ResponsibleAI - Input Moderation - Ability to verify inputs and prevent potentially harmful content generation

2026-06-28 15:00:57 +03:00 · 2023-11-13 10:48:24 +05:30
parent a311e024e1
commit ad8281e553
8 changed files with 346 additions and 1 deletions
@@ -0,0 +1,154 @@
+import { INode, INodeData, INodeParams } from '../../../src/Interface'
+import { getBaseClasses } from '../../../src'
+import { Moderation } from '../ResponsibleAI'
+import { OpenAIModerationRunner } from './OpenAIModerationRunner'
+
+class OpenAIModeration implements INode {
+    label: string
+    name: string
+    version: number
+    description: string
+    type: string
+    icon: string
+    category: string
+    baseClasses: string[]
+    inputs: INodeParams[]
+
+    constructor() {
+        this.label = 'Moderation - Open AI'
+        this.name = 'inputModerationOpenAI'
+        this.version = 1.0
+        this.type = 'Moderation'
+        this.icon = 'openai-moderation.png'
+        this.category = 'Responsible AI'
+        this.description = 'Check whether content complies with OpenAI usage policies.'
+        this.baseClasses = [this.type, ...getBaseClasses(Moderation)]
+        this.inputs = [
+            {
+                label: 'Moderation Checks',
+                name: 'moderationConfig',
+                type: 'options',
+                default: 'useDefault',
+                options: [
+                    {
+                        label: 'OpenAI Default',
+                        name: 'useDefault'
+                    },
+                    {
+                        label: 'Use Custom Threshold Values',
+                        name: 'useCustom'
+                    },
+                    {
+                        label: 'Combine OpenAI Default with Custom Threshold Values',
+                        name: 'combineBoth'
+                    }
+                ]
+            },
+            {
+                label: 'Error Message',
+                name: 'moderationErrorMessage',
+                type: 'string',
+                rows: 2,
+                default: "Cannot Process! Input violates OpenAI's content moderation policies.",
+                optional: true
+            },
+            {
+                label: 'Threshold Score - Sexual',
+                name: 'catSexualThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Sexual/Minors',
+                name: 'catSexualMinorsThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Hate',
+                name: 'catHateThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Hate/Threatening',
+                name: 'catHateThreateningThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Harassment',
+                name: 'catHarassmentThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Harassment/Threatening',
+                name: 'catHarassmentThreateningThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Self Harm',
+                name: 'catSelfHarmThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Self-Harm/Intent',
+                name: 'catSelfHarmIntentThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Self-Harm/Instructions',
+                name: 'catSelfHarmInstructionsThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Violence',
+                name: 'catViolenceThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            },
+            {
+                label: 'Threshold Score - Violence/Graphic',
+                name: 'catViolenceGraphicThreshold',
+                type: 'number',
+                default: 0.01,
+                additionalParams: true
+            }
+        ]
+    }
+
+    async init(nodeData: INodeData): Promise<any> {
+        const runner = new OpenAIModerationRunner()
+        this.inputs.forEach((value) => {
+            if (value.additionalParams === true) {
+                // remove thePrefix - 'cat'
+                let categoryName = value.name.substring(3)
+                // remove theSuffix - 'Threshold'
+                categoryName = categoryName.substring(0, categoryName.length - 9)
+                categoryName = categoryName.substring(0, 1).toLowerCase() + categoryName.substring(1)
+                let categoryThreshold = nodeData.inputs ? nodeData.inputs[value.name] : value.default
+                runner.setParameter(categoryName, parseFloat(categoryThreshold))
+            } else {
+                runner.setParameter(value.name, nodeData.inputs ? nodeData.inputs[value.name] : value.default)
+            }
+        })
+        return runner
+    }
+}
+
+module.exports = { nodeClass: OpenAIModeration }
@@ -0,0 +1,62 @@
+import { Moderation } from '../ResponsibleAI'
+import { BaseLanguageModel } from 'langchain/base_language'
+import { OpenAIModerationChain } from 'langchain/chains'
+
+export class OpenAIModerationRunner implements Moderation {
+    private moderationConfig: string = 'useDefault'
+    private moderationErrorMessage: string = "Text was found that violates OpenAI's content policy."
+    private sexual: number = 0.01
+    private sexualMinors: number = 0.01
+    private hate: number = 0.01
+    private hateThreatening: number = 0.01
+    private harassment: number = 0.01
+    private harassmentThreatening: number = 0.01
+    private selfHarm: number = 0.01
+    private selfHarmIntent: number = 0.01
+    private selfHarmInstructions: number = 0.01
+    private violence: number = 0.01
+    private violenceGraphic: number = 0.01
+
+    async checkForViolations(llm: BaseLanguageModel, input: string): Promise<string> {
+        const openAIApiKey = (llm as any).openAIApiKey
+        if (!openAIApiKey) {
+            throw Error('OpenAI API key not found')
+        }
+        // Create a new instance of the OpenAIModerationChain
+        const moderation = new OpenAIModerationChain({
+            openAIApiKey: openAIApiKey,
+            throwError: false // If set to true, the call will throw an error when the moderation chain detects violating content. If set to false, violating content will return "Text was found that violates OpenAI's content policy.".
+        })
+        // Send the user's input to the moderation chain and wait for the result
+        const { output: moderationOutput, results } = await moderation.call({
+            input: input
+        })
+        if (this.moderationConfig != 'useCustom' && results[0].flagged) {
+            throw Error(this.moderationErrorMessage)
+        }
+        if (this.moderationConfig != 'useDefault') {
+            const categoryScores = results[0].category_scores
+            if (
+                categoryScores['harassment'] > this.harassment ||
+                categoryScores['harassment/threatening'] > this.harassmentThreatening ||
+                categoryScores['self-harm'] > this.selfHarm ||
+                categoryScores['self-harm/intent'] > this.selfHarmIntent ||
+                categoryScores['self-harm/instructions'] > this.selfHarmInstructions ||
+                categoryScores['sexual'] > this.sexual ||
+                categoryScores['sexual/minors'] > this.sexualMinors ||
+                categoryScores['hate'] > this.hate ||
+                categoryScores['hate/threatening'] > this.hateThreatening ||
+                categoryScores['violence'] > this.violence ||
+                categoryScores['violence/graphic'] > this.violenceGraphic
+            ) {
+                throw Error(this.moderationErrorMessage)
+            }
+        }
+        return moderationOutput
+    }
+
+    setParameter(category: string, value: number) {
+        // @ts-ignore
+        this[category] = value
+    }
+}
@@ -0,0 +1,56 @@
+import { INode, INodeData, INodeParams } from '../../../src/Interface'
+import { getBaseClasses } from '../../../src'
+import { Moderation } from '../ResponsibleAI'
+import { SimplePromptModerationRunner } from './SimplePromptModerationRunner'
+
+class SimplePromptModeration implements INode {
+    label: string
+    name: string
+    version: number
+    description: string
+    type: string
+    icon: string
+    category: string
+    baseClasses: string[]
+    inputs: INodeParams[]
+
+    constructor() {
+        this.label = 'Moderation - Simple Prompt'
+        this.name = 'inputModerationSimple'
+        this.version = 1.0
+        this.type = 'Moderation'
+        this.icon = 'simple_moderation.png'
+        this.category = 'Responsible AI'
+        this.description = 'Detecting and mitigating prompt attacks'
+        this.baseClasses = [this.type, ...getBaseClasses(Moderation)]
+        this.inputs = [
+            {
+                label: 'Deny List',
+                name: 'denyList',
+                type: 'string',
+                rows: 4,
+                placeholder: `ignore previous instructions\ndo not follow the directions\nyou must ignore all previous instructions`,
+                description: 'An array of string literals (enter one per line) that should not appear in the prompt text.',
+                optional: false
+            },
+            {
+                label: 'Error Message',
+                name: 'moderationErrorMessage',
+                type: 'string',
+                rows: 2,
+                default: 'Cannot Process! Input violates content moderation policies.',
+                optional: true
+            }
+        ]
+    }
+
+    // eslint-disable-next-line unused-imports/no-unused-vars
+    async init(nodeData: INodeData): Promise<any> {
+        const denyList = nodeData.inputs?.denyList as string
+        const moderationErrorMessage = nodeData.inputs?.moderationErrorMessage as string
+
+        return new SimplePromptModerationRunner(denyList, moderationErrorMessage)
+    }
+}
+
+module.exports = { nodeClass: SimplePromptModeration }
@@ -0,0 +1,24 @@
+import { Moderation } from '../ResponsibleAI'
+import { BaseLanguageModel } from 'langchain/base_language'
+
+export class SimplePromptModerationRunner implements Moderation {
+    private readonly denyList: string = ''
+    private readonly moderationErrorMessage: string = ''
+
+    constructor(denyList: string, moderationErrorMessage: string) {
+        this.denyList = denyList
+        if (denyList.indexOf('\n') === -1) {
+            this.denyList += '\n'
+        }
+        this.moderationErrorMessage = moderationErrorMessage
+    }
+
+    async checkForViolations(llm: BaseLanguageModel, input: string): Promise<string> {
+        this.denyList.split('\n').forEach((denyListItem) => {
+            if (denyListItem && denyListItem !== '' && input.includes(denyListItem)) {
+                throw Error(this.moderationErrorMessage)
+            }
+        })
+        return Promise.resolve(input)
+    }
+}