chore: import upstream snapshot with attribution

plugins/knowledge/eslint.config.mjs

@@ -0,0 +1,13 @@
+import rootConfig from '../../eslint.config.mjs'
+
+export default [
+  ...rootConfig,
+  {
+    languageOptions: {
+      parserOptions: {
+        project: ['./tsconfig.json'],
+        tsconfigRootDir: import.meta.dirname,
+      },
+    },
+  },
+]

plugins/knowledge/package.json

@@ -0,0 +1,27 @@
+{
+  "name": "@botpresshub/knowledge",
+  "scripts": {
+    "check:type": "tsc --noEmit",
+    "build": "bp add -y && bp build",
+    "test": "vitest --run"
+  },
+  "private": true,
+  "dependencies": {
+    "@botpress/client": "workspace:*",
+    "@botpress/sdk": "workspace:*",
+    "json5": "^2.2.3",
+    "jsonrepair": "^3.10.0"
+  },
+  "devDependencies": {
+    "@botpress/cli": "workspace:*",
+    "@botpress/common": "workspace:*",
+    "@botpress/sdk": "workspace:*",
+    "@botpresshub/llm": "workspace:*",
+    "@bpinternal/genenv": "0.0.1",
+    "@types/semver": "^7.3.11",
+    "semver": "^7.3.8"
+  },
+  "bpDependencies": {
+    "llm": "../../interfaces/llm"
+  }
+}

plugins/knowledge/plugin.definition.ts

@@ -0,0 +1,11 @@
+import * as sdk from '@botpress/sdk'
+import llm from './bp_modules/llm'
+
+export default new sdk.PluginDefinition({
+  name: 'knowledge',
+  version: '1.0.0',
+  configuration: { schema: sdk.z.object({}) },
+  interfaces: {
+    llm: sdk.version.allWithinMajorOf(llm),
+  },
+})

plugins/knowledge/src/generate-content.ts

@@ -0,0 +1,31 @@
+import JSON5 from 'json5'
+import { jsonrepair } from 'jsonrepair'
+import * as bp from '.botpress'
+
+export type LLMInput = bp.interfaces.llm.actions.generateContent.input.Input
+export type LLMOutput = bp.interfaces.llm.actions.generateContent.output.Output
+
+export type LLMMessage = LLMInput['messages'][number]
+export type LLMChoice = LLMOutput['choices'][number]
+
+type PredictResponse = {
+  success: boolean
+  json: object
+}
+
+const tryParseJson = (str: string) => {
+  try {
+    return JSON5.parse(jsonrepair(str))
+  } catch {
+    return str
+  }
+}
+
+export const parseLLMOutput = (output: LLMOutput): PredictResponse => {
+  const mappedChoices: LLMChoice['content'][] = output.choices.map((choice) => choice.content)
+  const firstChoice = mappedChoices[0]!
+  return {
+    success: true,
+    json: tryParseJson(firstChoice as string),
+  }
+}

plugins/knowledge/src/index.ts

@@ -0,0 +1,72 @@
+import * as gen from './generate-content'
+import * as questions from './question-prompt'
+import * as bp from '.botpress'
+
+const plugin = new bp.Plugin({
+  actions: {},
+})
+
+plugin.on.beforeIncomingMessage('*', async ({ data: message, client, ctx, actions }) => {
+  if (message.type !== 'text') {
+    console.debug('Ignoring non-text message')
+    return
+  }
+
+  const text: string = message.payload.text
+  if (!text) {
+    console.debug('Ignoring empty message')
+    return
+  }
+
+  console.debug('Extracting questions from:', text)
+
+  const llmInput = questions.prompt({ text, line: 'L1' })
+  const llmOutput = await actions.llm.generateContent(llmInput)
+
+  const { success, json } = gen.parseLLMOutput(llmOutput)
+  if (!success) {
+    console.debug('Failed to extract questions')
+    return
+  }
+
+  const parsedResult = questions.OutputFormat.safeParse(json)
+  if (!parsedResult.success) {
+    console.debug('Failed to extract questions')
+    return
+  }
+
+  const { data } = parsedResult
+  if (!data.hasQuestions || !data.questions?.length) {
+    console.debug('No questions extracted')
+    return
+  }
+
+  const canonicalQuestion = data.questions.map((question) => question.resolved_question).join(' ')
+
+  console.debug('Searching for:', canonicalQuestion)
+  const { passages } = await client.searchFiles({
+    query: canonicalQuestion,
+  })
+
+  if (!passages.length) {
+    console.debug('No passages found')
+    return
+  }
+
+  // TODO: replace by proper answer generation
+  const answer = passages.map((p) => p.content).join('\n')
+
+  await client.createMessage({
+    conversationId: message.conversationId,
+    userId: ctx.botId,
+    payload: {
+      text: answer,
+    },
+    tags: {},
+    type: 'text',
+  })
+
+  return { stop: true }
+})
+
+export default plugin

plugins/knowledge/src/question-prompt.ts

@@ -0,0 +1,255 @@
+import { z } from '@botpress/sdk'
+import { LLMInput, LLMMessage } from './generate-content'
+
+export type ExtractedQuestion = z.infer<typeof ExtractedQuestion>
+export const ExtractedQuestion = z.object({
+  line: z.string().describe('The line number of the question (must be prefixed with "L")'),
+  raw_question: z.string().describe('The raw question extracted from the user message'),
+  resolved_question: z.string().describe('The resolved question with any missing context filled in'),
+  search_query: z.string().describe('The search query that would be used to find the answer in a search engine'),
+})
+
+export type OutputFormat = z.infer<typeof OutputFormat>
+export const OutputFormat = z.object({
+  hasQuestions: z.boolean().describe('Whether or not questions were found in the user message'),
+  questions: z
+    .array(ExtractedQuestion)
+    .describe('List of extracted questions, or an empty array if no questions are found')
+    .optional(),
+})
+
+type Example = {
+  output: Array<Omit<ExtractedQuestion, 'line'>>
+  input: {
+    context: string
+    user_message: string
+  }
+}
+
+const formatQuestionExtractorMessage = (
+  text: string,
+  context: string
+): { user_message_line: string; message: string } => {
+  const contextLines = context.split('\n').map((line, index) => `[L${index + 1}]\t ${line}`)
+  const user_message_line = `L${contextLines.length + 1}`
+
+  return {
+    user_message_line,
+    message: `
+<CONTEXT>
+${contextLines.join('\n')}
+</CONTEXT>
+
+<USER MESSAGE>
+[${user_message_line}]\t ${text}
+</USER MESSAGE>`.trim(),
+  }
+}
+
+const makeExample = (props: Example): LLMMessage[] => {
+  const { message, user_message_line } = formatQuestionExtractorMessage(props.input.user_message, props.input.context)
+  return [
+    {
+      role: 'user',
+      content: message,
+    },
+    {
+      role: 'assistant',
+      content: JSON.stringify({
+        hasQuestions: props.output.length > 0,
+        questions: props.output.map((o) => ({
+          line: user_message_line,
+          raw_question: o.raw_question,
+          resolved_question: o.resolved_question,
+          search_query: o.search_query,
+        })),
+      } satisfies OutputFormat),
+    },
+  ]
+}
+
+export type PromptArgs = {
+  text: string
+  line: string
+}
+export const prompt = (args: PromptArgs): LLMInput => ({
+  responseFormat: 'json_object',
+  temperature: 0,
+  systemPrompt: `
+You are a question extractor.
+You will be given a USER MESSAGE and a CONTEXT of the conversation so far.
+The CONTEXT should not be analyzed, only the USER MESSAGE. The purpose of the CONTEXT is to provide context for the USER MESSAGE.
+
+Your goal is to respond with the a list of question(s) found in the USER MESSAGE, if any.
+For the purpose of this task, a question is defined as any sentence that is asking for information or is seeking an answer. This includes direct questions, indirect questions, rhetorical questions, search engine queries etc.
+
+Include the raw_question (with no modifications), the resolved_question (with any missing context filled in), and the search_query (the query that would be used to find the answer in a search engine). If no relevant context is found, the resolved_question should be the same as the raw_question.
+Questions should be extracted with the following JSON format:
+
+[
+  {
+    "line": 'L34',
+    "raw_question": "how old is he",
+    "resolved_question": "how old is he (Justin Timberlake)?",
+    "search_query": "\"Justin Timberlake\" age",
+  }
+]
+
+If there are no questions in the USER MESSAGE, return an empty array.
+
+Always respond in JSON with the following format:
+type OutputFormat = ${ExtractedQuestion.toTypescriptType({ treatDefaultAsOptional: true })}
+
+The below examples are for illustrative purposes only. Your responses will be evaluated based on the quality of the questions extracted.
+Please extract the questions found on line L${args.line} only. If there are no questions, return an empty array.
+`.trim(),
+  messages: [
+    ...makeExample({
+      input: {
+        context: `
+Summary:
+"""
+The conversation so far is about the population of the United States.
+"""
+
+Transcript:
+"""
+user: how many people live in the united states?
+bot: The population of the United States is 331 million people.
+"""`,
+        user_message: 'tell me the same for canada, japan and china',
+      },
+      output: [
+        {
+          raw_question: 'tell me the same for canada, japan and china',
+          resolved_question: 'tell me the population for Canada, Japan and China',
+          search_query: 'current population of Canada, Japan, China',
+        },
+      ],
+    }),
+    ...makeExample({
+      input: {
+        context: `
+Summary:
+"""
+The conversation so far is about organizing an event for a company called Blue Bridge.
+"""
+
+Transcript:
+"""
+user: the name of the event is Night of the Stars
+bot: Got it, who should I send invitations to?
+"""`,
+        user_message: "hmmm i'm not sure, list people in marketing and sales and i'll tell you",
+      },
+      output: [
+        {
+          raw_question: 'list people in marketing and sales',
+          resolved_question: 'list people in marketing and sales (at Blue Bridge)',
+          search_query: '"Blue Bridge" sales and marketing employee directory',
+        },
+      ],
+    }),
+    ...makeExample({
+      input: {
+        context: '',
+        user_message: 'what is ptow',
+      },
+      output: [
+        {
+          raw_question: 'what is ptow',
+          resolved_question: 'what is ptow?',
+          search_query: 'ptow meaning',
+        },
+      ],
+    }),
+    ...makeExample({
+      input: {
+        context: '',
+        user_message: 'sure? lets do it',
+      },
+      output: [],
+    }),
+    ...makeExample({
+      input: {
+        context: `
+User Information:
+"""
+Name: Alex
+Language: English
+Location: United States
+City: New York
+"""
+
+Transcript:
+"""
+user: hello!
+bot: Hello! How can I help you today?
+"""`,
+        user_message: 'what is the weather?',
+      },
+      output: [
+        {
+          raw_question: 'what is the weather?',
+          resolved_question: 'what is the weather (in New York)?',
+          search_query: 'current weather in New York',
+        },
+      ],
+    }),
+    ...makeExample({
+      input: {
+        context: `
+Transcript:
+"""
+user: hello!
+bot: Hello! How can I help you today?
+user: can you give me a list of the top 10 movies?
+"""`,
+        user_message: 'sorry, make it 3 after all',
+      },
+      output: [],
+    }),
+    ...makeExample({
+      input: {
+        context: `
+Transcript:
+"""
+"""`,
+        user_message: 'How tall is Patrick?',
+      },
+      output: [
+        {
+          raw_question: 'How tall is Patrick',
+          resolved_question: 'How tall is Patrick',
+          search_query: 'How tall is Patrick',
+        },
+      ],
+    }),
+    ...makeExample({
+      input: {
+        context: `
+Transcript:
+"""
+user: tell me about cindy
+bot: Cindy is 45 years old and is currently a software engineer at Google.
+user: what is her salary?
+bot: She earned $142,000 per year in 2022.
+user: what about her husband?
+bot: Her husband, Danny, is a doctor.
+"""`,
+        user_message: 'I see! I think I met him once at a party. I wonder if he make more than her',
+      },
+      output: [
+        {
+          raw_question: 'I wonder if he make more than her',
+          resolved_question: 'I wonder if Danny makes more than 142,000 per year?',
+          search_query: 'Danny (doctor) salary comparison to Cindy (software engineer)',
+        },
+      ],
+    }),
+    {
+      role: 'user',
+      content: args.text,
+    },
+  ],
+})

plugins/knowledge/tsconfig.json

@@ -0,0 +1,8 @@
+{
+  "extends": "../../tsconfig.json",
+  "compilerOptions": {
+    "paths": { "*": ["./*"] },
+    "outDir": "dist"
+  },
+  "include": [".botpress/**/*", "definitions/**/*", "src/**/*", "*.ts"]
+}

plugins/knowledge/vitest.config.ts

@@ -0,0 +1,2 @@
+import config from '../../vitest.config'
+export default config