description.name
+            try:
+                info = client.get_collection(name)
+            except Exception:
+                continue
+
+            row_count = int(info.points_count or 0)
+            total_vectors += row_count
+
+            vector_params = self._extract_vector_config(info.config.params.vectors if info.config and info.config.params else None)
+            dimension = int(vector_params.size) if vector_params and vector_params.size else None
+            estimated_memory = row_count * dimension * 4 if dimension else 0
+            total_estimated_memory += estimated_memory
+            distance = str(vector_params.distance) if vector_params and vector_params.distance else None
+
+            indexed_rows = int(info.indexed_vectors_count or 0)
+            pending_rows = max(row_count - indexed_rows, 0)
+
+            collections.append(
+                {
+                    "name": name,
+                    "row_count": row_count,
+                    "dimension": dimension,
+                    "estimated_memory_bytes": estimated_memory,
+                    "is_vector_collection": dimension is not None and dimension > 1,
+                    "indexes": [
+                        {
+                            "index_name": "hnsw",
+                            "index_type": "HNSW",
+                            "metric_type": distance,
+                            "indexed_rows": indexed_rows,
+                            "pending_index_rows": pending_rows,
+                            "state": info.status,
+                        }
+                    ],
+                }
+            )
+
+        return {
+            "collections": collections,
+            "collection_count": len(collections),
+            "total_vectors": total_vectors,
+            "estimated_total_memory_bytes": total_estimated_memory,
+            "db_file_size_bytes": None,
+        }
+
+    def clear_all_data(self) -> None:
+        client = self._get_client()
+        try:
+            response = client.get_collections()
+        except Exception as exc:  # pragma: no cover
+            raise RuntimeError(f"Failed to list Qdrant collections: {exc}") from exc
+
+        for description in response.collections or []:
+            try:
+                client.delete_collection(description.name)
+            except Exception:
+                continue