fix: harden model catalog projections

2026-09-02 19:25:01 +02:00
parent ce4c31a6fb
commit a6a5bf2036
38 changed files with 573 additions and 83 deletions
@@ -214,7 +214,7 @@ embeddings: {provider: ollama_internal, base_url: http://embedding:11434, model:
    assert cfg.vectors.writer.api_key == "writer"


-def test_accepts_only_internal_ollama_embedding_contract(tmp_path):
+def test_accepts_catalog_selected_internal_ollama_embedding_contract(tmp_path):
    workspace = tmp_path / "workspace.yaml"
    workspace.write_text(
        """
@@ -225,8 +225,9 @@ resources:
  embeddings:
    provider: ollama_internal
    base_url: http://embedding:11434
-    model: qwen3-embedding:0.6b
-    dimensions: 1024
+    id: ollama/bge-m3
+    model: bge-m3
+    dimensions: 1536
 """
    )

@@ -234,8 +235,59 @@ resources:

    assert cfg.embeddings.provider == "ollama_internal"
    assert cfg.embeddings.base_url == "http://embedding:11434"
-    assert cfg.embeddings.model == "qwen3-embedding:0.6b"
-    assert cfg.embeddings.dim == 1024
+    assert cfg.embeddings.id == "ollama/bge-m3"
+    assert cfg.embeddings.model == "bge-m3"
+    assert cfg.embeddings.dim == 1536
+
+
+def test_installation_embedding_projection_completes_model_free_runtime_source(
+    tmp_path, monkeypatch,
+):
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_ID", "ollama/bge-m3")
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_MODEL", "bge-m3")
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_DIMENSIONS", "1536")
+    workspace = tmp_path / "workspace.yaml"
+    workspace.write_text(
+        """
+dwh:
+  type: postgres_direct
+  connection: {database: analytics, schema: mart, user: reader, password: secret}
+resources:
+  embeddings:
+    provider: ollama_internal
+    base_url: http://embedding:11434
+"""
+    )
+
+    cfg = load_config(workspace)
+
+    assert cfg.embeddings.id == "ollama/bge-m3"
+    assert cfg.embeddings.model == "bge-m3"
+    assert cfg.embeddings.dim == 1536
+
+
+def test_workspace_embedding_values_cannot_override_installation_projection(
+    tmp_path, monkeypatch,
+):
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_ID", "ollama/bge-m3")
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_MODEL", "bge-m3")
+    monkeypatch.setenv("THT_INTERNAL_EMBEDDING_DIMENSIONS", "1536")
+    workspace = tmp_path / "workspace.yaml"
+    workspace.write_text(
+        """
+dwh:
+  type: postgres_direct
+  connection: {database: analytics, schema: mart, user: reader, password: secret}
+resources:
+  embeddings:
+    provider: ollama_internal
+    base_url: http://embedding:11434
+    model: workspace-owned-model
+"""
+    )
+
+    with pytest.raises(ConfigError, match="proprietà dell'installazione|diverge"):
+        load_config(workspace)


 def test_accepts_internal_qdrant_resource_contract(tmp_path):
@@ -40,6 +40,7 @@ def test_manifest_contains_provenance_without_credentials():
        manifest_id="manifest:abc",
        created_at=datetime(2026, 7, 12, tzinfo=UTC),
        pipeline_version="evidence-v1",
+        embedding_id="ollama/nomic-embed-text",
        embedding_model="nomic-embed-text",
        embedding_dimensions=768,
        documents=[document()],
@@ -52,6 +53,7 @@ def test_manifest_contains_provenance_without_credentials():
    assert "etag:abc" in payload
    assert "evidence-v1" in payload
    assert "nomic-embed-text" in payload
+    assert '"embedding_id":"ollama/nomic-embed-text"' in payload
    assert "api_key" not in payload


@@ -93,6 +95,25 @@ def test_manifest_validates_embedding_compatibility_fields():
        )


+def test_schema_v1_manifest_without_canonical_embedding_id_remains_readable():
+    manifest = CorpusManifest.model_validate({
+        "schema_version": 1,
+        "embedding_model": "legacy-model",
+        "embedding_dimensions": 768,
+    })
+
+    assert manifest.embedding_id is None
+
+
+def test_schema_v2_embedding_generation_requires_canonical_id():
+    with pytest.raises(ValidationError, match="embedding_id"):
+        CorpusManifest(
+            schema_version=2,
+            embedding_model="model-v2",
+            embedding_dimensions=768,
+        )
+
+
 def test_canonical_metadata_rejects_secrets_and_non_json_values():
    with pytest.raises(ValidationError, match="credential-like"):
        CanonicalDocument.model_validate(
@@ -105,11 +105,13 @@ def item(name, fingerprint):
    )


-def pipeline(tmp_path, source, *, embedder=None, vectors=None, model="model-a", policy=None,
+def pipeline(tmp_path, source, *, embedder=None, vectors=None, model="model-a",
+             embedding_id=None, policy=None,
             retain=3, candidate_evaluator=None):
    return CorpusPipeline(
        store=CorpusStore(tmp_path / "corpus"), sources=[source],
        embedder=embedder or Embedder(), vector_store=vectors or Vectors(),
+        embedding_id=embedding_id,
        embedding_model=model, embedding_dimensions=3,
        chunk_policy=policy or ChunkPolicy(version="chunk-v1", max_chars=100),
        pipeline_version="evidence-v1",
@@ -875,6 +877,24 @@ def test_model_or_chunk_policy_change_forces_full_rebuild(tmp_path):
    assert source.acquire_calls == ["fs:one"]


+def test_canonical_embedding_id_change_forces_full_rebuild_and_is_persisted(tmp_path):
+    one = item("one", "a")
+    first = pipeline(
+        tmp_path, Source([(one, "hello")]), model="same-upstream",
+        embedding_id="ollama/catalog-a",
+    ).run()
+    assert first.manifest.embedding_id == "ollama/catalog-a"
+
+    source = Source([(one, "hello")])
+    changed = pipeline(
+        tmp_path, source, model="same-upstream", embedding_id="ollama/catalog-b",
+    ).run()
+
+    assert changed.changed == ("fs:one",)
+    assert changed.manifest.embedding_id == "ollama/catalog-b"
+    assert source.acquire_calls == ["fs:one"]
+
+
 def test_partial_vector_failure_never_changes_active_or_exposes_generation(tmp_path):
    one = item("one", "a")
    good = pipeline(tmp_path, Source([(one, "old")]))
@@ -21,7 +21,8 @@ def _write_cfg(tmp_path, raw):
    return cfg


-def _cfg(tmp_path, *, transport="thoth_rest", base_url="http://dwh.example.invalid", collection="psd", model="qwen3-embedding:0.6b"):
+def _cfg(tmp_path, *, transport="thoth_rest", base_url="http://dwh.example.invalid",
+         collection="psd", model="qwen3-embedding:0.6b", dimensions=1024):
    return {
        "schemaVersion": 1,
        "workspace": {"schema_version": 3, "id": "psd", "name": "PSD", "language": "it"},
@@ -34,7 +35,10 @@ def _cfg(tmp_path, *, transport="thoth_rest", base_url="http://dwh.example.inval
            "connection": {"host": "h", "port": 5432, "database": "warehouse", "schema": "dw", "user": "reader", "password": "secret"},
        },
        "vectors": {"type": "qdrant", "base_url": "http://qdrant:6333", "collection": collection, "collection_lifecycle": "self_heal"},
-        "embeddings": {"provider": "ollama_internal", "base_url": "http://embedding:11434", "model": model, "dimensions": 1024},
+        "embeddings": {
+            "provider": "ollama_internal", "base_url": "http://embedding:11434",
+            "id": f"ollama/{model}", "model": model, "dimensions": dimensions,
+        },
        "roots": {"artifacts": str(tmp_path / "artifacts"), "indexes": str(tmp_path / "indexes")},
        "paths": {"artifacts": str(tmp_path / "artifacts"), "indexes": str(tmp_path / "indexes"), "sessions": str(tmp_path / "sessions")},
    }
@@ -52,6 +56,16 @@ def test_canonical_json_is_deterministic_and_key_ordered(cfg):
    assert keys == ["schemaVersion", "dwh", "vector", "embedding", "roots"]


+def test_catalog_embedding_identity_and_dimensions_drive_effective_config(tmp_path):
+    cfg = _write_cfg(tmp_path, _cfg(tmp_path, model="bge-m3", dimensions=1536))
+    document = __import__("json").loads(canonical_effective_config_json(cfg))
+
+    assert document["embedding"] == {
+        "id": "ollama/bge-m3", "model": "bge-m3", "dimensions": 1536,
+    }
+    assert document["vector"]["dimensions"] == 1536
+
+
 def test_canonical_excludes_credentials_and_evidence(cfg):
    doc = canonical_effective_config_json(cfg)
    assert "secret" not in doc
@@ -149,6 +149,7 @@ def test_preprocessing_factory_forwards_only_evidence_pipeline_dependencies(monk
        "sources": [object()],
        "embedder": object(),
        "vector_store": object(),
+        "embedding_id": "ollama/model",
        "embedding_model": "model",
        "embedding_dimensions": 3,
        "chunk_policy": object(),
@@ -10,6 +10,8 @@ from tht.config import EmbeddingsConfig


 def _cfg(**kw):
+    kw.setdefault("model", "nomic-embed-text-v2-moe")
+    kw.setdefault("dim", 768)
    emb = EmbeddingsConfig(base_url="http://localhost:11434", **kw)
    return SimpleNamespace(embeddings=emb)

@@ -284,6 +284,7 @@ def test_run_from_config_uses_runtime_identity_workspace_id(monkeypatch, tmp_pat
            sources,
            embedder,
            vector_store,
+            embedding_id,
            embedding_model,
            embedding_dimensions,
            chunk_policy,
@@ -293,6 +294,7 @@ def test_run_from_config_uses_runtime_identity_workspace_id(monkeypatch, tmp_pat
            candidate_evaluator,
        ):
            calls["init"] = {
+                "embedding_id": embedding_id,
                "embedding_model": embedding_model,
                "embedding_dimensions": embedding_dimensions,
                "pipeline_version": pipeline_version,
@@ -314,6 +316,7 @@ def test_run_from_config_uses_runtime_identity_workspace_id(monkeypatch, tmp_pat

    command.run_from_config(config)

+    assert calls["init"]["embedding_id"] == "ollama/qwen3-embedding:0.6b"
    assert calls["init"]["sparse_language"] == "english"
    assert calls["init"]["candidate_evaluator"] is None
    assert calls["run_as_job"]["workspace_id"] == "psd-clinical"