feat(inference): add MiniMax native embedding adapter

src/powercontext/builtin/artifacts/memory/service.py

@@ -86,6 +86,7 @@
     EmbeddingVector,
     InferenceTimeoutError,
     InferenceUnavailableError,
+    embed_query,
 )
 from powercontext.builtin.tags import TagFilter
 from powercontext.errors import RevisionConflictError
@@ -539,7 +540,7 @@ async def _resolve_query_vector(
         if reuse is not None and reuse.embedding_profile == profile:
             return selected_mode, reuse.query_vector, reuse, 0
         try:
-            query_vector = (await self._embed_texts((query,), profile))[0]
+            query_vector = (await self._embed_texts((query,), profile, query=True))[0]
         except (InferenceUnavailableError, InferenceTimeoutError) as error:
             if requested_mode == "auto" and capabilities.fts:
                 return "fts", None, None, 1
@@ -987,11 +988,13 @@ async def _embed_texts(
         self,
         texts: tuple[str, ...],
         profile: EmbeddingProfile,
+        *,
+        query: bool = False,
     ) -> tuple[EmbeddingVector, ...]:
         embedding_model = self._embedding_model
         if embedding_model is None or embedding_model.profile != profile:
             raise CapabilityNotSupportedError("embedding-profile")
-        result = await embedding_model.embed(texts)
+        result = await embed_query(embedding_model, texts) if query else await embedding_model.embed(texts)
         vectors = result.vectors
         if len(vectors) != len(texts):
             raise InvalidEmbeddingError("count")

src/powercontext/builtin/inference/__init__.py

@@ -31,7 +31,12 @@
     GenerationResult,
     InferenceUsage,
 )
-from powercontext.builtin.inference.protocols import EmbeddingModel, StructuredGenerator
+from powercontext.builtin.inference.protocols import (
+    EmbeddingModel,
+    QueryEmbeddingModel,
+    StructuredGenerator,
+    embed_query,
+)
 from powercontext.builtin.inference.tokens import TokenEstimator, TokenEstimatorProfile, character_token_estimator
 
 __all__ = [
@@ -45,8 +50,10 @@
     "InferenceUnavailableError",
     "InferenceUsage",
     "InvalidInferenceOutputError",
+    "QueryEmbeddingModel",
     "StructuredGenerator",
     "TokenEstimator",
     "TokenEstimatorProfile",
     "character_token_estimator",
+    "embed_query",
 ]

src/powercontext/builtin/inference/minimax.py

@@ -0,0 +1,189 @@
+# Copyright (c) 2026 OceanBase.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""MiniMax embedding backend adapter.
+
+MiniMax publishes embeddings under the OpenAI path prefix but does not implement
+the OpenAI embeddings contract. Its ``/v1/embeddings`` endpoint requires the
+native request shape ``{"model", "texts": [...], "type"}`` instead of OpenAI's
+``{"input": [...]}``, and returns ``{"vectors": [...], "base_resp": {"status_code"}}``
+instead of ``{"data": [{"embedding": [...]}]}``. It also answers HTTP 200 with a
+non-zero ``base_resp.status_code`` on error rather than a 4xx/5xx. This adapter
+speaks the MiniMax-native shape directly so PowerContext can use MiniMax as an
+embedding provider without a custom Pydantic AI provider shim.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from collections.abc import Mapping, Sequence
+
+import httpx
+
+from powercontext.builtin.artifacts.memory.canonical import canonical_embedding
+from powercontext.builtin.artifacts.memory.models import EmbeddingProfile
+from powercontext.builtin.inference.errors import (
+    InferenceConfigurationError,
+    InferenceTimeoutError,
+    InferenceUnavailableError,
+    InvalidInferenceOutputError,
+)
+from powercontext.builtin.inference.models import EmbeddingResult, InferenceUsage
+
+_MINIMAX_DOCUMENT_TYPE = "db"
+_MINIMAX_QUERY_TYPE = "query"
+
+
+class MiniMaxEmbeddingModel:
+    """Embed an ordered text batch with MiniMax ``embo`` via ``/v1/embeddings``."""
+
+    def __init__(
+        self,
+        *,
+        base_url: str,
+        model: str,
+        headers: Mapping[str, str] | None = None,
+        profile: EmbeddingProfile,
+        batch_size: int = 10,
+        timeout_seconds: float = 30.0,
+        http_client: httpx.AsyncClient | None = None,
+    ) -> None:
+        if not model or not model.strip():
+            raise InferenceConfigurationError("embedding-model-empty")
+        if profile.dimension < 1:
+            raise InferenceConfigurationError("embedding-dimension-positive")
+        if batch_size < 1:
+            raise InferenceConfigurationError("embedding-batch-size-positive")
+        if not profile.profile_id.strip() or not profile.model.strip() or not profile.normalization.strip():
+            raise InferenceConfigurationError("embedding-profile-identifiers")
+        self.profile = profile
+        self._model = model
+        self._batch_size = batch_size
+        self._timeout = timeout_seconds
+        self._endpoint = f"{base_url.rstrip('/')}/embeddings"
+        request_headers: dict[str, str] = {"Content-Type": "application/json"}
+        if headers:
+            request_headers.update(headers)
+        self._headers = request_headers
+        self._client = http_client or httpx.AsyncClient()
+
+    async def aclose(self) -> None:
+        """Close the underlying HTTP client (registered with the runtime stack)."""
+
+        await self._client.aclose()
+
+    async def embed(self, texts: tuple[str, ...], /) -> EmbeddingResult:
+        """Embed documents, validating order, count, dimension, and finite values."""
+
+        return await self._embed(texts, embedding_type=_MINIMAX_DOCUMENT_TYPE)
+
+    async def embed_query(self, texts: tuple[str, ...], /) -> EmbeddingResult:
+        """Embed retrieval queries, validating order, count, dimension, and finite values."""
+
+        return await self._embed(texts, embedding_type=_MINIMAX_QUERY_TYPE)
+
+    async def _embed(self, texts: tuple[str, ...], *, embedding_type: str) -> EmbeddingResult:
+        """Embed one MiniMax document or query batch."""
+
+        if not texts:
+            return EmbeddingResult(vectors=())
+
+        try:
+            result = await asyncio.wait_for(
+                self._embed_batches(texts, embedding_type=embedding_type), timeout=self._timeout
+            )
+        except asyncio.CancelledError:
+            raise
+        except (InvalidInferenceOutputError, InferenceConfigurationError):
+            raise
+        except TimeoutError as error:
+            raise InferenceTimeoutError("embed", self._timeout) from error
+        except httpx.HTTPError as error:
+            raise InferenceUnavailableError("embed") from error
+        except Exception as error:
+            raise InferenceUnavailableError("embed") from error
+        return result
+
+    async def _embed_batches(self, texts: tuple[str, ...], *, embedding_type: str) -> EmbeddingResult:
+        vectors: list[tuple[float, ...]] = []
+        requests = 0
+        input_tokens = 0
+        for start in range(0, len(texts), self._batch_size):
+            batch = texts[start : start + self._batch_size]
+            rows, tokens = await self._embed_one(batch, embedding_type=embedding_type)
+            vectors.extend(self._validated_vectors(batch, rows))
+            requests += 1
+            input_tokens += tokens
+        return EmbeddingResult(
+            vectors=tuple(vectors),
+            usage=InferenceUsage(requests=requests, input_tokens=input_tokens, output_tokens=None),
+        )
+
+    async def _embed_one(self, batch: Sequence[str], *, embedding_type: str) -> tuple[list[list[float]], int]:
+        payload = {"model": self._model, "texts": list(batch), "type": embedding_type}
+        response = await self._client.post(self._endpoint, json=payload, headers=self._headers)
+        # MiniMax returns HTTP 200 with a non-zero base_resp.status_code on error;
+        # only a real transport/HTTP failure reaches raise_for_status first.
+        response.raise_for_status()
+        data = response.json()
+        if not isinstance(data, Mapping):
+            raise InvalidInferenceOutputError("embed", "provider response was not a JSON object")
+        base_resp = data.get("base_resp")
+        if isinstance(base_resp, Mapping):
+            status_code = base_resp.get("status_code")
+            if status_code is not None and status_code != 0:
+                raise InferenceUnavailableError("embed")
+        vectors = data.get("vectors")
+        if not isinstance(vectors, list) or len(vectors) != len(batch):
+            raise InvalidInferenceOutputError("embed", "provider returned no vectors or the wrong vector count")
+        total_tokens = data.get("total_tokens")
+        tokens = int(total_tokens) if isinstance(total_tokens, int) else 0
+        return vectors, tokens
+
+    def _validated_vectors(
+        self, texts: Sequence[str], rows: Sequence[Sequence[float]]
+    ) -> tuple[tuple[float, ...], ...]:
+        out: list[tuple[float, ...]] = []
+        for row in rows:
+            try:
+                out.append(
+                    canonical_embedding(
+                        tuple(row),
+                        dimension=self.profile.dimension,
+                        normalization=self.profile.normalization,
+                    )
+                )
+            except (TypeError, ValueError) as error:
+                raise InvalidInferenceOutputError("embed", str(error)) from error
+        return tuple(out)
+
+
+def _embedding_model_name(model: str | None) -> str:
+    """Strip a provider prefix such as ``openai:embo-01`` or ``minimax:embo-01``."""
+
+    if model is None:
+        return ""
+    return model.partition(":")[2] or model
+
+
+def is_minimax_embedding(base_url: str | None, model: str | None) -> bool:
+    """Detect a MiniMax embedding endpoint by host or explicit model prefix."""
+
+    base = base_url or ""
+    if any(fragment in base.lower() for fragment in ("minimaxi.com", "minimax.io")):
+        return True
+    return "minimax" in (model or "").lower()
+
+
+__all__ = ["MiniMaxEmbeddingModel", "is_minimax_embedding"]

src/powercontext/builtin/inference/protocols.py

@@ -16,7 +16,7 @@
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Protocol, TypeVar
+from typing import TYPE_CHECKING, Protocol, TypeVar, runtime_checkable
 
 from powercontext.builtin.inference.models import EmbeddingResult, GenerationResult
 
@@ -45,3 +45,21 @@ async def embed(self, texts: tuple[str, ...], /) -> EmbeddingResult:
         """Return exactly one ordered vector for each input text."""
 
         ...
+
+
+@runtime_checkable
+class QueryEmbeddingModel(Protocol):
+    """Embed retrieval queries when a provider distinguishes queries from documents."""
+
+    async def embed_query(self, texts: tuple[str, ...], /) -> EmbeddingResult:
+        """Return exactly one ordered query vector for each input text."""
+
+        ...
+
+
+async def embed_query(model: EmbeddingModel, texts: tuple[str, ...], /) -> EmbeddingResult:
+    """Embed retrieval queries, falling back to document embeddings for symmetric models."""
+
+    if isinstance(model, QueryEmbeddingModel):
+        return await model.embed_query(texts)
+    return await model.embed(texts)

src/powercontext/builtin/inference/pydantic_ai.py

@@ -20,7 +20,7 @@
 from collections.abc import Sequence
 from contextlib import nullcontext
 from copy import copy
-from typing import Generic, Self, TypeVar, cast
+from typing import Generic, Literal, Self, TypeVar, cast
 
 from pydantic import BaseModel, Field
 from typing_extensions import override
@@ -245,11 +245,23 @@ def _without_instrumentation(self) -> Self:
     async def embed(self, texts: tuple[str, ...], /) -> EmbeddingResult:
         """Embed documents and validate order, count, dimension, and finite values."""
 
+        return await self._embed(texts, input_type="document")
+
+    async def embed_query(self, texts: tuple[str, ...], /) -> EmbeddingResult:
+        """Embed retrieval queries and validate order, count, dimension, and finite values."""
+
+        return await self._embed(texts, input_type="query")
+
+    async def _embed(self, texts: tuple[str, ...], *, input_type: Literal["document", "query"]) -> EmbeddingResult:
+        """Embed one document or query batch through Pydantic AI."""
+
         if not texts:
             return EmbeddingResult(vectors=())
 
         try:
-            result = await asyncio.wait_for(self._embed_batches(texts), timeout=self._limits.timeout_seconds)
+            result = await asyncio.wait_for(
+                self._embed_batches(texts, input_type=input_type), timeout=self._limits.timeout_seconds
+            )
         except asyncio.CancelledError:
             raise
         except InvalidInferenceOutputError:
@@ -266,14 +278,18 @@ async def embed(self, texts: tuple[str, ...], /) -> EmbeddingResult:
 
         return result
 
-    async def _embed_batches(self, texts: tuple[str, ...]) -> EmbeddingResult:
+    async def _embed_batches(
+        self, texts: tuple[str, ...], *, input_type: Literal["document", "query"]
+    ) -> EmbeddingResult:
         vectors: list[tuple[float, ...]] = []
         requests = 0
         input_tokens = 0
         for start in range(0, len(texts), self._batch_size):
             batch = texts[start : start + self._batch_size]
-            result = await self._embedder.embed_documents(batch)
-            vectors.extend(self._validated_vectors(batch, result.inputs, result.input_type, result.embeddings))
+            result = await self._embedder.embed(batch, input_type=input_type)
+            vectors.extend(
+                self._validated_vectors(batch, result.inputs, result.input_type, result.embeddings, input_type)
+            )
             requests += 1
             input_tokens += result.usage.input_tokens
         return EmbeddingResult(
@@ -287,8 +303,9 @@ def _validated_vectors(
         inputs: Sequence[str],
         input_type: str,
         values: Sequence[Sequence[float]],
+        expected_input_type: Literal["document", "query"],
     ) -> tuple[tuple[float, ...], ...]:
-        if input_type != "document":
+        if input_type != expected_input_type:
             raise InvalidInferenceOutputError("embed", "provider returned the wrong input type")
         returned_inputs = tuple(inputs)
         rows = tuple(values)

src/powercontext/builtin/inference/usage.py

@@ -23,7 +23,7 @@
 from typing import Generic, TypeVar
 
 from powercontext.builtin.inference.models import EmbeddingResult, GenerationResult, InferenceUsage
-from powercontext.builtin.inference.protocols import EmbeddingModel, StructuredGenerator
+from powercontext.builtin.inference.protocols import EmbeddingModel, StructuredGenerator, embed_query
 from powercontext.builtin.statistics import ModelUsageOperation, ModelUsagePurpose
 
 InputT = TypeVar("InputT")
@@ -101,5 +101,10 @@ async def embed(self, texts: tuple[str, ...], /) -> EmbeddingResult:
         await _report(ModelUsageOperation.EMBEDDING, result.usage)
         return result
 
+    async def embed_query(self, texts: tuple[str, ...], /) -> EmbeddingResult:
+        result = await embed_query(self._delegate, texts)
+        await _report(ModelUsageOperation.EMBEDDING, result.usage)
+        return result
+
 
 __all__ = ["UsageReportingEmbeddingModel", "UsageReportingStructuredGenerator", "bind_usage_reporter"]

src/powercontext/builtin/runtime/application.py

@@ -125,6 +125,7 @@
     InferenceTimeoutError,
     InferenceUnavailableError,
     InvalidInferenceOutputError,
+    embed_query,
 )
 from powercontext.builtin.inference.models import InferenceUsage
 from powercontext.builtin.inference.usage import bind_usage_reporter
@@ -2612,7 +2613,7 @@ async def _search_with_embedding(
             )
             return result.model_copy(update={"query_embedding": query_embedding, "embedding_calls": 0}), False
         try:
-            embedded = await embedding.embed((request.query,))
+            embedded = await embed_query(embedding, (request.query,))
             if len(embedded.vectors) != 1:
                 raise InvalidInferenceOutputError("embed", "provider returned the wrong vector count")
         except (InferenceUnavailableError, InferenceTimeoutError) as error: