feat: add governed knowledge lifecycle

backend/apps/knowledge/tests/test_retrieval.py

@@ -1,5 +1,7 @@
 from apps.knowledge.models import KnowledgeBase, KnowledgeChunk, KnowledgeItem
-from apps.knowledge.services import retrieve_knowledge
+from apps.knowledge.services import rebuild_item_chunks, retrieve_knowledge
+from apps.modelhub.crypto import encrypt_secret
+from apps.modelhub.models import ModelDeployment, ModelProvider
 from apps.solutions.models import Solution
 
 
@@ -51,3 +53,154 @@ def test_published_solution_can_be_indexed_for_retrieval(db):
     assert len(chunks) == 1
     assert chunks[0].metadata["source_type"] == "solution"
     assert chunks[0].metadata["solution_id"] == solution.id
+
+
+def test_hybrid_retrieval_uses_configured_embedding_and_exposes_score_audit(
+    db,
+    monkeypatch,
+):
+    provider = ModelProvider.objects.create(
+        name="向量渠道",
+        base_url="https://models.example.com/v1",
+        api_key_encrypted=encrypt_secret("secret"),
+    )
+    deployment = ModelDeployment.objects.create(
+        provider=provider,
+        display_name="知识向量模型",
+        model_name="embedding-model",
+        capability="embedding",
+    )
+    vector = [0.0] * 1536
+    vector[0] = 1.0
+    calls = []
+
+    def request_embeddings(selected, contents, dimensions):
+        calls.append((selected.id, contents, dimensions))
+        return [vector[:] for _ in contents]
+
+    monkeypatch.setattr(
+        "apps.knowledge.embeddings._request_embeddings",
+        request_embeddings,
+    )
+    base = KnowledgeBase.objects.create(name="正式知识库")
+    item = KnowledgeItem.objects.create(
+        knowledge_base=base,
+        title="订单超时",
+        problem="订单接口返回 TIMEOUT-504",
+        solution="检查下游连接池",
+        status="published",
+    )
+    chunks = rebuild_item_chunks(item)
+
+    results = retrieve_knowledge("订单 TIMEOUT-504", limit=1, published_only=True)
+
+    assert chunks[0].embedding_model == "embedding-model"
+    assert chunks[0].embedding_vector is not None
+    assert results[0]["metadata"]["retrieval"]["query_embedding_model"] == "embedding-model"
+    assert results[0]["metadata"]["retrieval"]["query_embedding_deployment_id"] == deployment.id
+    assert results[0]["metadata"]["retrieval"]["query_embedding_fallback"] is False
+    assert results[0]["metadata"]["retrieval"]["channels"] == [
+        "portable_text",
+        "portable_vector",
+    ]
+    assert calls[0][2] == 1536
+
+
+def test_configured_rerank_model_controls_final_order_and_is_audited(db, monkeypatch):
+    provider = ModelProvider.objects.create(
+        name="重排渠道",
+        base_url="https://models.example.com/v1",
+        api_key_encrypted=encrypt_secret("secret"),
+    )
+    deployment = ModelDeployment.objects.create(
+        provider=provider,
+        display_name="知识重排模型",
+        model_name="rerank-model",
+        capability="rerank",
+    )
+    base = KnowledgeBase.objects.create(name="重排知识库")
+    first = KnowledgeItem.objects.create(
+        knowledge_base=base,
+        title="订单超时基础方案",
+        problem="订单 TIMEOUT-504",
+        solution="检查连接池",
+        status="published",
+    )
+    preferred = KnowledgeItem.objects.create(
+        knowledge_base=base,
+        title="订单超时权威方案",
+        problem="订单 TIMEOUT-504",
+        solution="按权威步骤恢复连接池",
+        status="published",
+    )
+    rebuild_item_chunks(first)
+    rebuild_item_chunks(preferred)
+
+    def rerank(selected, query, candidates):
+        assert selected.id == deployment.id
+        return {
+            item["id"]: 1.0 if item["title"] == preferred.title else 0.0
+            for item in candidates
+        }
+
+    monkeypatch.setattr("apps.knowledge.reranking._request_rerank", rerank)
+
+    results = retrieve_knowledge("订单 TIMEOUT-504", limit=2, published_only=True)
+
+    assert results[0]["metadata"]["item_id"] == preferred.id
+    audit = results[0]["metadata"]["retrieval"]["rerank"]
+    assert audit["status"] == "succeeded"
+    assert audit["model_name"] == "rerank-model"
+    assert audit["deployment_id"] == deployment.id
+    assert audit["model_weight"] == 0.6
+
+
+def test_configured_rerank_failure_falls_back_to_weighted_order_and_is_audited(
+    db,
+    monkeypatch,
+):
+    provider = ModelProvider.objects.create(
+        name="异常重排渠道",
+        base_url="https://models.example.com/v1",
+        api_key_encrypted=encrypt_secret("secret"),
+    )
+    deployment = ModelDeployment.objects.create(
+        provider=provider,
+        display_name="异常知识重排模型",
+        model_name="broken-rerank-model",
+        capability="rerank",
+    )
+    base = KnowledgeBase.objects.create(name="重排降级知识库")
+    relevant = KnowledgeItem.objects.create(
+        knowledge_base=base,
+        title="订单 TIMEOUT-504 处理",
+        problem="订单 TIMEOUT-504",
+        solution="检查连接池",
+        status="published",
+    )
+    unrelated = KnowledgeItem.objects.create(
+        knowledge_base=base,
+        title="账号登录失败",
+        problem="账号密码错误",
+        solution="重置密码",
+        status="published",
+    )
+    rebuild_item_chunks(relevant)
+    rebuild_item_chunks(unrelated)
+
+    def failed_rerank(*_args, **_kwargs):
+        raise TimeoutError("rerank unavailable")
+
+    monkeypatch.setattr(
+        "apps.knowledge.reranking._request_rerank",
+        failed_rerank,
+    )
+
+    results = retrieve_knowledge("订单 TIMEOUT-504", limit=2, published_only=True)
+
+    assert results[0]["metadata"]["item_id"] == relevant.id
+    audit = results[0]["metadata"]["retrieval"]["rerank"]
+    assert audit["status"] == "failed_fallback"
+    assert audit["model_name"] == "broken-rerank-model"
+    assert audit["deployment_id"] == deployment.id
+    assert audit["model_weight"] == 0.0