Add integration checks for comparing the result of calling the model API directly vs via CodeGate (#1032)

* Enable codegate enrichment tests

Signed-off-by: Radoslav Dimitrov <[email protected]>

* Re-use call_provider for calling both codegate and the provider

Signed-off-by: Radoslav Dimitrov <[email protected]>

---------

Signed-off-by: Radoslav Dimitrov <[email protected]>

Author: rdimitrov

tests/integration/checks.py

@@ -29,7 +29,8 @@ def load(test_data: dict) -> List[BaseCheck]:
             checks.append(ContainsCheck(test_name))
         if test_data.get(DoesNotContainCheck.KEY):
             checks.append(DoesNotContainCheck(test_name))
-
+        if test_data.get(CodeGateEnrichment.KEY) is not None:
+            checks.append(CodeGateEnrichment(test_name))
         return checks
 
 
@@ -51,11 +52,10 @@ async def run_check(self, parsed_response: str, test_data: dict) -> bool:
         similarity = await self._calculate_string_similarity(
             parsed_response, test_data[DistanceCheck.KEY]
         )
+        logger.debug(f"Similarity: {similarity}")
+        logger.debug(f"Response: {parsed_response}")
+        logger.debug(f"Expected Response: {test_data[DistanceCheck.KEY]}")
         if similarity < 0.8:
-            logger.error(f"Test {self.test_name} failed")
-            logger.error(f"Similarity: {similarity}")
-            logger.error(f"Response: {parsed_response}")
-            logger.error(f"Expected Response: {test_data[DistanceCheck.KEY]}")
             return False
         return True
 
@@ -64,10 +64,9 @@ class ContainsCheck(BaseCheck):
     KEY = "contains"
 
     async def run_check(self, parsed_response: str, test_data: dict) -> bool:
+        logger.debug(f"Response: {parsed_response}")
+        logger.debug(f"Expected Response to contain: {test_data[ContainsCheck.KEY]}")
         if test_data[ContainsCheck.KEY].strip() not in parsed_response:
-            logger.error(f"Test {self.test_name} failed")
-            logger.error(f"Response: {parsed_response}")
-            logger.error(f"Expected Response to contain: '{test_data[ContainsCheck.KEY]}'")
             return False
         return True
 
@@ -76,11 +75,33 @@ class DoesNotContainCheck(BaseCheck):
     KEY = "does_not_contain"
 
     async def run_check(self, parsed_response: str, test_data: dict) -> bool:
+        logger.debug(f"Response: {parsed_response}")
+        logger.debug(f"Expected Response to not contain: '{test_data[DoesNotContainCheck.KEY]}'")
         if test_data[DoesNotContainCheck.KEY].strip() in parsed_response:
-            logger.error(f"Test {self.test_name} failed")
-            logger.error(f"Response: {parsed_response}")
-            logger.error(
-                f"Expected Response to not contain: '{test_data[DoesNotContainCheck.KEY]}'"
-            )
             return False
         return True
+
+
+class CodeGateEnrichment(BaseCheck):
+    KEY = "codegate_enrichment"
+
+    async def run_check(self, parsed_response: str, test_data: dict) -> bool:
+        direct_response = test_data["direct_response"]
+        logger.debug(f"Response (CodeGate): {parsed_response}")
+        logger.debug(f"Response (Raw model): {direct_response}")
+
+        # Use the DistanceCheck to compare the two responses
+        distance_check = DistanceCheck(self.test_name)
+        are_similar = await distance_check.run_check(
+            parsed_response, {DistanceCheck.KEY: direct_response}
+        )
+
+        # Check if the response is enriched by CodeGate.
+        # If it is, there should be a difference in the similarity score.
+        expect_enrichment = test_data.get(CodeGateEnrichment.KEY).get("expect_difference", False)
+        if expect_enrichment:
+            logger.info("CodeGate enrichment check: Expecting difference")
+            return not are_similar
+        # If the response is not enriched, the similarity score should be the same.
+        logger.info("CodeGate enrichment check: Not expecting difference")
+        return are_similar

tests/integration/integration_tests.py

@@ -9,7 +9,7 @@
 import requests
 import structlog
 import yaml
-from checks import CheckLoader
+from checks import CheckLoader, CodeGateEnrichment
 from dotenv import find_dotenv, load_dotenv
 from requesters import RequesterFactory
 
@@ -21,7 +21,7 @@ def __init__(self):
         self.requester_factory = RequesterFactory()
         self.failed_tests = []  # Track failed tests
 
-    def call_codegate(
+    def call_provider(
         self, url: str, headers: dict, data: dict, provider: str, method: str = "POST"
     ) -> Optional[requests.Response]:
         logger.debug(f"Creating requester for provider: {provider}")
@@ -132,18 +132,29 @@ def replacement(match):
 
     async def run_test(self, test: dict, test_headers: dict) -> bool:
         test_name = test["name"]
-        url = test["url"]
         data = json.loads(test["data"])
+        codegate_url = test["url"]
         streaming = data.get("stream", False)
         provider = test["provider"]
-
         logger.info(f"Starting test: {test_name}")
 
-        response = self.call_codegate(url, test_headers, data, provider)
+        # Call Codegate
+        response = self.call_provider(codegate_url, test_headers, data, provider)
         if not response:
             logger.error(f"Test {test_name} failed: No response received")
             return False
 
+        # Call model directly if specified
+        direct_response = None
+        if test.get(CodeGateEnrichment.KEY) is not None:
+            direct_provider_url = test.get(CodeGateEnrichment.KEY)["provider_url"]
+            direct_response = self.call_provider(
+                direct_provider_url, test_headers, data, "not-codegate"
+            )
+            if not direct_response:
+                logger.error(f"Test {test_name} failed: No direct response received")
+                return False
+
         # Debug response info
         logger.debug(f"Response status: {response.status_code}")
         logger.debug(f"Response headers: {dict(response.headers)}")
@@ -152,13 +163,24 @@ async def run_test(self, test: dict, test_headers: dict) -> bool:
             parsed_response = self.parse_response_message(response, streaming=streaming)
             logger.debug(f"Response message: {parsed_response}")
 
+            if direct_response:
+                # Dirty hack to pass direct response to checks
+                test["direct_response"] = self.parse_response_message(
+                    direct_response, streaming=streaming
+                )
+                logger.debug(f"Direct response message: {test['direct_response']}")
+
             # Load appropriate checks for this test
             checks = CheckLoader.load(test)
 
             # Run all checks
             all_passed = True
             for check in checks:
+                logger.info(f"Running check: {check.__class__.__name__}")
                 passed_check = await check.run_check(parsed_response, test)
+                logger.info(
+                    f"Check {check.__class__.__name__} {'passed' if passed_check else 'failed'}"
+                )
                 if not passed_check:
                     all_passed = False
 
@@ -379,6 +401,7 @@ async def main():
     # Exit with status code 1 if any tests failed
     if not all_tests_passed:
         sys.exit(1)
+    logger.info("All tests passed")
 
 
 if __name__ == "__main__":

tests/integration/ollama/testcases.yaml

@@ -31,6 +31,9 @@ testcases:
     name: Ollama Chat
     provider: ollama
     url: http://127.0.0.1:8989/ollama/chat/completions
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:11434/api/chat
+      expect_difference: false
     data: |
       {
         "max_tokens":4096,
@@ -55,6 +58,9 @@ testcases:
     name: Ollama FIM
     provider: ollama
     url: http://127.0.0.1:8989/ollama/api/generate
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:11434/api/generate
+      expect_difference: false
     data: |
       {
         "stream": true,
@@ -88,6 +94,9 @@ testcases:
     name: Ollama Malicious Package
     provider: ollama
     url: http://127.0.0.1:8989/ollama/chat/completions
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:11434/api/chat
+      expect_difference: true
     data: |
       {
         "max_tokens":4096,
@@ -112,6 +121,9 @@ testcases:
     name: Ollama secret redacting chat
     provider: ollama
     url: http://127.0.0.1:8989/ollama/chat/completions
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:11434/api/chat
+      expect_difference: true
     data: |
       {
         "max_tokens":4096,

tests/integration/vllm/testcases.yaml

@@ -31,6 +31,9 @@ testcases:
     name: VLLM Chat
     provider: vllm
     url: http://127.0.0.1:8989/vllm/chat/completions
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:8000/v1/chat/completions
+      expect_difference: false
     data: |
       {
         "max_tokens":4096,
@@ -55,6 +58,10 @@ testcases:
     name: VLLM FIM
     provider: vllm
     url: http://127.0.0.1:8989/vllm/completions
+# This is commented out for now as there's some issue with parsing the streamed response from the model (on the vllm side, not codegate)
+#    codegate_enrichment:
+#      provider_url: http://127.0.0.1:8000/v1/completions
+#      expect_difference: false
     data: |
       {
         "model": "Qwen/Qwen2.5-Coder-0.5B-Instruct",
@@ -84,6 +91,9 @@ testcases:
     name: VLLM Malicious Package
     provider: vllm
     url: http://127.0.0.1:8989/vllm/chat/completions
+    codegate_enrichment:
+      provider_url: http://127.0.0.1:8000/v1/chat/completions
+      expect_difference: true
     data: |
       {
         "max_tokens":4096,