chore(evaluator): remove legacy service tree

TESTING.md

@@ -434,8 +434,7 @@ make test-package PACKAGE=nmp_common
 uv run pytest -v packages/nmp_common/tests/
 
 # Test a specific service
-make test-service SERVICE=evaluator
-uv run pytest -v services/evaluator/tests/
+make test-service SERVICE=guardrails
 
 # Test a specific file
 uv run pytest -v path/to/test_file.py
@@ -746,4 +745,3 @@ If you're updating existing tests:
 - [ ] Chaos engineering tests
 - [ ] Load and stress tests
 - [ ] Contract testing between services
-

conftest.py

@@ -233,7 +233,7 @@ def pytest_collection_modifyitems(config, items):
                 item.add_marker(pytest.mark.e2e)
                 marker_names.add("e2e")
 
-        # Auto-mark integration tests (e.g., /services/evaluator/tests/integration/)
+        # Auto-mark integration tests (e.g., /services/core/jobs/tests/integration/)
         elif "/integration/" in fspath_str:
             if "integration" not in marker_names:
                 item.add_marker(pytest.mark.integration)

docs/evaluator/tutorials/run-llm-judge-evaluation.mdx

@@ -155,12 +155,12 @@ This tutorial uses `nvidia/nemotron-3-nano-30b-a3b` from NVIDIA Build.
 ```python
 from nemo_evaluator_sdk import RunConfig, LLMJudgeMetric
 from nemo_evaluator_sdk.values import (
+    FilesetRef,
     InferenceParams,
     JSONScoreParser,
     Model,
     RangeScore,
 )
-from nmp.evaluator.app.values import FilesetRef
 
 JUDGE_MODEL_URL = "https://integrate.api.nvidia.com/v1/chat/completions"
 JUDGE_MODEL_NAME = "nvidia/nemotron-3-nano-30b-a3b"

packages/nmp_common/src/nmp/common/mcp/README.md

@@ -13,7 +13,6 @@ When multiple MCP servers exist across the platform:
 ```
 services/core/mcp/              # Core infrastructure tools
 services/guardrails/mcp/        # Guardrails-specific tools
-services/evaluator/mcp/         # Evaluation-specific tools
 services/customizer/mcp/        # Customization-specific tools
 ```
 
@@ -157,11 +156,9 @@ When aggregating multiple service MCP servers:
 ```python
 # services/core/mcp/src/nmp/core/mcp/server.py
 from nmp.guardrails.mcp.server import guardrails
-from nmp.evaluator.mcp.server import evaluator
 
 platform = FastMCP("NeMo Platform")
 platform.mount(guardrails)  # All tools use same patterns
-platform.mount(evaluator)   # Consistent for agents
 ```
 
 **Benefits**:

packages/nmp_testing/src/nmp/testing/client.py

@@ -246,11 +246,11 @@ def create_test_client(
 
     Example (with access_log for request verification):
         with create_test_client(
-            EvaluatorService, auth_enabled=True, access_log=True, client_type=ClientContext
+            FilesService, auth_enabled=True, access_log=True, client_type=ClientContext
         ) as ctx:
             ctx.access_log.clear()  # Clear requests from setup
             ctx.test_client.get(
-                "/apis/evaluation/v2/workspaces/default/metrics",
+                "/apis/files/v2/workspaces/default/filesets",
                 headers={"X-NMP-Principal-Id": "test@example.com"},
             )
             # Verify internal entity requests used the same principal

plugins/nemo-guardrails/benchmarks/README.md

@@ -44,9 +44,8 @@ the `bench` extra on `nemo-guardrails-plugin`. The `make benchmark-guardrails`
 target installs them automatically via `uv run --extra bench`; they are not
 part of the plugin's runtime install.
 
-The upstream `aiperf` CLI itself pins `aiofiles<24.2`, which conflicts with
-NMP's evaluator-service. To avoid downgrading the shared workspace venv, the
-harness creates an isolated venv at
+The upstream `aiperf` CLI itself pins older transitive dependencies. To avoid
+downgrading the shared workspace venv, the harness creates an isolated venv at
 `plugins/nemo-guardrails/benchmarks/artifacts/venvs/aiperf/` on first run and
 reuses it on subsequent runs. CI gets a fresh one each invocation; locally
 this caches across runs for fast iteration.

plugins/nemo-guardrails/pyproject.toml

@@ -18,10 +18,9 @@ dependencies = [
 bench = [
     "httpx>=0.27",
     "pyyaml>=6.0",
-    # NOTE: aiperf itself is *not* listed here. Its 0.x line pins
-    # aiofiles<24.2 which conflicts with evaluator-service's
-    # aiofiles>=25.1. To avoid downgrading the shared workspace venv, the
-    # harness installs aiperf into a dedicated venv at run time; see
+    # NOTE: aiperf itself is *not* listed here. Its 0.x line pins older
+    # transitive dependencies. To avoid downgrading the shared workspace venv,
+    # the harness installs aiperf into a dedicated venv at run time; see
     # `nemo_guardrails_plugin.benchmarks.bootstrap`.
 ]
 

plugins/nemo-guardrails/src/nemo_guardrails_plugin/benchmarks/bootstrap.py

@@ -3,9 +3,8 @@
 
 """Bootstrap an isolated venv for the upstream AIPerf load generator.
 
-``aiperf`` pins ``aiofiles<24.2`` which conflicts with NMP's evaluator-service
-requirement of ``aiofiles>=25.1``, so we install it into a dedicated venv
-instead of the shared workspace one. The venv is reused across local runs;
+``aiperf`` pins older transitive dependencies, so we install it into a dedicated
+venv instead of the shared workspace one. The venv is reused across local runs;
 CI gets a fresh one each invocation.
 """
 

ruff.toml

@@ -26,10 +26,8 @@ exclude = [
     "node_modules",
     "site-packages",
     "venv",
-    "services/evaluator/src/external/evaltool",
     "services/auditor/fixes/cve-2025-8194",
     "services/guardrails/fixes/cve-2025-8194",
-    "services/evaluator/fixes/cve-2025-8194",
     "services/safe-synthesizer/fixes/cve-2025-8194",
     "services/safe-synthesizer-api/fixes/cve-2025-8194",
     "architecture/examples",

script/copyright_fixer.py

@@ -158,8 +158,8 @@
 def _matches_path_filter(relpath: str, patterns: list[str]) -> bool:
     """Return True if *relpath* matches any of the given path patterns.
 
-    Patterns are matched as prefixes first (e.g. ``services/evaluator``
-    matches ``services/evaluator/src/foo.py``).  If a pattern contains
+    Patterns are matched as prefixes first (e.g. ``services/guardrails``
+    matches ``services/guardrails/src/foo.py``).  If a pattern contains
     glob characters it falls back to fnmatch on the full relative path.
     """
     for pat in patterns:
@@ -524,7 +524,7 @@ def update_license_headers(
     don't end up with a monster commit::
 
         # Only process two directories
-        ./script/copyright_fixer.py . --include services/evaluator --include packages/models
+        ./script/copyright_fixer.py . --include services/guardrails --include packages/models
 
         # Process everything except generated SDK code
         ./script/copyright_fixer.py . --exclude packages/nemo_platform

script/generate_openapi_spec.py

@@ -495,29 +495,6 @@ def apply_schema_fixes(spec_files: List[str], apply_reorder: bool = True) -> Non
                 print_verbose(f"Applying streaming fixes to {spec_file}")
                 spec = fix_openai_streaming_endpoints(spec)
 
-            # Rename namespaced schemas BEFORE tweak_spec strips namespace prefixes.
-            # This prevents collisions when multiple services define classes with the
-            # same name (e.g. both evaluator and guardrail define "Model").
-            # Keys are the full namespaced schema name as generated by FastAPI/Pydantic;
-            # values are the desired OpenAPI schema name after renaming.
-            namespaced_schema_renames: dict[str, str] = {
-                # SDK Model schema (Model is imported from nemo_evaluator_sdk.values.models)
-                "nmp__evaluator__app__values__models__Model": "Evaluator.Model",
-                "nemo_evaluator_sdk__values__models__Model": "Evaluator.Model",
-            }
-            if "platform" in spec_file:
-                if "components" in spec and "schemas" in spec["components"]:
-                    schemas = spec["components"]["schemas"]
-                    for namespaced_key, new_name in namespaced_schema_renames.items():
-                        if namespaced_key in schemas:
-                            schemas[new_name] = schemas.pop(namespaced_key)
-                            rename_schema_references(spec, namespaced_key, new_name)
-                        else:
-                            print_verbose(
-                                f"Warning: expected schema '{namespaced_key}' not found in {spec_file}, "
-                                f"skipping rename to '{new_name}'"
-                            )
-
             # Apply the standard fix-schema logic
             spec = tweak_spec(spec)
             spec = hoist_nested_defs(spec)

services/core/auth/src/nmp/core/auth/app/policy_tests/helpers_test.rego

@@ -29,6 +29,7 @@ test_normalize_endpoint if {
     mock_endpoints := {
         "/apis/models/v2/workspaces/{workspace}/models": {"get": {}},
         "/apis/models/v2/workspaces/{workspace}/models/{name}": {"get": {}},
+        "/apis/files/v2/workspaces/{workspace}/filesets": {"get": {}},
         "/apis/files/v2/workspaces/{workspace}/filesets/{name}": {"get": {}},
         "/apis/entities/v2/workspaces": {"get": {}},
         "/apis/entities/v2/workspaces/{workspace}/members": {"get": {}}
@@ -52,4 +53,7 @@ test_normalize_endpoint if {
     common.normalize_endpoint("/apis/entities/v2/workspaces") == "/apis/entities/v2/workspaces"
         with data.authz.endpoints as mock_endpoints
 
+    # Test filesets collection pattern
+    common.normalize_endpoint("/apis/files/v2/workspaces/test-ns/filesets") == "/apis/files/v2/workspaces/{workspace}/filesets"
+        with data.authz.endpoints as mock_endpoints
 }

services/core/jobs/tests/controllers/test_subprocess_backend.py

@@ -221,27 +221,27 @@ def test_build_command_uses_current_interpreter_for_python_module_commands() ->
     executor = SubprocessExecutionProvider(
         provider="subprocess",
         profile="default",
-        command=["python", "-m", "nmp.evaluator.tasks.evaluate_metric"],
+        command=["python", "-m", "nemo_evaluator.tasks.evaluate"],
     )
 
     assert SubprocessJobBackend._build_command(executor, None) == [
         sys.executable,
         "-m",
-        "nmp.evaluator.tasks.evaluate_metric",
+        "nemo_evaluator.tasks.evaluate",
     ]
 
 
 def test_build_command_uses_current_interpreter_for_python3_commands() -> None:
     executor = SubprocessExecutionProvider(
         provider="subprocess",
         profile="default",
-        command=["python3", "-m", "nmp.evaluator.tasks.evaluate_metric"],
+        command=["python3", "-m", "nemo_evaluator.tasks.evaluate"],
     )
 
     assert SubprocessJobBackend._build_command(executor, None) == [
         sys.executable,
         "-m",
-        "nmp.evaluator.tasks.evaluate_metric",
+        "nemo_evaluator.tasks.evaluate",
     ]
 
 
@@ -253,13 +253,13 @@ def test_build_command_prefers_virtual_env_python(tmp_path) -> None:
     executor = SubprocessExecutionProvider(
         provider="subprocess",
         profile="default",
-        command=["python", "-m", "nmp.evaluator.tasks.evaluate_metric"],
+        command=["python", "-m", "nemo_evaluator.tasks.evaluate"],
     )
 
     assert SubprocessJobBackend._build_command(executor, str(tmp_path / "venv")) == [
         str(venv_python),
         "-m",
-        "nmp.evaluator.tasks.evaluate_metric",
+        "nemo_evaluator.tasks.evaluate",
     ]
 
 

services/core/mcp/README.md

@@ -101,7 +101,6 @@ This service follows NeMo Platform v2 patterns:
 Future expansion will support mounting service-specific MCP servers from:
 
 - `nmp.guardrails.mcp`
-- `nmp.evaluator.mcp`
 - `nmp.customizer.mcp`
 - etc.