feat: add embedding types, port and openai-compat transport

src/polygateway/ports.py

@@ -10,7 +10,14 @@ from dataclasses import dataclass
 from enum import StrEnum
 from typing import Any, Protocol, runtime_checkable
 
-from .types import ChatRequest, LLMResponse, SourceConfig, SourceStats, TransportResult
+from .types import (
+    ChatRequest,
+    EmbeddingTransportResult,
+    LLMResponse,
+    SourceConfig,
+    SourceStats,
+    TransportResult,
+)
 
 CallNext = Callable[[ChatRequest], Awaitable[LLMResponse]]
 
@@ -37,6 +44,15 @@ class Transport(Protocol):
     ) -> TransportResult: ...
 
 
+@runtime_checkable
+class EmbeddingTransport(Protocol):
+    """一次原始 embedding 调用的协议细节(M2 §7);不含任何治理。"""
+
+    async def embed(
+        self, *, texts: list[str], source: SourceConfig, call_id: str
+    ) -> EmbeddingTransportResult: ...
+
+
 @runtime_checkable
 class Permit(Protocol):
     """限流入场许可;settle/release 均幂等,finally 中必然执行。"""

src/polygateway/transports/openai_compat.py

@@ -15,10 +15,15 @@ from typing import TYPE_CHECKING, Any
 
 import httpx
 
-from polygateway.errors import RequestRejectedError, SourceDeadError, TransientError
+from polygateway.errors import (
+    RequestRejectedError,
+    ResultInvalidError,
+    SourceDeadError,
+    TransientError,
+)
 from polygateway.providers import ProviderProfile, get_provider
 from polygateway.streaming import StreamLivenessTimeout, stream_with_liveness_timeouts
-from polygateway.types import SourceConfig, TransportResult
+from polygateway.types import EmbeddingTransportResult, SourceConfig, TransportResult
 
 if TYPE_CHECKING:
     from collections.abc import AsyncIterator, Callable, Mapping
@@ -141,6 +146,56 @@ def _resolve_usage(usage: dict[str, Any], source: SourceConfig) -> tuple[int, in
     return 0, source.est_tokens, "estimated"
 
 
+def _extract_vectors(
+    data: dict[str, Any], source: SourceConfig, expected_count: int, ctx: dict[str, Any]
+) -> list[list[float]]:
+    """按 data[].index 重排提取向量并校验条数/维度一致性。"""
+    try:
+        vectors = [
+            [float(x) for x in item["embedding"]]
+            for item in sorted(data["data"], key=lambda it: int(it["index"]))
+        ]
+    except (KeyError, TypeError, ValueError) as exc:
+        raise ResultInvalidError(f"{source.name} embedding 响应形态异常: {exc}", **ctx) from exc
+    if len(vectors) != expected_count:
+        raise ResultInvalidError(
+            f"{source.name} 返回 {len(vectors)} 条向量,与输入 {expected_count} 条不符", **ctx
+        )
+    if len({len(v) for v in vectors}) != 1 or not vectors[0]:
+        raise ResultInvalidError(
+            f"{source.name} 向量维度异常: {sorted({len(v) for v in vectors})}", **ctx
+        )
+    return vectors
+
+
+def _resolve_embedding_usage(data: dict[str, Any], source: SourceConfig) -> tuple[int, str]:
+    """usage 读取;缺失/非法按 est_tokens 保守兜底并标 estimated(与 chat 同口径)。"""
+    prompt = (data.get("usage") or {}).get("prompt_tokens")
+    if isinstance(prompt, int) and prompt > 0:
+        return prompt, "measured"
+    return source.est_tokens, "estimated"
+
+
+def _parse_embedding_payload(
+    resp: httpx.Response, source: SourceConfig, expected_count: int
+) -> EmbeddingTransportResult:
+    """解析 /embeddings 响应;一切形态异常归 ResultInvalidError(坏结果≠坏服务)。"""
+    ctx: dict[str, Any] = {"source_name": source.name, "operation": "embedding"}
+    try:
+        data = resp.json()
+    except json.JSONDecodeError as exc:
+        raise ResultInvalidError(f"{source.name} embedding 响应非 JSON: {exc}", **ctx) from exc
+    vectors = _extract_vectors(data, source, expected_count, ctx)
+    prompt_tokens, usage_source = _resolve_embedding_usage(data, source)
+    return EmbeddingTransportResult(
+        vectors=vectors,
+        dim=len(vectors[0]),
+        prompt_tokens=prompt_tokens,
+        usage_source=usage_source,
+        raw={"id": data.get("id")},
+    )
+
+
 def _default_client_factory(source: SourceConfig) -> httpx.AsyncClient:
     return httpx.AsyncClient(
         headers={"Authorization": f"Bearer {source.api_key}"},
@@ -217,6 +272,29 @@ class OpenAICompatTransport:
             # VT 宽集: 覆盖断连/协议错误/读写失败(设计 §9 行 8)
             raise TransientError(f"{source.name} 网络错误: {exc}", **ctx) from exc
 
+    async def embed(
+        self, *, texts: list[str], source: SourceConfig, call_id: str
+    ) -> EmbeddingTransportResult:
+        """一次原始 embedding 调用(M2 §7): POST /embeddings,错误翻译同 chat。
+
+        响应按 data[].index 重排保序(GovDoc embedding.py:149 / VT :164 同款);
+        空 data/长度不符/维度不一致 → ResultInvalidError(坏结果不熔断)。
+        """
+        if not texts:
+            raise ValueError("texts 不能为空(空输入由 EmbeddingClient 短路)")
+        url = source.base_url.rstrip("/") + "/embeddings"
+        client = self._client_for(source)
+        ctx: dict[str, Any] = {"source_name": source.name, "operation": "embedding"}
+        try:
+            resp = await client.post(url, json={"model": source.model, "input": texts})
+        except httpx.TimeoutException as exc:
+            raise TransientError(f"{source.name} 超时: {exc}", **ctx) from exc
+        except httpx.TransportError as exc:
+            raise TransientError(f"{source.name} 网络错误: {exc}", **ctx) from exc
+        if resp.status_code != 200:
+            raise _status_to_error(source, resp.status_code, resp.text, resp.headers)
+        return _parse_embedding_payload(resp, source, len(texts))
+
     async def _complete_stream(
         self,
         client: httpx.AsyncClient,

src/polygateway/types.py

@@ -186,3 +186,30 @@ class GlobalLimits:
     def __post_init__(self) -> None:
         if self.max_concurrency < 0 or self.rpm < 0 or self.tpm < 0:
             raise ValueError("全局限额不能为负(0 表示不启用)")
+
+
+@dataclass(frozen=True)
+class EmbeddingTransportResult:
+    """一次原始 embedding 调用的解析结果(M2 设计 §7.2;transport → client)。"""
+
+    vectors: list[list[float]]
+    dim: int
+    prompt_tokens: int
+    usage_source: str  # measured | estimated
+    raw: dict[str, Any]
+
+
+@dataclass(frozen=True)
+class EmbeddingResponse:
+    """一次治理 embedding 调用的统一响应(多批合并;与输入等长保序)。"""
+
+    vectors: list[list[float]]
+    dim: int
+    model: str
+    provider: str
+    prompt_tokens: int
+    usage_source: str
+    latency_ms: int
+    call_id: str
+    source_name: str
+    cost: float | None = None