feat(models): support mainstream model protocols

2026-07-17 20:58:27 +08:00 · 2026-07-17 20:58:27 +08:00 · cc79d3a625
commit cc79d3a625
parent 457dfed252
15 changed files with 622 additions and 51 deletions
--- a/backend/agenteval/model_gateway.py
+++ b/backend/agenteval/model_gateway.py
@ -1,11 +1,11 @@
-"""Shared OpenAI-compatible model transport used by evaluation features."""
+"""Shared multi-protocol model transport used by evaluation features."""

 from typing import Any

 import httpx

+from agenteval.model_protocols import ProtocolAdapterError, get_protocol_adapter
 from agenteval.services.model_configs import ModelRuntimeConfig
-from agenteval.utils.llm import extract_content_from_llm_response


 class ModelGatewayError(RuntimeError):
@ -13,22 +13,17 @@ class ModelGatewayError(RuntimeError):


 class ModelGateway:
-    def __init__(self, timeout: float = 60.0):
+    def __init__(self, timeout: float = 60.0, transport: httpx.AsyncBaseTransport | None = None):
        self.timeout = timeout
-
-    @staticmethod
-    def _headers(config: ModelRuntimeConfig) -> dict[str, str]:
-        headers = {"Content-Type": "application/json"}
-        if config.api_key:
-            headers["Authorization"] = f"Bearer {config.api_key}"
-        return headers
+        self.transport = transport

    async def _post(self, config: ModelRuntimeConfig, payload: dict[str, Any]) -> dict[str, Any]:
+        adapter = self._adapter(config)
        try:
-            async with httpx.AsyncClient(timeout=self.timeout) as client:
+            async with httpx.AsyncClient(timeout=self.timeout, transport=self.transport) as client:
                response = await client.post(
                    config.endpoint_url,
-                    headers=self._headers(config),
+                    headers=adapter.headers(config.api_key),
                    json=payload,
                )
            response.raise_for_status()
@ -44,42 +39,41 @@ class ModelGateway:
        except Exception as exc:
            raise ModelGatewayError(f"模型接口调用失败: {exc}") from exc

+    @staticmethod
+    def _adapter(config: ModelRuntimeConfig):
+        try:
+            return get_protocol_adapter(config.provider)
+        except ProtocolAdapterError as exc:
+            raise ModelGatewayError(str(exc)) from exc
+
    async def chat(
        self,
        config: ModelRuntimeConfig,
        messages: list[dict[str, str]],
        temperature: float = 0.2,
    ) -> str:
-        payload = {
-            "model": config.model_name,
-            "messages": messages,
-            "temperature": temperature,
-        }
-        content = extract_content_from_llm_response(await self._post(config, payload))
-        if not content:
-            raise ModelGatewayError("模型接口返回内容为空")
-        return content
+        adapter = self._adapter(config)
+        try:
+            payload = adapter.chat_payload(config.model_name, messages, temperature)
+            return adapter.parse_chat(await self._post(config, payload))
+        except ProtocolAdapterError as exc:
+            raise ModelGatewayError(str(exc)) from exc

    async def embed(self, config: ModelRuntimeConfig, inputs: str | list[str]) -> list[list[float]]:
-        data = await self._post(config, {"model": config.model_name, "input": inputs})
+        adapter = self._adapter(config)
        try:
-            rows = sorted(data["data"], key=lambda item: item.get("index", 0))
-            return [row["embedding"] for row in rows]
-        except (KeyError, TypeError) as exc:
-            raise ModelGatewayError("Embedding 接口返回格式不正确") from exc
+            payload = adapter.embedding_payload(config.model_name, inputs)
+            return adapter.parse_embeddings(await self._post(config, payload))
+        except ProtocolAdapterError as exc:
+            raise ModelGatewayError(str(exc)) from exc

    async def moderate(self, config: ModelRuntimeConfig, text: str) -> dict[str, Any]:
-        payload: dict[str, Any] = {"input": text}
-        if config.model_name:
-            payload["model"] = config.model_name
-        data = await self._post(config, payload)
+        adapter = self._adapter(config)
        try:
-            result = data["results"][0]
-        except (KeyError, IndexError, TypeError) as exc:
-            raise ModelGatewayError("Moderation 接口返回格式不正确") from exc
-        if not isinstance(result, dict):
-            raise ModelGatewayError("Moderation 接口返回格式不正确")
-        return result
+            payload = adapter.moderation_payload(config.model_name, text)
+            return adapter.parse_moderation(await self._post(config, payload))
+        except ProtocolAdapterError as exc:
+            raise ModelGatewayError(str(exc)) from exc

    async def test_connection(self, config: ModelRuntimeConfig) -> str:
        if config.capability.value == "chat":
--- a/backend/agenteval/model_protocols/init.py
+++ b/backend/agenteval/model_protocols/init.py
@ -0,0 +1,39 @@
+"""Registry of supported external model protocols."""
+
+from agenteval.models import ModelCapability, ModelProtocol
+
+from .anthropic import AnthropicAdapter
+from .base import ModelProtocolAdapter, ProtocolAdapterError
+from .dashscope import DashScopeAdapter
+from .gemini import GoogleGeminiAdapter
+from .openai import OpenAICompatibleAdapter
+
+_ADAPTERS: dict[ModelProtocol, ModelProtocolAdapter] = {
+    adapter.protocol: adapter
+    for adapter in (
+        OpenAICompatibleAdapter(),
+        AnthropicAdapter(),
+        GoogleGeminiAdapter(),
+        DashScopeAdapter(),
+    )
+}
+
+
+def get_protocol_adapter(provider: str | ModelProtocol) -> ModelProtocolAdapter:
+    try:
+        protocol = ModelProtocol(provider)
+        return _ADAPTERS[protocol]
+    except (ValueError, KeyError) as exc:
+        raise ProtocolAdapterError(f"不支持的模型协议: {provider}") from exc
+
+
+def get_protocol_capabilities(provider: str | ModelProtocol) -> frozenset[ModelCapability]:
+    return get_protocol_adapter(provider).supported_capabilities
+
+
+__all__ = [
+    "ModelProtocolAdapter",
+    "ProtocolAdapterError",
+    "get_protocol_adapter",
+    "get_protocol_capabilities",
+]
--- a/backend/agenteval/model_protocols/anthropic.py
+++ b/backend/agenteval/model_protocols/anthropic.py
@ -0,0 +1,57 @@
+"""Anthropic Messages API protocol."""
+
+from typing import Any
+
+from agenteval.models import ModelCapability, ModelProtocol
+
+from .base import ModelProtocolAdapter, ProtocolAdapterError
+
+
+class AnthropicAdapter(ModelProtocolAdapter):
+    protocol = ModelProtocol.ANTHROPIC
+    supported_capabilities = frozenset({ModelCapability.CHAT})
+
+    def headers(self, api_key: str | None) -> dict[str, str]:
+        headers = super().headers(api_key)
+        headers["anthropic-version"] = "2023-06-01"
+        if api_key:
+            headers["x-api-key"] = api_key
+        return headers
+
+    def chat_payload(
+        self,
+        model_name: str | None,
+        messages: list[dict[str, str]],
+        temperature: float,
+    ) -> dict[str, Any]:
+        system_parts: list[str] = []
+        api_messages: list[dict[str, str]] = []
+        for message in messages:
+            role = message.get("role", "user")
+            content = message.get("content", "")
+            if role in {"system", "developer"}:
+                system_parts.append(content)
+            else:
+                api_messages.append({"role": "assistant" if role == "assistant" else "user", "content": content})
+        payload: dict[str, Any] = {
+            "model": self.require_model(model_name),
+            "max_tokens": 1024,
+            "messages": api_messages,
+            "temperature": temperature,
+        }
+        if system_parts:
+            payload["system"] = "\n\n".join(system_parts)
+        return payload
+
+    def parse_chat(self, data: dict[str, Any]) -> str:
+        blocks = data.get("content")
+        if not isinstance(blocks, list):
+            raise ProtocolAdapterError("Anthropic 接口返回格式不正确")
+        content = "\n".join(
+            str(block.get("text", ""))
+            for block in blocks
+            if isinstance(block, dict) and block.get("type") == "text" and block.get("text")
+        )
+        if not content:
+            raise ProtocolAdapterError("模型接口返回内容为空")
+        return content
--- a/backend/agenteval/model_protocols/base.py
+++ b/backend/agenteval/model_protocols/base.py
@ -0,0 +1,49 @@
+"""Protocol adapter contract for external model APIs."""
+
+from typing import Any
+
+from agenteval.models import ModelCapability, ModelProtocol
+
+
+class ProtocolAdapterError(ValueError):
+    pass
+
+
+class ModelProtocolAdapter:
+    protocol: ModelProtocol
+    supported_capabilities: frozenset[ModelCapability] = frozenset()
+
+    def headers(self, api_key: str | None) -> dict[str, str]:
+        return {"Content-Type": "application/json"}
+
+    def chat_payload(
+        self,
+        model_name: str | None,
+        messages: list[dict[str, str]],
+        temperature: float,
+    ) -> dict[str, Any]:
+        self._unsupported(ModelCapability.CHAT)
+
+    def parse_chat(self, data: dict[str, Any]) -> str:
+        self._unsupported(ModelCapability.CHAT)
+
+    def embedding_payload(self, model_name: str | None, inputs: str | list[str]) -> dict[str, Any]:
+        self._unsupported(ModelCapability.EMBEDDING)
+
+    def parse_embeddings(self, data: dict[str, Any]) -> list[list[float]]:
+        self._unsupported(ModelCapability.EMBEDDING)
+
+    def moderation_payload(self, model_name: str | None, text: str) -> dict[str, Any]:
+        self._unsupported(ModelCapability.MODERATION)
+
+    def parse_moderation(self, data: dict[str, Any]) -> dict[str, Any]:
+        self._unsupported(ModelCapability.MODERATION)
+
+    def _unsupported(self, capability: ModelCapability) -> None:
+        raise ProtocolAdapterError(f"{self.protocol.value} 协议不支持 {capability.value} 能力")
+
+    @staticmethod
+    def require_model(model_name: str | None) -> str:
+        if not model_name:
+            raise ProtocolAdapterError("模型名称不能为空")
+        return model_name
--- a/backend/agenteval/model_protocols/dashscope.py
+++ b/backend/agenteval/model_protocols/dashscope.py
@ -0,0 +1,45 @@
+"""Alibaba Cloud Model Studio native DashScope protocol."""
+
+from typing import Any
+
+from agenteval.models import ModelCapability, ModelProtocol
+
+from .base import ModelProtocolAdapter, ProtocolAdapterError
+
+
+class DashScopeAdapter(ModelProtocolAdapter):
+    protocol = ModelProtocol.DASHSCOPE
+    supported_capabilities = frozenset({ModelCapability.CHAT})
+
+    def headers(self, api_key: str | None) -> dict[str, str]:
+        headers = super().headers(api_key)
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+        return headers
+
+    def chat_payload(
+        self,
+        model_name: str | None,
+        messages: list[dict[str, str]],
+        temperature: float,
+    ) -> dict[str, Any]:
+        return {
+            "model": self.require_model(model_name),
+            "input": {"messages": messages},
+            "parameters": {"result_format": "message", "temperature": temperature},
+        }
+
+    def parse_chat(self, data: dict[str, Any]) -> str:
+        try:
+            content = data["output"]["choices"][0]["message"]["content"]
+        except (KeyError, IndexError, TypeError) as exc:
+            raise ProtocolAdapterError("DashScope 接口返回格式不正确") from exc
+        if isinstance(content, list):
+            content = "\n".join(
+                str(block.get("text") or block.get("content") or "")
+                for block in content
+                if isinstance(block, dict)
+            )
+        if not content:
+            raise ProtocolAdapterError("模型接口返回内容为空")
+        return str(content)
--- a/backend/agenteval/model_protocols/gemini.py
+++ b/backend/agenteval/model_protocols/gemini.py
@ -0,0 +1,56 @@
+"""Google Gemini generateContent REST protocol."""
+
+from typing import Any
+
+from agenteval.models import ModelCapability, ModelProtocol
+
+from .base import ModelProtocolAdapter, ProtocolAdapterError
+
+
+class GoogleGeminiAdapter(ModelProtocolAdapter):
+    protocol = ModelProtocol.GOOGLE_GEMINI
+    supported_capabilities = frozenset({ModelCapability.CHAT})
+
+    def headers(self, api_key: str | None) -> dict[str, str]:
+        headers = super().headers(api_key)
+        if api_key:
+            headers["x-goog-api-key"] = api_key
+        return headers
+
+    def chat_payload(
+        self,
+        model_name: str | None,
+        messages: list[dict[str, str]],
+        temperature: float,
+    ) -> dict[str, Any]:
+        self.require_model(model_name)
+        system_parts: list[dict[str, str]] = []
+        contents: list[dict[str, Any]] = []
+        for message in messages:
+            role = message.get("role", "user")
+            part = {"text": message.get("content", "")}
+            if role in {"system", "developer"}:
+                system_parts.append(part)
+            else:
+                contents.append({"role": "model" if role == "assistant" else "user", "parts": [part]})
+        payload: dict[str, Any] = {
+            "contents": contents,
+            "generationConfig": {"temperature": temperature},
+        }
+        if system_parts:
+            payload["systemInstruction"] = {"parts": system_parts}
+        return payload
+
+    def parse_chat(self, data: dict[str, Any]) -> str:
+        try:
+            parts = data["candidates"][0]["content"]["parts"]
+        except (KeyError, IndexError, TypeError) as exc:
+            raise ProtocolAdapterError("Gemini 接口返回格式不正确") from exc
+        content = "\n".join(
+            str(part.get("text", ""))
+            for part in parts
+            if isinstance(part, dict) and part.get("text")
+        )
+        if not content:
+            raise ProtocolAdapterError("模型接口返回内容为空")
+        return content
--- a/backend/agenteval/model_protocols/openai.py
+++ b/backend/agenteval/model_protocols/openai.py
@ -0,0 +1,61 @@
+"""OpenAI-compatible chat, embedding, and moderation protocol."""
+
+from typing import Any
+
+from agenteval.models import ModelCapability, ModelProtocol
+from agenteval.utils.llm import extract_content_from_llm_response
+
+from .base import ModelProtocolAdapter, ProtocolAdapterError
+
+
+class OpenAICompatibleAdapter(ModelProtocolAdapter):
+    protocol = ModelProtocol.OPENAI_COMPATIBLE
+    supported_capabilities = frozenset(ModelCapability)
+
+    def headers(self, api_key: str | None) -> dict[str, str]:
+        headers = super().headers(api_key)
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+        return headers
+
+    def chat_payload(
+        self,
+        model_name: str | None,
+        messages: list[dict[str, str]],
+        temperature: float,
+    ) -> dict[str, Any]:
+        return {"model": self.require_model(model_name), "messages": messages, "temperature": temperature}
+
+    def parse_chat(self, data: dict[str, Any]) -> str:
+        content = extract_content_from_llm_response(data)
+        if not content:
+            raise ProtocolAdapterError("模型接口返回内容为空")
+        return content
+
+    def embedding_payload(self, model_name: str | None, inputs: str | list[str]) -> dict[str, Any]:
+        return {"model": self.require_model(model_name), "input": inputs}
+
+    def parse_embeddings(self, data: dict[str, Any]) -> list[list[float]]:
+        try:
+            rows = sorted(data["data"], key=lambda item: item.get("index", 0))
+            vectors = [row["embedding"] for row in rows]
+        except (KeyError, TypeError) as exc:
+            raise ProtocolAdapterError("Embedding 接口返回格式不正确") from exc
+        if not all(isinstance(vector, list) for vector in vectors):
+            raise ProtocolAdapterError("Embedding 接口返回格式不正确")
+        return vectors
+
+    def moderation_payload(self, model_name: str | None, text: str) -> dict[str, Any]:
+        payload: dict[str, Any] = {"input": text}
+        if model_name:
+            payload["model"] = model_name
+        return payload
+
+    def parse_moderation(self, data: dict[str, Any]) -> dict[str, Any]:
+        try:
+            result = data["results"][0]
+        except (KeyError, IndexError, TypeError) as exc:
+            raise ProtocolAdapterError("Moderation 接口返回格式不正确") from exc
+        if not isinstance(result, dict):
+            raise ProtocolAdapterError("Moderation 接口返回格式不正确")
+        return result
--- a/backend/agenteval/models.py
+++ b/backend/agenteval/models.py
@ -36,6 +36,13 @@ class ModelCapability(str, Enum):
    MODERATION = "moderation"


+class ModelProtocol(str, Enum):
+    OPENAI_COMPATIBLE = "openai_compatible"
+    ANTHROPIC = "anthropic"
+    GOOGLE_GEMINI = "google_gemini"
+    DASHSCOPE = "dashscope"
+
+
 class ModelPurpose(str, Enum):
    GENERATOR = "generator"
    JUDGE = "judge"
--- a/backend/agenteval/services/model_configs.py
+++ b/backend/agenteval/services/model_configs.py
@ -8,6 +8,7 @@ from cryptography.fernet import Fernet, InvalidToken
 from sqlmodel import Session

 from agenteval.config import get_settings
+from agenteval.model_protocols import ProtocolAdapterError, get_protocol_adapter
 from agenteval.models import ModelCapability, ModelPurpose
 from agenteval.storage.db import ModelConfigDB, ScenarioDB
 from agenteval.storage.model_config_repository import ModelConfigRepository
@ -94,11 +95,17 @@ class ModelConfigService:
        self.cipher = cipher or SecretCipher()

    @staticmethod
-    def validate_fields(capability: str, endpoint_url: str, model_name: str | None) -> None:
+    def validate_fields(provider: str, capability: str, endpoint_url: str, model_name: str | None) -> None:
        try:
            capability_value = ModelCapability(capability)
        except ValueError as exc:
            raise ModelConfigError(f"不支持的模型能力: {capability}") from exc
+        try:
+            adapter = get_protocol_adapter(provider)
+        except ProtocolAdapterError as exc:
+            raise ModelConfigError(str(exc)) from exc
+        if capability_value not in adapter.supported_capabilities:
+            raise ModelConfigError(f"{provider} 协议不支持 {capability_value.value} 能力")
        parsed = urlparse(endpoint_url)
        if parsed.scheme not in {"http", "https"} or not parsed.netloc:
            raise ModelConfigError("Endpoint 必须是有效的 HTTP 或 HTTPS URL")
@ -109,8 +116,10 @@ class ModelConfigService:
    def validate_common(name: str, provider: str, enabled: bool, is_default: bool) -> None:
        if not name.strip():
            raise ModelConfigError("模型配置名称不能为空")
-        if provider != "openai_compatible":
-            raise ModelConfigError(f"不支持的模型协议: {provider}")
+        try:
+            get_protocol_adapter(provider)
+        except ProtocolAdapterError as exc:
+            raise ModelConfigError(str(exc)) from exc
        if is_default and not enabled:
            raise ModelConfigError("停用的模型配置不能设为默认")

@ -129,7 +138,7 @@ class ModelConfigService:
    ) -> ModelConfigDB:
        name = name.strip()
        self.validate_common(name, provider, enabled, is_default)
-        self.validate_fields(capability, endpoint_url.strip(), model_name)
+        self.validate_fields(provider, capability, endpoint_url.strip(), model_name)
        if self.repo.get_by_name(name):
            raise ModelConfigError("模型配置名称已存在")
        config = ModelConfigDB(
@ -163,7 +172,7 @@ class ModelConfigService:
        config = self.require(config_id)
        name = name.strip()
        self.validate_common(name, provider, enabled, is_default)
-        self.validate_fields(capability, endpoint_url.strip(), model_name)
+        self.validate_fields(provider, capability, endpoint_url.strip(), model_name)
        references = self.repo.list_references(config_id)
        if references and not enabled:
            raise ModelConfigError("模型配置正在被场景引用，不能停用")
--- a/backend/agenteval/web/model_config_schemas.py
+++ b/backend/agenteval/web/model_config_schemas.py
@ -4,13 +4,13 @@ from datetime import datetime

 from pydantic import BaseModel, Field

-from agenteval.models import ModelCapability
+from agenteval.models import ModelCapability, ModelProtocol
 from agenteval.storage.db import ModelConfigDB, iso_utc


 class ModelConfigCreate(BaseModel):
    name: str = Field(min_length=1, max_length=100)
-    provider: str = "openai_compatible"
+    provider: ModelProtocol = ModelProtocol.OPENAI_COMPATIBLE
    capability: ModelCapability
    endpoint_url: str
    model_name: str | None = None
@ -27,7 +27,7 @@ class ModelConfigUpdate(ModelConfigCreate):
 class ModelConfigResponse(BaseModel):
    id: str
    name: str
-    provider: str
+    provider: ModelProtocol
    capability: ModelCapability
    endpoint_url: str
    model_name: str | None
@ -43,7 +43,7 @@ class ModelConfigResponse(BaseModel):
        return cls(
            id=config.id or "",
            name=config.name,
-            provider=config.provider,
+            provider=ModelProtocol(config.provider),
            capability=ModelCapability(config.capability),
            endpoint_url=config.endpoint_url,
            model_name=config.model_name,
--- a/frontend/web/src/api.ts
+++ b/frontend/web/src/api.ts
@ -41,11 +41,12 @@ export interface Scenario {
 }

 export type ModelCapability = 'chat' | 'embedding' | 'moderation'
+export type ModelProtocol = 'openai_compatible' | 'anthropic' | 'google_gemini' | 'dashscope'

 export interface ModelConfig {
  id: string
  name: string
-  provider: string
+  provider: ModelProtocol
  capability: ModelCapability
  endpoint_url: string
  model_name: string | null
@ -59,7 +60,7 @@ export interface ModelConfig {

 export interface ModelConfigPayload {
  name: string
-  provider: string
+  provider: ModelProtocol
  capability: ModelCapability
  endpoint_url: string
  model_name?: string | null
--- a/frontend/web/src/pages/ModelConfigs.tsx
+++ b/frontend/web/src/pages/ModelConfigs.tsx
@ -6,7 +6,7 @@ import {
  ApiOutlined, CheckCircleOutlined, DeleteOutlined, EditOutlined, PlusOutlined, ReloadOutlined,
 } from '@ant-design/icons'
 import {
-  modelConfigsApi, type ModelCapability, type ModelConfig, type ModelConfigPayload,
+  modelConfigsApi, type ModelCapability, type ModelConfig, type ModelConfigPayload, type ModelProtocol,
 } from '../api'
 import PageWrapper from '../components/PageWrapper'
 import { formatDateTime } from '../utils/date'
@ -24,6 +24,33 @@ const capabilityColors: Record<ModelCapability, string> = {
  moderation: 'orange',
 }

+const protocolOptions: Record<ModelProtocol, {
+  label: string
+  capabilities: ModelCapability[]
+  endpointPlaceholder: string
+}> = {
+  openai_compatible: {
+    label: 'OpenAI 兼容',
+    capabilities: ['chat', 'embedding', 'moderation'],
+    endpointPlaceholder: 'https://api.example.com/v1/chat/completions',
+  },
+  anthropic: {
+    label: 'Anthropic Messages',
+    capabilities: ['chat'],
+    endpointPlaceholder: 'https://api.anthropic.com/v1/messages',
+  },
+  google_gemini: {
+    label: 'Google Gemini',
+    capabilities: ['chat'],
+    endpointPlaceholder: 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent',
+  },
+  dashscope: {
+    label: 'DashScope 原生',
+    capabilities: ['chat'],
+    endpointPlaceholder: 'https://dashscope.aliyuncs.com/api/v1/services/aigc/text-generation/generation',
+  },
+}
+
 export default function ModelConfigsPage() {
  const [configs, setConfigs] = useState<ModelConfig[]>([])
  const [loading, setLoading] = useState(false)
@ -33,6 +60,8 @@ export default function ModelConfigsPage() {
  const [capability, setCapability] = useState<ModelCapability | undefined>()
  const [enabled, setEnabled] = useState<boolean | undefined>()
  const [form] = Form.useForm<ModelConfigPayload>()
+  const selectedProtocol = Form.useWatch('provider', form) || 'openai_compatible'
+  const selectedProtocolMeta = protocolOptions[selectedProtocol]

  const load = async () => {
    setLoading(true)
@ -72,6 +101,13 @@ export default function ModelConfigsPage() {
    setDrawerOpen(true)
  }

+  const changeProtocol = (provider: ModelProtocol) => {
+    const supportedCapabilities = protocolOptions[provider].capabilities
+    if (!supportedCapabilities.includes(form.getFieldValue('capability'))) {
+      form.setFieldValue('capability', supportedCapabilities[0])
+    }
+  }
+
  const submit = async () => {
    const values = await form.validateFields()
    const payload: ModelConfigPayload = {
@ -120,6 +156,10 @@ export default function ModelConfigsPage() {
        </Space>
      ),
    },
+    {
+      title: '协议', dataIndex: 'provider', key: 'provider', width: 150,
+      render: (value: ModelProtocol) => <Tag>{protocolOptions[value]?.label || value}</Tag>,
+    },
    {
      title: '能力', dataIndex: 'capability', key: 'capability', width: 110,
      render: (value: ModelCapability) => <Tag color={capabilityColors[value]}>{capabilityLabels[value]}</Tag>,
@ -219,14 +259,19 @@ export default function ModelConfigsPage() {
          </Form.Item>
          <div style={{ display: 'grid', gridTemplateColumns: '1fr 1fr', gap: 12 }}>
            <Form.Item name="provider" label="协议" rules={[{ required: true }]}>
-              <Select options={[{ value: 'openai_compatible', label: 'OpenAI Compatible' }]} />
+              <Select
+                onChange={changeProtocol}
+                options={Object.entries(protocolOptions).map(([value, option]) => ({ value, label: option.label }))}
+              />
            </Form.Item>
            <Form.Item name="capability" label="能力" rules={[{ required: true }]}>
-              <Select options={Object.entries(capabilityLabels).map(([value, label]) => ({ value, label }))} />
+              <Select
+                options={selectedProtocolMeta.capabilities.map((value) => ({ value, label: capabilityLabels[value] }))}
+              />
            </Form.Item>
          </div>
          <Form.Item name="endpoint_url" label="完整 Endpoint URL" rules={[{ required: true, message: '请输入 Endpoint URL' }, { type: 'url', message: '请输入有效 URL' }]}>
-            <Input placeholder="https://api.example.com/v1/chat/completions" />
+            <Input placeholder={selectedProtocolMeta.endpointPlaceholder} />
          </Form.Item>
          <Form.Item shouldUpdate noStyle>
            {() => form.getFieldValue('capability') !== 'moderation' && (
--- a/tests/integration/test_model_configs_api.py
+++ b/tests/integration/test_model_configs_api.py
@ -126,3 +126,35 @@ def test_scenario_rejects_capability_mismatch(model_client):
    )
    assert response.status_code == 400
    assert "不能用于 chat" in response.json()["detail"]
+
+
+def test_model_config_api_supports_mainstream_protocols_and_validates_capability(model_client):
+    client, _ = model_client
+
+    for provider in ("anthropic", "google_gemini", "dashscope"):
+        response = client.post(
+            "/api/model-configs",
+            json=_payload(
+                name=f"{provider}-chat",
+                provider=provider,
+                endpoint_url=f"https://models.example.com/{provider}",
+                api_key=None,
+                is_default=False,
+            ),
+        )
+        assert response.status_code == 200
+        assert response.json()["provider"] == provider
+
+    mismatch = client.post(
+        "/api/model-configs",
+        json=_payload(
+            name="anthropic-embedding",
+            provider="anthropic",
+            capability="embedding",
+            endpoint_url="https://models.example.com/anthropic/embeddings",
+            api_key=None,
+            is_default=False,
+        ),
+    )
+    assert mismatch.status_code == 400
+    assert "不支持 embedding 能力" in mismatch.json()["detail"]
--- a/tests/unit/test_model_configs.py
+++ b/tests/unit/test_model_configs.py
@ -108,3 +108,33 @@ def test_failed_binding_does_not_create_scenario(db_session):
    with pytest.raises(ModelConfigError):
        ScenarioRepository(db_session).create(scenario)
    assert ScenarioRepository(db_session).get(scenario.id) is None
+
+
+@pytest.mark.parametrize("provider", ["anthropic", "google_gemini", "dashscope"])
+def test_chat_only_protocols_accept_chat_and_reject_other_capabilities(db_session, provider):
+    service = _service(db_session)
+    created = service.create(
+        name=f"{provider}-chat",
+        provider=provider,
+        capability="chat",
+        endpoint_url=f"https://models.example.com/{provider}",
+        model_name="chat-model",
+        api_key=None,
+        enabled=True,
+        is_default=False,
+        description="",
+    )
+    assert created.provider == provider
+
+    with pytest.raises(ModelConfigError, match="不支持 embedding 能力"):
+        service.create(
+            name=f"{provider}-embedding",
+            provider=provider,
+            capability="embedding",
+            endpoint_url=f"https://models.example.com/{provider}/embeddings",
+            model_name="embedding-model",
+            api_key=None,
+            enabled=True,
+            is_default=False,
+            description="",
+        )
--- a/tests/unit/test_model_gateway.py
+++ b/tests/unit/test_model_gateway.py
@ -0,0 +1,146 @@
+"""Request and response contract tests for supported model protocols."""
+
+import json
+from datetime import datetime
+
+import httpx
+import pytest
+from agenteval.model_gateway import ModelGateway, ModelGatewayError
+from agenteval.models import ModelCapability
+from agenteval.services.model_configs import ModelRuntimeConfig
+
+
+def _config(provider: str, capability: ModelCapability = ModelCapability.CHAT) -> ModelRuntimeConfig:
+    return ModelRuntimeConfig(
+        id=f"{provider}-config",
+        name=provider,
+        provider=provider,
+        capability=capability,
+        endpoint_url=f"https://models.example.com/{provider}",
+        model_name="test-model",
+        api_key="test-key",
+        updated_at=datetime(2026, 7, 17),
+    )
+
+
+def _gateway(handler) -> ModelGateway:
+    return ModelGateway(transport=httpx.MockTransport(handler))
+
+
+async def test_openai_compatible_chat_embedding_and_moderation_contracts():
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        payload = json.loads(request.content)
+        if "messages" in payload:
+            return httpx.Response(200, json={"choices": [{"message": {"content": "chat reply"}}]})
+        if payload.get("input") == ["first", "second"]:
+            return httpx.Response(
+                200,
+                json={"data": [{"index": 1, "embedding": [2.0]}, {"index": 0, "embedding": [1.0]}]},
+            )
+        return httpx.Response(200, json={"results": [{"flagged": False}]})
+
+    gateway = _gateway(handler)
+    config = _config("openai_compatible")
+
+    assert await gateway.chat(config, [{"role": "user", "content": "hello"}], temperature=0.3) == "chat reply"
+    assert await gateway.embed(config, ["first", "second"]) == [[1.0], [2.0]]
+    assert await gateway.moderate(config, "safe text") == {"flagged": False}
+
+    assert all(request.headers["authorization"] == "Bearer test-key" for request in requests)
+    assert json.loads(requests[0].content) == {
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hello"}],
+        "temperature": 0.3,
+    }
+
+
+async def test_anthropic_messages_contract():
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.headers["x-api-key"] == "test-key"
+        assert request.headers["anthropic-version"] == "2023-06-01"
+        assert "authorization" not in request.headers
+        assert json.loads(request.content) == {
+            "model": "test-model",
+            "max_tokens": 1024,
+            "messages": [
+                {"role": "user", "content": "hello"},
+                {"role": "assistant", "content": "hi"},
+            ],
+            "temperature": 0.1,
+            "system": "be concise",
+        }
+        return httpx.Response(
+            200,
+            json={"content": [{"type": "text", "text": "first"}, {"type": "text", "text": "second"}]},
+        )
+
+    result = await _gateway(handler).chat(
+        _config("anthropic"),
+        [
+            {"role": "system", "content": "be concise"},
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi"},
+        ],
+        temperature=0.1,
+    )
+
+    assert result == "first\nsecond"
+
+
+async def test_google_gemini_generate_content_contract():
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.headers["x-goog-api-key"] == "test-key"
+        assert json.loads(request.content) == {
+            "contents": [
+                {"role": "user", "parts": [{"text": "hello"}]},
+                {"role": "model", "parts": [{"text": "hi"}]},
+            ],
+            "generationConfig": {"temperature": 0.2},
+            "systemInstruction": {"parts": [{"text": "be concise"}]},
+        }
+        return httpx.Response(
+            200,
+            json={"candidates": [{"content": {"parts": [{"text": "Gemini reply"}]}}]},
+        )
+
+    result = await _gateway(handler).chat(
+        _config("google_gemini"),
+        [
+            {"role": "system", "content": "be concise"},
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi"},
+        ],
+    )
+
+    assert result == "Gemini reply"
+
+
+async def test_dashscope_native_contract():
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.headers["authorization"] == "Bearer test-key"
+        assert json.loads(request.content) == {
+            "model": "test-model",
+            "input": {"messages": [{"role": "user", "content": "hello"}]},
+            "parameters": {"result_format": "message", "temperature": 0.4},
+        }
+        return httpx.Response(
+            200,
+            json={"output": {"choices": [{"message": {"content": "DashScope reply"}}]}},
+        )
+
+    result = await _gateway(handler).chat(
+        _config("dashscope"),
+        [{"role": "user", "content": "hello"}],
+        temperature=0.4,
+    )
+
+    assert result == "DashScope reply"
+
+
+@pytest.mark.parametrize("provider", ["anthropic", "google_gemini", "dashscope"])
+async def test_chat_only_protocols_reject_embedding(provider: str):
+    with pytest.raises(ModelGatewayError, match="不支持 embedding 能力"):
+        await _gateway(lambda _: httpx.Response(500)).embed(_config(provider), "hello")