📝 doc: Azure content safety Proxy usage

Signed-off-by: Lunik <lunik@tiwabbit.fr>

docs/my-website/docs/proxy/logging.md

@@ -17,6 +17,7 @@ Log Proxy Input, Output, Exceptions using Custom Callbacks, Langfuse, OpenTeleme
 - [Logging to Sentry](#logging-proxy-inputoutput---sentry)
 - [Logging to Traceloop (OpenTelemetry)](#logging-proxy-inputoutput-traceloop-opentelemetry)
 - [Logging to Athina](#logging-proxy-inputoutput-athina)
+- [Moderation with Azure Content-Safety](#moderation-with-azure-content-safety)
 
 ## Custom Callback Class [Async]
 Use this when you want to run custom callbacks in `python`
@@ -1003,4 +1004,87 @@ curl --location 'http://0.0.0.0:4000/chat/completions' \
         }
     ]
     }'
-```
+```
+
+## Moderation with Azure Content Safety
+
+[Azure Content-Safety](https://azure.microsoft.com/en-us/products/ai-services/ai-content-safety) is a Microsoft Azure service that provides content moderation APIs to detect potential offensive, harmful, or risky content in text.
+
+We will use the `--config` to set `litellm.success_callback = ["azure_content_safety"]` this will moderate all LLM calls using Azure Content Safety.
+
+**Step 0** Deploy Azure Content Safety
+
+Deploy an Azure Content-Safety instance from the Azure Portal and get the `endpoint` and `key`.
+
+**Step 1** Set Athina API key
+
+```shell
+AZURE_CONTENT_SAFETY_KEU = "<your-azure-content-safety-key>"
+```
+
+**Step 2**: Create a `config.yaml` file and set `litellm_settings`: `success_callback`
+```yaml
+model_list:
+  - model_name: gpt-3.5-turbo
+    litellm_params:
+      model: gpt-3.5-turbo
+litellm_settings:
+  callbacks: ["azure_content_safety"]
+  azure_content_safety_params:
+    endpoint: "<your-azure-content-safety-endpoint>"
+    key: "os.environ/AZURE_CONTENT_SAFETY_KEY"
+```
+
+**Step 3**: Start the proxy, make a test request
+
+Start proxy
+```shell
+litellm --config config.yaml --debug
+```
+
+Test Request
+```
+curl --location 'http://0.0.0.0:4000/chat/completions' \
+    --header 'Content-Type: application/json' \
+    --data ' {
+        "model": "gpt-3.5-turbo",
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hi, how are you?"
+            }
+        ]
+    }'
+```
+
+An HTTP 400 error will be returned if the content is detected with a value greater than the threshold set in the `config.yaml`.
+The details of the response will describe :
+- The `source` : input text or llm generated text
+- The `category` : the category of the content that triggered the moderation
+- The `severity` : the severity from 0 to 10
+
+**Step 4**: Customizing Azure Content Safety Thresholds
+
+You can customize the thresholds for each category by setting the `thresholds` in the `config.yaml`
+
+```yaml
+model_list:
+  - model_name: gpt-3.5-turbo
+    litellm_params:
+      model: gpt-3.5-turbo
+litellm_settings:
+  callbacks: ["azure_content_safety"]
+  azure_content_safety_params:
+    endpoint: "<your-azure-content-safety-endpoint>"
+    key: "os.environ/AZURE_CONTENT_SAFETY_KEY"
+    thresholds:
+      Hate: 6
+      SelfHarm: 8
+      Sexual: 6
+      Violence: 4
+```
+
+:::info
+`thresholds` are not required by default, but you can tune the values to your needs.
+Default values is `4` for all categories
+:::

litellm/proxy/hooks/azure_content_safety.py

@@ -59,16 +59,6 @@ class _PROXY_AzureContentSafety(
         except:
             pass
 
-    def _severity(self, severity):
-        if severity >= 6:
-            return "high"
-        elif severity >= 4:
-            return "medium"
-        elif severity >= 2:
-            return "low"
-        else:
-            return "safe"
-
     def _compute_result(self, response):
         result = {}
 
@@ -80,7 +70,7 @@ class _PROXY_AzureContentSafety(
             if severity is not None:
                 result[category] = {
                     "filtered": severity >= self.thresholds[category],
-                    "severity": self._severity(severity),
+                    "severity": severity,
                 }
 
         return result
@@ -148,10 +138,10 @@ class _PROXY_AzureContentSafety(
                 content=response.choices[0].message.content, source="output"
             )
 
-    #async def async_post_call_streaming_hook(
+    # async def async_post_call_streaming_hook(
     #    self,
     #    user_api_key_dict: UserAPIKeyAuth,
     #    response: str,
-    #):
+    # ):
     #    self.print_verbose(f"Inside Azure Content-Safety Call-Stream Hook")
     #    await self.test_violation(content=response, source="output")

litellm/tests/test_azure_content_safety.py

@@ -50,7 +50,7 @@ async def test_strict_input_filtering_01():
 
     assert exc_info.value.detail["source"] == "input"
     assert exc_info.value.detail["category"] == "Hate"
-    assert exc_info.value.detail["severity"] == "low"
+    assert exc_info.value.detail["severity"] == 2
 
 
 @pytest.mark.asyncio
@@ -168,7 +168,7 @@ async def test_strict_output_filtering_01():
 
     assert exc_info.value.detail["source"] == "output"
     assert exc_info.value.detail["category"] == "Hate"
-    assert exc_info.value.detail["severity"] == "low"
+    assert exc_info.value.detail["severity"] == 2
 
 
 @pytest.mark.asyncio