10915879151

Committed 18 Sep 2024 05:03AM UTC coverage: 31.855% (+0.005%) from 31.85%

Build # 10915879151

Build Type

Pull #9765

github

Committed by

s-cheng

Commit Message

fix migration, dont drop the old field yet

Pull Request Pull Request #9765: [EA Forum only] updates to the criticism tips bot, temporarily pause showing the card to users

Run Details

5077 of 41370 branches covered (12.27%)

Branch coverage included in aggregate %.

17 of 51 new or added lines in 5 files covered. (33.33%)

4 existing lines in 2 files now uncovered.

29452 of 67024 relevant lines covered (43.94%)

184.63 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

16.67

/packages/lesswrong/server/languageModels/criticismTipsBot.ts

import { getOpenAI } from './languageModelIntegration';
import { isAnyTest, isProduction } from '../../lib/executionEnvironment';
import sanitizeHtml from 'sanitize-html';
import { isEAForum } from '../../lib/instanceSettings';
import OpenAI from 'openai';
import { captureEvent } from '../../lib/analyticsEvents';
import type { PostIsCriticismRequest } from '../resolvers/postResolvers';
import { sanitizeHtmlOptions } from './modGPT';


const criticismTipsBotPrompt = `
  You are an assistant for the Effective Altruism Forum.
  Decide if the given post should be tagged as "Criticism of work in effective altruism". This tag applies to posts critically examining the work, projects, or methodologies of specific individuals, organizations, or initiatives affiliated with the effective altruism (EA) movement or community.
  Give your reasoning, referring to the above criteria. Then end your response with "yes" or "no" (one word, no quotes) indicating your overall judgement on whether the post should be tagged.
`

const checkIsCriticism = async (api: OpenAI, text: string) => {
  return await api.chat.completions.create({
    model: 'gpt-4o-mini',
    messages: [
      {role: 'system', content: criticismTipsBotPrompt},
      {role: 'user', content: text},
    ],
  })
}

/**
 * On the EA Forum, we're using an LLM to check if the post could be classified
 * as "criticism of work in effective altruism". We check while the editor is open,
 * because if this returns true, then we want to show the author a little card
 * with tips on how to make it more likely to go well (see PostsEditBotTips).
 */
export async function postIsCriticism(post: PostIsCriticismRequest, currentUserId?: string): Promise<boolean> {
  // Only run this on the EA Forum on production, since it costs money.
  if (!isEAForum || !isProduction) return false
  
  if (!post.body) {
    if (!isAnyTest) {
      //eslint-disable-next-line no-console
      console.log("Skipping criticism tips bot (no contents in this post)")
    }
    return false
  }

  const api = await getOpenAI();
  if (!api) {
    if (!isAnyTest) {
      //eslint-disable-next-line no-console
      console.log("Skipping criticism tips bot (API not configured)")
    }
    return false
  }
  
  const postBody = post.contentType === 'markdown' ? post.body : sanitizeHtml(post.body, sanitizeHtmlOptions)
  const userText = `${post.title}\n\n${postBody}`
  
  const analyticsData = {
    userId: currentUserId,
    postId: post._id,
    postTitle: post.title,
    postBody,
    fullMessage: userText,
  }

  try {
    const response = await checkIsCriticism(api, userText)
    const topResult = response.choices[0].message?.content
    if (!topResult) return false
    
    // We ask it to explain the decision, but the frontend only cares about the final decision
    const finalWord = topResult.trim().toLowerCase().split(/[\n\s]+/).pop()
    
    captureEvent("criticismTipsBotResponse", {
      ...analyticsData,
      analysis: topResult,
      decision: finalWord
    })
    return finalWord === "yes"
    
  } catch (error) {
    if (error instanceof OpenAI.APIError) {
      captureEvent("criticismTipsBotError", {
        ...analyticsData,
        status: error.status,
        error: error.message
      })
    } else {
      //eslint-disable-next-line no-console
      console.error(error)
    }
    return false
  }
}


1	import { getOpenAI } from './languageModelIntegration';	57✔
2	import { isAnyTest, isProduction } from '../../lib/executionEnvironment';	19✔
3	import sanitizeHtml from 'sanitize-html';	19✔
4	import { isEAForum } from '../../lib/instanceSettings';	19✔
5	import OpenAI from 'openai';	19✔
6	import { captureEvent } from '../../lib/analyticsEvents';	19✔
7	import type { PostIsCriticismRequest } from '../resolvers/postResolvers';
8	import { sanitizeHtmlOptions } from './modGPT';	19✔
9
10
11	const criticismTipsBotPrompt = `	19✔
12	You are an assistant for the Effective Altruism Forum.
13	Decide if the given post should be tagged as "Criticism of work in effective altruism". This tag applies to posts critically examining the work, projects, or methodologies of specific individuals, organizations, or initiatives affiliated with the effective altruism (EA) movement or community.
14	Give your reasoning, referring to the above criteria. Then end your response with "yes" or "no" (one word, no quotes) indicating your overall judgement on whether the post should be tagged.
15	`
16
17	const checkIsCriticism = async (api: OpenAI, text: string) => {	19✔
NEW 18	return await api.chat.completions.create({	×
19	model: 'gpt-4o-mini',
20	messages: [
21	{role: 'system', content: criticismTipsBotPrompt},
22	{role: 'user', content: text},
23	],
24	})
25	}
26
27	/**
28	* On the EA Forum, we're using an LLM to check if the post could be classified
29	* as "criticism of work in effective altruism". We check while the editor is open,
30	* because if this returns true, then we want to show the author a little card
31	* with tips on how to make it more likely to go well (see PostsEditBotTips).
32	*/
33	export async function postIsCriticism(post: PostIsCriticismRequest, currentUserId?: string): Promise<boolean> {
34	// Only run this on the EA Forum on production, since it costs money.
NEW 35	if (!isEAForum \|\| !isProduction) return false	×
36
NEW 37	if (!post.body) {	×
NEW 38	if (!isAnyTest) {	×
39	//eslint-disable-next-line no-console
NEW 40	console.log("Skipping criticism tips bot (no contents in this post)")	×
41	}
NEW 42	return false	×
43	}
44
NEW 45	const api = await getOpenAI();	×
NEW 46	if (!api) {	×
NEW 47	if (!isAnyTest) {	×
48	//eslint-disable-next-line no-console
NEW 49	console.log("Skipping criticism tips bot (API not configured)")	×
50	}
NEW 51	return false	×
52	}
53
NEW 54	const postBody = post.contentType === 'markdown' ? post.body : sanitizeHtml(post.body, sanitizeHtmlOptions)	×
NEW 55	const userText = `${post.title}\n\n${postBody}`	×
56
NEW 57	const analyticsData = {	×
58	userId: currentUserId,
59	postId: post._id,
60	postTitle: post.title,
61	postBody,
62	fullMessage: userText,
63	}
64
NEW 65	try {	×
NEW 66	const response = await checkIsCriticism(api, userText)	×
NEW 67	const topResult = response.choices[0].message?.content	×
NEW 68	if (!topResult) return false	×
69
70	// We ask it to explain the decision, but the frontend only cares about the final decision
NEW 71	const finalWord = topResult.trim().toLowerCase().split(/[\n\s]+/).pop()	×
72
NEW 73	captureEvent("criticismTipsBotResponse", {	×
74	...analyticsData,
75	analysis: topResult,
76	decision: finalWord
77	})
NEW 78	return finalWord === "yes"	×
79
80	} catch (error) {
NEW 81	if (error instanceof OpenAI.APIError) {	×
NEW 82	captureEvent("criticismTipsBotError", {	×
83	...analyticsData,
84	status: error.status,
85	error: error.message
86	})
87	} else {
88	//eslint-disable-next-line no-console
NEW 89	console.error(error)	×
90	}
NEW 91	return false	×
92	}
93	}
94

ForumMagnum / ForumMagnum / 10915879151

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous