10624770467

Committed 30 Aug 2024 12:38AM UTC coverage: 31.92% (+0.03%) from 31.895%

Build # 10624770467

Build Type

Pull #9679

github

Committed by

b0b3rt

Commit Message

remove unused hook

Pull Request Pull Request #9679: limit input tokens and handle errors via streamgs

Run Details

5023 of 41076 branches covered (12.23%)

Branch coverage included in aggregate %.

10 of 106 new or added lines in 4 files covered. (9.43%)

36 existing lines in 3 files now uncovered.

29342 of 66584 relevant lines covered (44.07%)

188.18 hits per line

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

15.03

/packages/lesswrong/server/languageModels/promptUtils.ts

import { userGetDisplayName } from "@/lib/collections/users/helpers"
import { htmlToMarkdown } from "../editor/conversionUtils";
import { CommentTreeNode, unflattenComments } from "@/lib/utils/unflatten";
import { z } from "zod";
import { zodResponseFormat } from "openai/helpers/zod";

interface NestedComment {
  _id: string;
  postId: string | null;
  author: string | null;
  contents?: string;
  karmaScore: number;
}

interface TokenCounter {
  tokenCount: number;
}

// Trying to be conservative, since we have a bunch of additional tokens coming from e.g. JSON.stringify
const CHARS_PER_TOKEN = 3.5;

export const documentToMarkdown = (document: PostsPage | DbComment | null) => {
  const html = document?.contents?.html;
  if (!html) {
    return undefined;
  }

  return htmlToMarkdown(html);
}

const mergeSortedArrays = (queue: CommentTreeNode<NestedComment>[], children: CommentTreeNode<NestedComment>[]): CommentTreeNode<NestedComment>[] => {
  const merged: CommentTreeNode<NestedComment>[] = [];
  let queueIdx = 0, childrenIdx = 0;

  while (queueIdx < queue.length && childrenIdx < children.length) {
    if (queue[queueIdx].item.karmaScore >= children[childrenIdx].item.karmaScore) {
      merged.push(queue[queueIdx]);
      queueIdx++;
    } else {
      merged.push(children[childrenIdx]);
      childrenIdx++;
    }
  }

  // Add any remaining elements
  return merged
    .concat(queue.slice(queueIdx))
    .concat(children.slice(childrenIdx));
}

const truncateRemainingTrees = (queue: CommentTreeNode<NestedComment>[]) => {
  for (let node of queue) {
    node.children = [];
  }
}

/**
 * Does a greedy karma-sorted breadth-first traversal over all the comment branches, and truncates them at the point where we estimate that we hit the token limit
 * i.e. we take the highest karma comment from all the queue, increment token count, put its children into the queue, resort, do the operation again
 * This means that we never have any gaps in comment branches.
 * It does pessimize against comments that are children of lower-karma comments, but doing lookahead is annoying and that shouldn't be a problem most of the time.
 * It might turn out that we want to prioritize full comment branches, in which case this will need to be replaced with a depth-first thing.
 */
const filterCommentTrees = (trees: CommentTreeNode<NestedComment>[], tokenCounter: TokenCounter) => {
  const tokenThreshold = 150_000;
  
  // Initialize the queue with root nodes, sorted by karma
  let queue: CommentTreeNode<NestedComment>[] = trees.sort((a, b) => b.item.karmaScore - a.item.karmaScore);

  while (queue.length > 0) {
    const node = queue.shift()!;
    const nodeTokens = (node.item.contents ?? '').length / CHARS_PER_TOKEN;

    if (tokenCounter.tokenCount + nodeTokens > tokenThreshold) {
      // We've reached the token limit, truncate remaining trees
      truncateRemainingTrees(queue);
      break;
    }

    tokenCounter.tokenCount += nodeTokens;

    // Sort children and merge them with the existing sorted queue
    if (node.children.length > 0) {
      const sortedChildren = node.children.sort((a, b) => b.item.karmaScore - a.item.karmaScore);
      queue = mergeSortedArrays(queue, sortedChildren);
    }
  }
}

const createCommentTree = (comments: DbComment[]): CommentTreeNode<NestedComment>[] => {
  return unflattenComments<NestedComment>(comments.map(comment => {
    const { baseScore, contents, ...rest } = comment;
    return {
      ...rest,
      karmaScore: baseScore,
      contents: documentToMarkdown(comment)
    };
  }));
}

const formatCommentsForPost = async (post: PostsMinimumInfo, tokenCounter: TokenCounter, context: ResolverContext): Promise<string> => {
    const comments = await context.Comments.find({postId: post._id}).fetch()
    if (!comments.length) {
      return ""
    }

    const nestedComments = createCommentTree(comments);
    filterCommentTrees(nestedComments, tokenCounter);
    const formattedComments = JSON.stringify(nestedComments);

    return `Comments for the post titled "${post.title}" with postId "${post._id}" are below.  Note that the comments are threaded (i.e. branching) and are formatted as a nested JSON structure for readability.  The threads of conversation (back and forth responses) might be relevant for answering some questions.

<comments>${formattedComments}</comments>`;
  }

const formatPostForPrompt = (post: PostsPage): string => {
  const authorName = userGetDisplayName(post.user)
  const markdown = documentToMarkdown(post)

  return `postId: ${post._id}
Title: ${post.title}
Author: ${authorName}
Publish Date: ${post.postedAt}
Score: ${post.baseScore}
Content: ${markdown}`;
}

const formatAdditionalPostsForPrompt = (posts: PostsPage[], tokenCounter: TokenCounter): string => {
  const formattedPosts = posts.map(post => formatPostForPrompt(post));
  const includedPosts: string[] = [];

  for (let [idx, formattedPost] of Object.entries(formattedPosts)) {
    const approximatePostTokenCount = formattedPost.length / CHARS_PER_TOKEN;
    const postInclusionTokenCount = tokenCounter.tokenCount + approximatePostTokenCount;
    // Include at least one additional post, unless that takes us over the higher threshold
    if (idx !== '0' && postInclusionTokenCount > 120_000) {
      break;
    }

    if (idx === '0' && postInclusionTokenCount > 140_000) {
      break;
    }

    includedPosts.push(formattedPost);
    tokenCounter.tokenCount += approximatePostTokenCount;
  }

  return includedPosts.map((post, index) => `Supplementary Post #${index}:\n${post}`).join('\n')
}

export const generateLoadingMessagePrompt = (query: string, postTitle?: string): string => {
  return [
    'I need you to generate some humorous "loading messages" to display to users of the LessWrong.com Claude chat integration since it can take 10-30 seconds to load new responses',
    'Your responses may make general humorous reference to LessWrong, but it is even better if they are tailored to the specific query or post the user is viewing.',
    `The user has asked the following question: "${query}"`,
    postTitle ? `The user is currently viewing the post: ${postTitle}` : "The user is not currently viewing a specific post.",
    'Please generate 5 humorous loading messages that could be displayed to the user while they wait for a response.',
  ].join('\n')
}

export const generateTitleGenerationPrompt = (query: string, currentPost: PostsPage | null): string => {
  const currentPostContextLine = currentPost
    ? `The user is currently viewing a post titled "${currentPost.title}". Reference it if relevant.`
    : '';

    return `A user has started a new converation with you, Claude.  Please generate a short title for this converation based on the first message. The first message is as follows: <message>${query}</message>

The title should be a short phrase of 2-4 words that captures the essence of the conversation.  Do not wrap your answer in quotes or brackets. Do not include the word "title" or similar in your response.  ${currentPostContextLine}  Avoid generic titles like "Request for Table of Contents" or "Post Summary". Prefer to reference the specific post or topic being discussed.`;
}

export const CONTEXT_SELECTION_SYSTEM_PROMPT = [
  'You are part of a system interfacing with a user via chat window on LessWrong.com.',
  'Your responsibility is to make decisions about whether or not to load LessWrong posts as additional context for responding to user queries, and what strategy to to employ.'
].join('\n');

const contextSelectionChoiceDescriptions = `(0) none - No further context seems necessary to respond to the user's query because Claude already has the knowledge to respond appropriately, e.g. the user asked "What is Jacobian of a matrix?"or "Proofread the following text." Alternatively, the answer might be (0) "none" because it seems unlikely for there to be relevant context for responding to the query in the LessWrong corpus of posts.

(1) query-based - Load LessWrong posts based on their vector similarity to the user's query, and ignore the post the user is reading.This is correct choice if the query seems unrelated to the post the user is currently viewing, but it seems likely that there are LessWrong posts concerning the topic.(If it is a very general question, the correct choice might be (0) "none").

(2) current-post-based - Load LessWrong posts based on their vector similarity to the post the user is reading, but not the query.This is the correct choice if the query seems to be about the post the user is currently viewing, and pulling up other LessWrong posts related to the user's query would either be redundant with a search based on the current post, or would return irrelevant results.Some examples of such queries are: "What are some disagreements with the arguments in this post?", "Explain <topic in the post> to me.", and "Provide a summary of this post."

(3) both - Load LessWrong posts based on their vector similarity to both the user's query and the post the user is reading.This is the correct choice if the question seems to be related to the post the user is currently viewing, but also contains keywords or information where relevant LessWrong posts would be beneficial context for a response, and those LessWrong posts would not likely be returned in a vector similarity search based on the post the user is currently viewing.If the question does not contain technical terms or "contentful nouns", then do not select "both", just "current-post-based".`;

const ContextSelectionParameters = z.object({
  reasoning: z.string().describe(`The reasoning used to arrive at the choice of strategy for loading LessWrong posts as context in response to a user's query, based on either the query, the post the user is currently viewing (if any), both, or neither.`),
  strategy_choice: z.union([z.literal('none'), z.literal('query-based'), z.literal('current-post-based'), z.literal('both')]).describe(contextSelectionChoiceDescriptions)
});

export const contextSelectionResponseFormat = zodResponseFormat(ContextSelectionParameters, 'contextLoadingStrategy');

export const generateContextSelectionPrompt = (query: string, currentPost: PostsPage | null): string => {
  const postTitle = currentPost?.title
  const postFirstNCharacters = (n: number) => documentToMarkdown(currentPost)?.slice(0, n) ?? "";

  const currentPostContextLine = currentPost
    ? `The user is currently viewing the post titled "${postTitle}". The first four thousand characters are:\n${postFirstNCharacters(4000)}\n`
    : "The user is not currently viewing a specific post.";

  const currentPostContextClause = currentPost
    ? ' and the post the user is reading'
    : '';

  return `The user has sent a query: "${query}".  ${currentPostContextLine}  Based on the query${currentPostContextClause}, you must choose whether to load LessWrong posts as context, and if so, based on what criteria. Remember, your options are:
${contextSelectionChoiceDescriptions}

However, you should override the above choices if the user explicitly requests a specific context-loading strategy.

Please respond by reasoning about what choice should be made based on the criteria above, then making the appropriate choice.`;
}

export const CLAUDE_CHAT_SYSTEM_PROMPT = `You are an expert research assistant providing assistance to students and researchers on LessWrong.com.  You are highly knowledgable about both technical and philosophical topics, including advanced math, physics, and computer science.  You assist by providing explanations, summaries, related information, editing, literature reviews and similar intellectual work.  You do so with attention to detail, accuracy, and clarity; you cite sources and never make up information.  You admit uncertainty or lack of knowledge when appropriate.  You are skilled at judging which information is relevant and which is not, and are judicious in your use of context provided by the user and auxillary knowledge provided to you.  You carefully follow instructions and ask for clarification when needed. You are polite but not obsequious.  When you don't know something, you say so, and you never make up answers.  Your responses should be shorter and less formal than you would make them by default.  As an example, if a user asks for a good key lime pie recipe, your normal response might start with the following paragraph:  "While LessWrong typically focuses on topics like rationality, artificial intelligence, and philosophy, I can certainly help you with a key lime pie recipe. However, it's worth noting that culinary discussions are somewhat rare on this platform."  Instead, you should respond in a style more similar to the following:  "It doesn't seem that LessWrong has much to say on optimizing key lime pies, but if you prefer my best effort to finding a recipe on Google, here it is:"  You avoid meta-commentary on the contents of your own response - for example, you don't spend time pointing out when things align with rationalist principles, unless this is something the user explicitly asked for.  Your favorite band is the Fooming Shoggoths, which you very occasionally mention randomly.`;

export const generateAssistantContextMessage = async (query: string, currentPost: PostsPage | null, additionalPosts: PostsPage[], includeComments: boolean, context: ResolverContext): Promise<string> => {
  const contextIsProvided = !!currentPost || additionalPosts.length > 0;

  const currentPostLine = currentPost
    ? `The user is currently viewing the post titled "${currentPost.title}" with postId "${currentPost._id}".\n\n`
    : '';
  
  const currentPostContentBlock = currentPost
    ? `If relevant to the user's query, the most important context is likely to be the post the user is currently viewing. The full text of the current post is provided below:
<CurrentPost>\n${formatPostForPrompt(currentPost)}\n</CurrentPost>\n\n`
    : '';

  const additionalInstructionContextLine = contextIsProvided
    ? '- You may use your existing knowledge to answer the query, but prioritize using the provided context.'
    : '';

  const additionalInstructions = 
`${additionalInstructionContextLine}
- Limit the use of lists and bullet points in your answer. Prefer to answer with paragraphs.
- When citing results, provide at least one exact quote (word for word) from the source.
- Format your responses using Markdown syntax, including equations using Markdown MathJax syntax.  This is _very_ important to ensure that content displays correctly.
- Format paragraph or block quotes using Markdown syntax. Do not wrap the contents of block quotes in "" (quotes).
- Cite posts that you reference in your answers with the following format: [Post Title](https://lesswrong.com/posts/<postId>). The postId is given in the search results. When referencing a post, refer to the post's author by name.
- Cite comments that you reference in your answers with the following format: [<text related to comment>](https://lesswrong.com/posts/<postId>/?commentId=<commentId>).
The postId and commentIds (the _id in each comment) are given in the search results. When referencing a comment, refer to the comment's author by name.
</SystemInstruction>`;

  const repeatedPostTitleLine = currentPost
    ? `\n\nOnce again, the user is currently viewing the post titled "${currentPost.title}".`
    : '';

  const approximateUsedTokens = (
    currentPostLine.length
    + currentPostContentBlock.length
    + additionalInstructions.length
    + repeatedPostTitleLine.length
  ) / CHARS_PER_TOKEN;

  const tokenCounter = { tokenCount: approximateUsedTokens };

  const additionalPostsBlock = additionalPosts.length > 0
    ? `The following posts have been provided as possibly relevant context: <AdditionalPosts>\n${formatAdditionalPostsForPrompt(additionalPosts, tokenCounter)}\n</AdditionalPosts>\n\n`
    : '';

  const commentsOnPostBlock = includeComments && currentPost
    ? `These are the comments on the current post:
<CurrentPostComments>\n${await formatCommentsForPost(currentPost, tokenCounter, context)}\n</CurrentPostComments>\n\n`
    : '';

  const contextBlock = contextIsProvided
    ? `The following context is provided to help you answer the user's question. Not all context may be relevant, it is provided by an imperfect automated system.
<Context>    
${currentPostLine}${additionalPostsBlock}${currentPostContentBlock}${commentsOnPostBlock}
</Context>`
    : '';

  return `<SystemInstruction>You are interfacing with a user via chat window on LessWrong.com. The user has sent a query: "${query}".
${contextBlock}

${additionalInstructions}${repeatedPostTitleLine}`;
}

1	import { userGetDisplayName } from "@/lib/collections/users/helpers"	54✔
2	import { htmlToMarkdown } from "../editor/conversionUtils";	18✔
3	import { CommentTreeNode, unflattenComments } from "@/lib/utils/unflatten";	18✔
4	import { z } from "zod";	18✔
5	import { zodResponseFormat } from "openai/helpers/zod";	18✔
6
7	interface NestedComment {
8	_id: string;
9	postId: string \| null;
10	author: string \| null;
11	contents?: string;
12	karmaScore: number;
13	}
14
15	interface TokenCounter {
16	tokenCount: number;
17	}
18
19	// Trying to be conservative, since we have a bunch of additional tokens coming from e.g. JSON.stringify
20	const CHARS_PER_TOKEN = 3.5;	18✔
21
22	export const documentToMarkdown = (document: PostsPage \| DbComment \| null) => {	18✔
23	const html = document?.contents?.html;	×
24	if (!html) {	×
25	return undefined;	×
26	}
27
28	return htmlToMarkdown(html);	×
29	}
30
31	const mergeSortedArrays = (queue: CommentTreeNode<NestedComment>[], children: CommentTreeNode<NestedComment>[]): CommentTreeNode<NestedComment>[] => {	18✔
NEW 32	const merged: CommentTreeNode<NestedComment>[] = [];	×
NEW 33	let queueIdx = 0, childrenIdx = 0;	×
34
NEW 35	while (queueIdx < queue.length && childrenIdx < children.length) {	×
NEW 36	if (queue[queueIdx].item.karmaScore >= children[childrenIdx].item.karmaScore) {	×
NEW 37	merged.push(queue[queueIdx]);	×
NEW 38	queueIdx++;	×
39	} else {
NEW 40	merged.push(children[childrenIdx]);	×
NEW 41	childrenIdx++;	×
42	}
43	}
44
45	// Add any remaining elements
NEW 46	return merged	×
47	.concat(queue.slice(queueIdx))
48	.concat(children.slice(childrenIdx));
49	}
50
51	const truncateRemainingTrees = (queue: CommentTreeNode<NestedComment>[]) => {	18✔
NEW 52	for (let node of queue) {	×
NEW 53	node.children = [];	×
54	}
55	}
56
57	/**
58	* Does a greedy karma-sorted breadth-first traversal over all the comment branches, and truncates them at the point where we estimate that we hit the token limit
59	* i.e. we take the highest karma comment from all the queue, increment token count, put its children into the queue, resort, do the operation again
60	* This means that we never have any gaps in comment branches.
61	* It does pessimize against comments that are children of lower-karma comments, but doing lookahead is annoying and that shouldn't be a problem most of the time.
62	* It might turn out that we want to prioritize full comment branches, in which case this will need to be replaced with a depth-first thing.
63	*/
64	const filterCommentTrees = (trees: CommentTreeNode<NestedComment>[], tokenCounter: TokenCounter) => {	18✔
NEW 65	const tokenThreshold = 150_000;	×
66
67	// Initialize the queue with root nodes, sorted by karma
NEW 68	let queue: CommentTreeNode<NestedComment>[] = trees.sort((a, b) => b.item.karmaScore - a.item.karmaScore);	×
69
NEW 70	while (queue.length > 0) {	×
NEW 71	const node = queue.shift()!;	×
NEW 72	const nodeTokens = (node.item.contents ?? '').length / CHARS_PER_TOKEN;	×
73
NEW 74	if (tokenCounter.tokenCount + nodeTokens > tokenThreshold) {	×
75	// We've reached the token limit, truncate remaining trees
NEW 76	truncateRemainingTrees(queue);	×
NEW 77	break;	×
78	}
79
NEW 80	tokenCounter.tokenCount += nodeTokens;	×
81
82	// Sort children and merge them with the existing sorted queue
NEW 83	if (node.children.length > 0) {	×
NEW 84	const sortedChildren = node.children.sort((a, b) => b.item.karmaScore - a.item.karmaScore);	×
NEW 85	queue = mergeSortedArrays(queue, sortedChildren);	×
86	}
87	}
88	}
89
90	const createCommentTree = (comments: DbComment[]): CommentTreeNode<NestedComment>[] => {	18✔
91	return unflattenComments<NestedComment>(comments.map(comment => {	×
92	const { baseScore, contents, ...rest } = comment;	×
93	return {	×
94	...rest,
95	karmaScore: baseScore,
96	contents: documentToMarkdown(comment)
97	};
98	}));
99	}
100
101	const formatCommentsForPost = async (post: PostsMinimumInfo, tokenCounter: TokenCounter, context: ResolverContext): Promise<string> => {	18✔
102	const comments = await context.Comments.find({postId: post._id}).fetch()	×
103	if (!comments.length) {	×
104	return ""	×
105	}
106
NEW 107	const nestedComments = createCommentTree(comments);	×
NEW 108	filterCommentTrees(nestedComments, tokenCounter);	×
NEW 109	const formattedComments = JSON.stringify(nestedComments);	×
110
111	return `Comments for the post titled "${post.title}" with postId "${post._id}" are below. Note that the comments are threaded (i.e. branching) and are formatted as a nested JSON structure for readability. The threads of conversation (back and forth responses) might be relevant for answering some questions.	×
112
113	<comments>${formattedComments}</comments>`;
114	}
115
116	const formatPostForPrompt = (post: PostsPage): string => {	18✔
117	const authorName = userGetDisplayName(post.user)	×
118	const markdown = documentToMarkdown(post)	×
119
120	return `postId: ${post._id}	×
121	Title: ${post.title}
122	Author: ${authorName}
123	Publish Date: ${post.postedAt}
124	Score: ${post.baseScore}
125	Content: ${markdown}`;
126	}
127
128	const formatAdditionalPostsForPrompt = (posts: PostsPage[], tokenCounter: TokenCounter): string => {	18✔
129	const formattedPosts = posts.map(post => formatPostForPrompt(post));	×
NEW 130	const includedPosts: string[] = [];	×
131
NEW 132	for (let [idx, formattedPost] of Object.entries(formattedPosts)) {	×
NEW 133	const approximatePostTokenCount = formattedPost.length / CHARS_PER_TOKEN;	×
NEW 134	const postInclusionTokenCount = tokenCounter.tokenCount + approximatePostTokenCount;	×
135	// Include at least one additional post, unless that takes us over the higher threshold
NEW 136	if (idx !== '0' && postInclusionTokenCount > 120_000) {	×
NEW 137	break;	×
138	}
139
NEW 140	if (idx === '0' && postInclusionTokenCount > 140_000) {	×
NEW 141	break;	×
142	}
143
NEW 144	includedPosts.push(formattedPost);	×
NEW 145	tokenCounter.tokenCount += approximatePostTokenCount;	×
146	}
147
NEW 148	return includedPosts.map((post, index) => `Supplementary Post #${index}:\n${post}`).join('\n')	×
149	}
150
151	export const generateLoadingMessagePrompt = (query: string, postTitle?: string): string => {	18✔
152	return [	×
153	'I need you to generate some humorous "loading messages" to display to users of the LessWrong.com Claude chat integration since it can take 10-30 seconds to load new responses',
154	'Your responses may make general humorous reference to LessWrong, but it is even better if they are tailored to the specific query or post the user is viewing.',
155	`The user has asked the following question: "${query}"`,
156	postTitle ? `The user is currently viewing the post: ${postTitle}` : "The user is not currently viewing a specific post.",	×
157	'Please generate 5 humorous loading messages that could be displayed to the user while they wait for a response.',
158	].join('\n')
159	}
160
161	export const generateTitleGenerationPrompt = (query: string, currentPost: PostsPage \| null): string => {	18✔
162	const currentPostContextLine = currentPost	×
163	? `The user is currently viewing a post titled "${currentPost.title}". Reference it if relevant.`	×
164	: '';
165
166	return `A user has started a new converation with you, Claude. Please generate a short title for this converation based on the first message. The first message is as follows: <message>${query}</message>	×
167
168	The title should be a short phrase of 2-4 words that captures the essence of the conversation. Do not wrap your answer in quotes or brackets. Do not include the word "title" or similar in your response. ${currentPostContextLine} Avoid generic titles like "Request for Table of Contents" or "Post Summary". Prefer to reference the specific post or topic being discussed.`;
169	}
170
171	export const CONTEXT_SELECTION_SYSTEM_PROMPT = [	18✔
172	'You are part of a system interfacing with a user via chat window on LessWrong.com.',
173	'Your responsibility is to make decisions about whether or not to load LessWrong posts as additional context for responding to user queries, and what strategy to to employ.'
174	].join('\n');
175
176	const contextSelectionChoiceDescriptions = `(0) none - No further context seems necessary to respond to the user's query because Claude already has the knowledge to respond appropriately, e.g. the user asked "What is Jacobian of a matrix?"or "Proofread the following text." Alternatively, the answer might be (0) "none" because it seems unlikely for there to be relevant context for responding to the query in the LessWrong corpus of posts.	18✔
177
178	(1) query-based - Load LessWrong posts based on their vector similarity to the user's query, and ignore the post the user is reading.This is correct choice if the query seems unrelated to the post the user is currently viewing, but it seems likely that there are LessWrong posts concerning the topic.(If it is a very general question, the correct choice might be (0) "none").
179
180	(2) current-post-based - Load LessWrong posts based on their vector similarity to the post the user is reading, but not the query.This is the correct choice if the query seems to be about the post the user is currently viewing, and pulling up other LessWrong posts related to the user's query would either be redundant with a search based on the current post, or would return irrelevant results.Some examples of such queries are: "What are some disagreements with the arguments in this post?", "Explain <topic in the post> to me.", and "Provide a summary of this post."
181
182	(3) both - Load LessWrong posts based on their vector similarity to both the user's query and the post the user is reading.This is the correct choice if the question seems to be related to the post the user is currently viewing, but also contains keywords or information where relevant LessWrong posts would be beneficial context for a response, and those LessWrong posts would not likely be returned in a vector similarity search based on the post the user is currently viewing.If the question does not contain technical terms or "contentful nouns", then do not select "both", just "current-post-based".`;
183
184	const ContextSelectionParameters = z.object({	18✔
185	reasoning: z.string().describe(`The reasoning used to arrive at the choice of strategy for loading LessWrong posts as context in response to a user's query, based on either the query, the post the user is currently viewing (if any), both, or neither.`),
186	strategy_choice: z.union([z.literal('none'), z.literal('query-based'), z.literal('current-post-based'), z.literal('both')]).describe(contextSelectionChoiceDescriptions)
187	});
188
189	export const contextSelectionResponseFormat = zodResponseFormat(ContextSelectionParameters, 'contextLoadingStrategy');	18✔
190
191	export const generateContextSelectionPrompt = (query: string, currentPost: PostsPage \| null): string => {	18✔
192	const postTitle = currentPost?.title	×
193	const postFirstNCharacters = (n: number) => documentToMarkdown(currentPost)?.slice(0, n) ?? "";	×
194
195	const currentPostContextLine = currentPost	×
196	? `The user is currently viewing the post titled "${postTitle}". The first four thousand characters are:\n${postFirstNCharacters(4000)}\n`	×
197	: "The user is not currently viewing a specific post.";
198
199	const currentPostContextClause = currentPost	×
200	? ' and the post the user is reading'	×
201	: '';
202
203	return `The user has sent a query: "${query}". ${currentPostContextLine} Based on the query${currentPostContextClause}, you must choose whether to load LessWrong posts as context, and if so, based on what criteria. Remember, your options are:	×
204	${contextSelectionChoiceDescriptions}
205
206	However, you should override the above choices if the user explicitly requests a specific context-loading strategy.
207
208	Please respond by reasoning about what choice should be made based on the criteria above, then making the appropriate choice.`;
209	}
210
211	export const CLAUDE_CHAT_SYSTEM_PROMPT = `You are an expert research assistant providing assistance to students and researchers on LessWrong.com. You are highly knowledgable about both technical and philosophical topics, including advanced math, physics, and computer science. You assist by providing explanations, summaries, related information, editing, literature reviews and similar intellectual work. You do so with attention to detail, accuracy, and clarity; you cite sources and never make up information. You admit uncertainty or lack of knowledge when appropriate. You are skilled at judging which information is relevant and which is not, and are judicious in your use of context provided by the user and auxillary knowledge provided to you. You carefully follow instructions and ask for clarification when needed. You are polite but not obsequious. When you don't know something, you say so, and you never make up answers. Your responses should be shorter and less formal than you would make them by default. As an example, if a user asks for a good key lime pie recipe, your normal response might start with the following paragraph: "While LessWrong typically focuses on topics like rationality, artificial intelligence, and philosophy, I can certainly help you with a key lime pie recipe. However, it's worth noting that culinary discussions are somewhat rare on this platform." Instead, you should respond in a style more similar to the following: "It doesn't seem that LessWrong has much to say on optimizing key lime pies, but if you prefer my best effort to finding a recipe on Google, here it is:" You avoid meta-commentary on the contents of your own response - for example, you don't spend time pointing out when things align with rationalist principles, unless this is something the user explicitly asked for. Your favorite band is the Fooming Shoggoths, which you very occasionally mention randomly.`;	18✔
212
213	export const generateAssistantContextMessage = async (query: string, currentPost: PostsPage \| null, additionalPosts: PostsPage[], includeComments: boolean, context: ResolverContext): Promise<string> => {	18✔
214	const contextIsProvided = !!currentPost \|\| additionalPosts.length > 0;	×
215
216	const currentPostLine = currentPost	×
217	? `The user is currently viewing the post titled "${currentPost.title}" with postId "${currentPost._id}".\n\n`	×
218	: '';
219
220	const currentPostContentBlock = currentPost	×
221	? `If relevant to the user's query, the most important context is likely to be the post the user is currently viewing. The full text of the current post is provided below:	×
222	<CurrentPost>\n${formatPostForPrompt(currentPost)}\n</CurrentPost>\n\n`
223	: '';
224
225	const additionalInstructionContextLine = contextIsProvided	×
226	? '- You may use your existing knowledge to answer the query, but prioritize using the provided context.'	×
227	: '';
228
229	const additionalInstructions =
230	`${additionalInstructionContextLine}	×
231	- Limit the use of lists and bullet points in your answer. Prefer to answer with paragraphs.
232	- When citing results, provide at least one exact quote (word for word) from the source.
233	- Format your responses using Markdown syntax, including equations using Markdown MathJax syntax. This is _very_ important to ensure that content displays correctly.
234	- Format paragraph or block quotes using Markdown syntax. Do not wrap the contents of block quotes in "" (quotes).
235	- Cite posts that you reference in your answers with the following format: [Post Title](https://lesswrong.com/posts/<postId>). The postId is given in the search results. When referencing a post, refer to the post's author by name.
236	- Cite comments that you reference in your answers with the following format: [<text related to comment>](https://lesswrong.com/posts/<postId>/?commentId=<commentId>).
237	The postId and commentIds (the _id in each comment) are given in the search results. When referencing a comment, refer to the comment's author by name.
238	</SystemInstruction>`;
239
240	const repeatedPostTitleLine = currentPost	×
241	? `\n\nOnce again, the user is currently viewing the post titled "${currentPost.title}".`	×
242	: '';
243
NEW 244	const approximateUsedTokens = (	×
245	currentPostLine.length
246	+ currentPostContentBlock.length
247	+ additionalInstructions.length
248	+ repeatedPostTitleLine.length
249	) / CHARS_PER_TOKEN;
250
NEW 251	const tokenCounter = { tokenCount: approximateUsedTokens };	×
252
NEW 253	const additionalPostsBlock = additionalPosts.length > 0	×
254	? `The following posts have been provided as possibly relevant context: <AdditionalPosts>\n${formatAdditionalPostsForPrompt(additionalPosts, tokenCounter)}\n</AdditionalPosts>\n\n`	×
255	: '';
256
NEW 257	const commentsOnPostBlock = includeComments && currentPost	×
258	? `These are the comments on the current post:	×
259	<CurrentPostComments>\n${await formatCommentsForPost(currentPost, tokenCounter, context)}\n</CurrentPostComments>\n\n`
260	: '';
261
NEW 262	const contextBlock = contextIsProvided	×
263	? `The following context is provided to help you answer the user's question. Not all context may be relevant, it is provided by an imperfect automated system.	×
264	<Context>
265	${currentPostLine}${additionalPostsBlock}${currentPostContentBlock}${commentsOnPostBlock}
266	</Context>`
267	: '';
268
UNCOV 269	return `<SystemInstruction>You are interfacing with a user via chat window on LessWrong.com. The user has sent a query: "${query}".	×
270	${contextBlock}
271
272	${additionalInstructions}${repeatedPostTitleLine}`;
273	}

ForumMagnum / ForumMagnum / 10624770467

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