feat: Add convenience function for LlamaGuard Filter (#555)

* chore: add new filter utility for LlamaGuard * chore: update default * chore: fix test * chore: cleanup * chore: refactor * chore: minor doc fix * chore: adapt changes as per slack discussion * chore: add orchestration client test * chore: add documentation * chore: add release note * fix: Changes from lint * chore: fox typo * chore: add mutiple filters in sample code * chore: fix typo * Update .changeset/hip-tools-smile.md Co-authored-by: Deeksha Sinha <[email protected]> * review * minor * fix: type tests * fix: Changes from lint * reivew * readme --------- Co-authored-by: cloud-sdk-js <[email protected]> Co-authored-by: Zhongpin Wang <[email protected]> Co-authored-by: Deeksha Sinha <[email protected]>
SAP · Feb 24, 2025 · bc51f59 · bc51f59
1 parent b1d244c
commit bc51f59
Show file tree

Hide file tree

Showing 10 changed files with 287 additions and 51 deletions.
diff --git a/.changeset/hip-tools-smile.md b/.changeset/hip-tools-smile.md
@@ -0,0 +1,5 @@
+---
+'@sap-ai-sdk/orchestration': minor
+---
+
+[New Functionality] Introduce `buildLlamaGuardFilter()` convenience function to build Llama guard filters.
diff --git a/packages/orchestration/README.md b/packages/orchestration/README.md
@@ -336,20 +336,12 @@ Use the orchestration client with filtering to restrict content that is passed t
 
 This feature allows filtering both the [input](https://help.sap.com/docs/sap-ai-core/sap-ai-core-service-guide/input-filtering) and [output](https://help.sap.com/docs/sap-ai-core/sap-ai-core-service-guide/output-filtering) of a model based on content safety criteria.
 
-#### Azure Content Filter
-
-Use `buildAzureContentSafetyFilter()` function to build an Azure content filter for both input and output.
-Each category of the filter can be assigned a specific severity level, which corresponds to an Azure threshold value.
-
-| Severity Level          | Azure Threshold Value |
-| ----------------------- | --------------------- |
-| `ALLOW_SAFE`            | 0                     |
-| `ALLOW_SAFE_LOW`        | 2                     |
-| `ALLOW_SAFE_LOW_MEDIUM` | 4                     |
-| `ALLOW_ALL`             | 6                     |
+The following example demonstrates how to use content filtering with the orchestration client.
+See the sections below for details on the available content filters and how to build them.
 
 ```ts
-import { OrchestrationClient, ContentFilters } from '@sap-ai-sdk/orchestration';
+import { OrchestrationClient } from '@sap-ai-sdk/orchestration';
+
 const llm = {
   model_name: 'gpt-4o',
   model_params: { max_tokens: 50, temperature: 0.1 }
@@ -358,16 +350,14 @@ const templating = {
   template: [{ role: 'user', content: '{{?input}}' }]
 };
 
-const filter = buildAzureContentSafetyFilter({
-  Hate: 'ALLOW_SAFE_LOW',
-  Violence: 'ALLOW_SAFE_LOW_MEDIUM'
-});
+const filter = ... // Use a build function to create a content filter
+
 const orchestrationClient = new OrchestrationClient({
   llm,
   templating,
   filtering: {
     input: {
-      filters: [filter]
+      filters: [filter] // Multiple filters can be applied
     },
     output: {
       filters: [filter]
@@ -385,6 +375,38 @@ try {
 }
 ```
 
+Multiple filters can be applied at the same time for both input and output filtering.
+
+#### Azure Content Filter
+
+Use `buildAzureContentSafetyFilter()` function to build an Azure content filter.
+Each category of the filter can be assigned a specific severity level, which corresponds to an Azure threshold value.
+
+| Severity Level          | Azure Threshold Value |
+| ----------------------- | --------------------- |
+| `ALLOW_SAFE`            | 0                     |
+| `ALLOW_SAFE_LOW`        | 2                     |
+| `ALLOW_SAFE_LOW_MEDIUM` | 4                     |
+| `ALLOW_ALL`             | 6                     |
+
+```ts
+const filter = buildAzureContentSafetyFilter({
+  Hate: 'ALLOW_SAFE_LOW',
+  Violence: 'ALLOW_SAFE_LOW_MEDIUM'
+});
+```
+
+#### Llama Guard Filter
+
+Use `buildLlamaGuardFilter()` function to build a Llama Guard content filter.
+
+Available categories can be found with autocompletion.
+Pass the categories as arguments to the function to enable them.
+
+```ts
+const filter = buildLlamaGuardFilter('hate', 'violent_crimes');
+```
+
 #### Error Handling
 
 Both `chatCompletion()` and `getContent()` methods can throw errors.

diff --git a/packages/orchestration/src/index.ts b/packages/orchestration/src/index.ts
@@ -9,6 +9,7 @@ export type {
   DocumentGroundingServiceConfig,
   DocumentGroundingServiceFilter,
   LlmModelParams,
+  LlamaGuardCategory,
   AzureContentFilter,
   AzureFilterThreshold
 } from './orchestration-types.js';
@@ -24,6 +25,7 @@ export { OrchestrationClient } from './orchestration-client.js';
 export {
   buildAzureContentFilter,
   buildAzureContentSafetyFilter,
+  buildLlamaGuardFilter,
   buildDocumentGroundingConfig
 } from './util/index.js';
 

diff --git a/packages/orchestration/src/orchestration-client.test.ts b/packages/orchestration/src/orchestration-client.test.ts
@@ -13,7 +13,8 @@ import { OrchestrationResponse } from './orchestration-response.js';
 import {
   constructCompletionPostRequestFromJsonModuleConfig,
   constructCompletionPostRequest,
-  buildAzureContentSafetyFilter
+  buildAzureContentSafetyFilter,
+  buildLlamaGuardFilter
 } from './util/index.js';
 import type { CompletionPostResponse } from './client/api/schema/index.js';
 import type {
@@ -206,6 +207,59 @@ describe('orchestration service client', () => {
     expect(response.data).toEqual(mockResponse);
   });
 
+  it('calls chatCompletion with filter configuration supplied using multiple convenience functions', async () => {
+    const llamaFilter = buildLlamaGuardFilter('self_harm');
+    const azureContentFilter = buildAzureContentSafetyFilter({
+      Sexual: 'ALLOW_SAFE'
+    });
+    const config: OrchestrationModuleConfig = {
+      llm: {
+        model_name: 'gpt-35-turbo-16k',
+        model_params: { max_tokens: 50, temperature: 0.1 }
+      },
+      templating: {
+        template: [
+          {
+            role: 'user',
+            content: 'Create {{?number}} paraphrases of {{?phrase}}'
+          }
+        ]
+      },
+      filtering: {
+        input: {
+          filters: [llamaFilter, azureContentFilter]
+        },
+        output: {
+          filters: [llamaFilter, azureContentFilter]
+        }
+      }
+    };
+    const prompt = {
+      inputParams: { phrase: 'I like myself.', number: '20' }
+    };
+    const mockResponse = await parseMockResponse<CompletionPostResponse>(
+      'orchestration',
+      'orchestration-chat-completion-multiple-filter-config.json'
+    );
+
+    mockInference(
+      {
+        data: constructCompletionPostRequest(config, prompt)
+      },
+      {
+        data: mockResponse,
+        status: 200
+      },
+      {
+        url: 'inference/deployments/1234/completion'
+      }
+    );
+    const response = await new OrchestrationClient(config).chatCompletion(
+      prompt
+    );
+    expect(response.data).toEqual(mockResponse);
+  });
+
   it('calls chatCompletion with filtering configuration', async () => {
     const config: OrchestrationModuleConfig = {
       llm: {

diff --git a/packages/orchestration/src/orchestration-types.ts b/packages/orchestration/src/orchestration-types.ts
@@ -8,6 +8,7 @@ import type {
   FilteringStreamOptions,
   GlobalStreamOptions,
   GroundingModuleConfig,
+  LlamaGuard38B,
   MaskingModuleConfig,
   LlmModuleConfig as OriginalLlmModuleConfig,
   TemplatingModuleConfig
@@ -201,3 +202,8 @@ export const supportedAzureFilterThresholds = {
  *
  */
 export type AzureFilterThreshold = keyof typeof supportedAzureFilterThresholds;
+
+/**
+ * The filter categories supported for Llama guard filter.
+ */
+export type LlamaGuardCategory = keyof LlamaGuard38B;
diff --git a/packages/orchestration/src/util/filtering.test.ts b/packages/orchestration/src/util/filtering.test.ts
@@ -1,6 +1,7 @@
 import {
   buildAzureContentFilter,
-  buildAzureContentSafetyFilter
+  buildAzureContentSafetyFilter,
+  buildLlamaGuardFilter
 } from './filtering.js';
 import { constructCompletionPostRequest } from './module-config.js';
 import type { OrchestrationModuleConfig } from '../orchestration-types.js';
@@ -211,4 +212,34 @@ describe('Content filter util', () => {
       );
     });
   });
+
+  describe('Llama Guard filter', () => {
+    it('builds filter config with custom config', async () => {
+      const filterConfig = buildLlamaGuardFilter('elections', 'hate');
+      const expectedFilterConfig = {
+        type: 'llama_guard_3_8b',
+        config: {
+          elections: true,
+          hate: true
+        }
+      };
+      expect(filterConfig).toEqual(expectedFilterConfig);
+    });
+
+    it('builds filter config without duplicates', async () => {
+      const filterConfig = buildLlamaGuardFilter(
+        'non_violent_crimes',
+        'privacy',
+        'non_violent_crimes'
+      );
+      const expectedFilterConfig = {
+        type: 'llama_guard_3_8b',
+        config: {
+          non_violent_crimes: true,
+          privacy: true
+        }
+      };
+      expect(filterConfig).toEqual(expectedFilterConfig);
+    });
+  });
 });
diff --git a/packages/orchestration/src/util/filtering.ts b/packages/orchestration/src/util/filtering.ts
@@ -3,15 +3,17 @@ import type {
   AzureContentSafety,
   AzureContentSafetyFilterConfig,
   InputFilteringConfig,
+  LlamaGuard38BFilterConfig,
   OutputFilteringConfig
 } from '../client/api/schema/index.js';
 import type {
   AzureContentFilter,
-  AzureFilterThreshold
+  AzureFilterThreshold,
+  LlamaGuardCategory
 } from '../orchestration-types.js';
 
 /**
- * Convenience function to create Azure content filters.
+ * Convenience function to build Azure content filter.
  * @param filter - Filtering configuration for Azure filter. If skipped, the default Azure content filter configuration is used.
  * @returns An object with the Azure filtering configuration.
  * @deprecated Since 1.8.0. Use {@link buildAzureContentSafetyFilter()} instead.
@@ -33,10 +35,11 @@ export function buildAzureContentFilter(
 }
 
 /**
- * Convenience function to create Azure content filters.
+ * Convenience function to build Azure content filter.
  * @param config - Configuration for Azure content safety filter.
  * If skipped, the default configuration of `ALLOW_SAFE_LOW` is used for all filter categories.
  * @returns Filter config object.
+ * @example "buildAzureContentSafetyFilter({ Hate: 'ALLOW_SAFE', Violence: 'ALLOW_SAFE_LOW_MEDIUM' })"
  */
 export function buildAzureContentSafetyFilter(
   config?: AzureContentFilter
@@ -58,3 +61,20 @@ export function buildAzureContentSafetyFilter(
     })
   };
 }
+
+/**
+ * Convenience function to build Llama guard filter.
+ * @param categories - Categories to be enabled for filtering. Provide at least one category.
+ * @returns Filter config object.
+ * @example "buildLlamaGuardFilter('self_harm', 'hate')"
+ */
+export function buildLlamaGuardFilter(
+  ...categories: [LlamaGuardCategory, ...LlamaGuardCategory[]]
+): LlamaGuard38BFilterConfig {
+  return {
+    type: 'llama_guard_3_8b',
+    config: Object.fromEntries(
+      [...categories].map(category => [category, true])
+    )
+  };
+}