diff options
Diffstat (limited to 'src/client/apis/gpt/GPT.ts')
| -rw-r--r-- | src/client/apis/gpt/GPT.ts | 144 |
1 files changed, 120 insertions, 24 deletions
diff --git a/src/client/apis/gpt/GPT.ts b/src/client/apis/gpt/GPT.ts index 03380e4d6..29b6ab989 100644 --- a/src/client/apis/gpt/GPT.ts +++ b/src/client/apis/gpt/GPT.ts @@ -1,14 +1,22 @@ import { ChatCompletionMessageParam, Image } from 'openai/resources'; import { openai } from './setup'; +export enum GPTDocCommand { + AssignTags = 1, + Filter = 2, + GetInfo = 3, + Sort = 4, +} + +export const DescriptionSeperator = '======'; +export const DocSeperator = '------'; + enum GPTCallType { SUMMARY = 'summary', COMPLETION = 'completion', EDIT = 'edit', CHATCARD = 'chatcard', // a single flashcard style response to a question FLASHCARD = 'flashcard', // a set of flashcard qustion/answer responses to a topic - QUIZ = 'quiz', - SORT = 'sort', DESCRIBE = 'describe', MERMAID = 'mermaid', DATA = 'data', @@ -16,15 +24,17 @@ enum GPTCallType { PRONUNCIATION = 'pronunciation', DRAW = 'draw', COLOR = 'color', - RUBRIC = 'rubric', - TYPE = 'type', - SUBSET = 'subset', - INFO = 'info', TEMPLATE = 'template', VIZSUM = 'vizsum', VIZSUM2 = 'vizsum2', FILL = 'fill', COMPLETEPROMPT = 'completeprompt', + QUIZDOC = 'quiz_doc', + MAKERUBRIC = 'make_rubric', // create a definition rubric for a document to be used when quizzing the user + COMMANDTYPE = 'command_type', // Determine the type of command being made (GPTQueryType - eg., AssignTags, Sort, Filter, DocInfo, GenInfo) and possibly some parameters (eg, Tag type for Tags) + SUBSETDOCS = 'subset_docs', // select a subset of documents based on their descriptions + DOCINFO = 'doc_info', // provide information about a document + SORTDOCS = 'sort_docs', } type GPTCallOpts = { @@ -34,7 +44,7 @@ type GPTCallOpts = { prompt: string; }; -const callTypeMap: { [type: string]: GPTCallOpts } = { +const callTypeMap: { [type in GPTCallType]: GPTCallOpts } = { // newest model: gpt-4 summary: { model: 'gpt-4-turbo', maxTokens: 256, temp: 0.5, prompt: 'Summarize the text given in simpler terms.' }, edit: { model: 'gpt-4-turbo', maxTokens: 256, temp: 0.5, prompt: 'Reword the text.' }, @@ -42,7 +52,7 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { model: 'gpt-4o', maxTokens: 2048, temp: 0.7, - prompt: 'Create a stack of flashcards out of this text with each question and answer labeled as question and answer. Create a title that represents the question in just a few words and label it "title". For some questions, ask "what is this image of" but tailored to stacks theme and the image and write a keyword that represents the image and label it "keyword". Otherwise, write none. Do not label each flashcard and do not include asterisks.', + prompt: 'Create a stack of at least 10 flashcards out of this text with each question and answer labeled as question and answer. Each flashcard should have a title that represents the question in just a few words and label it "title". For some questions, ask "what is this image of" but tailored to stacks theme and the image and write a keyword that represents the image and label it "keyword". Otherwise, write none. Do not label each flashcard and do not include asterisks.', }, completion: { model: 'gpt-4-turbo', maxTokens: 256, temp: 0.5, prompt: "You are a helpful assistant. Answer the user's prompt." }, @@ -58,11 +68,17 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { temp: 0.5, prompt: "You are a helpful resarch assistant. Analyze the user's data to find meaningful patterns and/or correlation. Please only return a JSON with a correlation column 1 propert, a correlation column 2 property, and an analysis property. ", }, - sort: { + sort_docs: { model: 'gpt-4o', maxTokens: 2048, temp: 0.25, - prompt: "The user is going to give you a list of descriptions. Each one is separated by `======` on either side. Descriptions will vary in length, so make sure to only separate when you see `======`. Sort them by the user's specifications. Make sure each description is only in the list once. Each item should be separated by `======`. Immediately afterward, surrounded by `------` on BOTH SIDES, provide some insight into your reasoning for the way you sorted (and mention nothing about the formatting details given in this description). It is VERY important that you format it exactly as described, ensuring the proper number of `=` and `-` (6 of each) and NO commas", + prompt: `The user is going to give you a list of descriptions. + Each one is separated by '${DescriptionSeperator}' on either side. + Descriptions will vary in length, so make sure to only separate when you see '${DescriptionSeperator}'. + Sort them by the user's specifications. + Make sure each description is only in the list once. Each item should be separated by '${DescriptionSeperator}'. + Immediately afterward, surrounded by '${DocSeperator}' on BOTH SIDES, provide some insight into your reasoning for the way you sorted (and mention nothing about the formatting details given in this description). + It is VERY important that you format it exactly as described, ensuring the proper number of '${DescriptionSeperator[0]}' and '${DocSeperator[0]}' (${DescriptionSeperator.length} of each) and NO commas`, }, describe: { model: 'gpt-4-vision-preview', maxTokens: 2048, temp: 0, prompt: 'Describe these images in 3-5 words' }, flashcard: { @@ -72,7 +88,7 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { prompt: 'Make flashcards out of this text with each question and answer labeled as question and answer. Create a title for each question and asnwer that is labeled as "title". Do not label each flashcard and do not include asterisks: ', }, chatcard: { model: 'gpt-4-turbo', maxTokens: 512, temp: 0.5, prompt: 'Answer the following question as a short flashcard response. Do not include a label.' }, - quiz: { + quiz_doc: { model: 'gpt-4-turbo', maxTokens: 1024, temp: 0, @@ -88,7 +104,7 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { model: 'gpt-4-turbo', maxTokens: 512, temp: 0.5, - prompt: 'You will be given a list of field descriptions for multiple templates in the format {field #0: “description”}{field #1: “description”}{...}, and a list of column descriptions in the format {“title”: “description”}{...}. Your job is to match columns with fields based on their descriptions. Your output should be in the following JSON format: {“Template title”:{“#”: “title”, “#”: “title”, “#”: “title” …}, “Template title”:{“#”: “title”, “#”: “title”, “#”: “title” …}} where “Template title” represents the template, # represents the field # and “title” the title of the column assigned to it. A filled out example might look like {“fivefield2”:{“0”:”Name”, “1”:”Image”, “2”:”Caption”, “3”:”Position”, “4”:”Stats”}, “fivefield3”:{0:”Image”, 1:”Name”, 2:”Caption”, 3:”Stats”, 4:”Position”}. Include one object for each template. IT IS VERY IMPORTANT THAT YOU ONLY INCLUDE TEXT IN THE FORMAT ABOVE, WITH NO ADDITIONS WHATSOEVER. Do not include extraneous ‘#’ characters, ‘column’ for columns, or ‘template’ for templates: ONLY THE TITLES AND NUMBERS. There should never be one column assigned to more than one field (ie. if the “name” column is assigned to field 1, it can’t be assigned to any other fields) . Do this for each template whose fields are described. The descriptions are as follows:', + prompt: 'You will be given a list of field descriptions for one or more templates in the format {field #0: “description”}{field #1: “description”}{...}, and a list of column descriptions in the format {“title”: “description”}{...}. Your job is to match columns with fields based on their descriptions. Your output should be in the following JSON format: {“template_title”:{“#”: “title”, “#”: “title”, “#”: “title” …}, “template_title”:{“#”: “title”, “#”: “title”, “#”: “title” …}} where “template_title” is the templates title as specified in the description provided, # represents the field # and “title” the title of the column assigned to it. A filled out example might look like {“fivefield2”:{“0”:”Name”, “1”:”Image”, “2”:”Caption”, “3”:”Position”, “4”:”Stats”}, “fivefield3”:{0:”Image”, 1:”Name”, 2:”Caption”, 3:”Stats”, 4:”Position”}. Include one object for each template. IT IS VERY IMPORTANT THAT YOU ONLY INCLUDE TEXT IN THE FORMAT ABOVE, WITH NO ADDITIONS WHATSOEVER. Do not include extraneous ‘#’ characters, ‘column’ for columns, or ‘template’ for templates: ONLY THE TITLES AND NUMBERS. There should never be one column assigned to more than one field (ie. if the “name” column is assigned to field 1, it can’t be assigned to any other fields) . Do this for each template whose fields are described. The descriptions are as follows:', }, vizsum: { model: 'gpt-4-turbo', @@ -103,12 +119,12 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { prompt: 'Your job is to provide structured information on columns in a dataset based on example rows. You will categorize each column in two ways: by type and size. The size categories are as follows: tiny (one or two words), small (a sentence/multiple words), medium (a few sentences), large (a longer paragraph), and huge (a very long or multiple paragraphs). The type categories are as follows: visual (links/file paths to images, pdfs, maps, or any other visual media type), and text (plain text that isn’t a link/file path). Visual media should be assumed to have size “medium” “large” or “huge”. You will give your responses in JSON format, like so: {“title (of column)”:{“type”:”text”, “size”:”small”}, “title (of column)”:{“type”:”visual”, “size”:”medium”}, …}. DO NOT INCLUDE ANY OTHER TEXT, ONLY THE JSON.', }, fill: { - model: 'gpt-4-turbo', + model: 'gpt-4o', maxTokens: 512, temp: 0.5, prompt: 'Your job is to generate content for fields based on a user prompt and background context given to you. You will be given the content of the other fields present in the format: ---- Field # (field title): content ---- Field # (field title): content ----- (etc.) You will be given info on the columns to generate for in the format ---- title: , prompt: , word limit: , assigned field: ----. For each column, based on the prompt, word limit, and the context of existing fields, you should generate a short response in the following JSON format: {“___”(where ___ is the title from the column description with no additions): {“number”:”#” (where # is the assigned field of the column), “content”:”response” (where response is your response to the prompt in the column info)}}. Here’s another example of the format with only one column: {“position”: {“number”:”2”, “content”:”*your response goes here*”}}. ONLY INCLUDE THE JSON TEXT WITH NO OTHER ADDED TEXT. YOUR RESPONSE MUST BE VALID JSON. The word limit for each column applies only to that column’s response. Do not include speculation or information that you can’t glean from your factual knowledge or the content of the other fields (no description of images you can’t see, for example). You should include one object per column you are provided info on.', }, - completeprompt: { model: 'gpt-4-turbo', maxTokens: 512, temp: 0.5, prompt: 'Your prompt is as follows:' }, + completeprompt: { model: 'gpt-4o', maxTokens: 512, temp: 0.5, prompt: 'Your prompt is as follows:' }, draw: { model: 'gpt-4o', maxTokens: 1024, @@ -121,6 +137,46 @@ const callTypeMap: { [type: string]: GPTCallOpts } = { temp: 0.5, prompt: 'You will be coloring drawings. You will be given what the drawing is, then a list of descriptions for parts of the drawing. Based on each description, respond with the stroke and fill color that it should be. Follow the rules: 1. Avoid using black for stroke color 2. Make the stroke color 1-3 shades darker than the fill color 3. Use the same colors when possible. Format as {#abcdef #abcdef}, making sure theres a color for each description, and do not include any additional text.', }, + command_type: { + model: 'gpt-4-turbo', + maxTokens: 1024, + temp: 0, + prompt: `I'm going to provide you with a question. + Based on the question, is the user asking you to + ${GPTDocCommand.AssignTags}. Assigns docs with tags(like star / heart etc)/labels. + ${GPTDocCommand.GetInfo}. Provide information about a specific doc. + ${GPTDocCommand.Filter}. Filter docs based on a question/information. + ${GPTDocCommand.Sort}. Put docs in a specific order. + Answer with only the number for ${GPTDocCommand.GetInfo}-${GPTDocCommand.Sort}. + For number one, provide the number (${GPTDocCommand.AssignTags}) and the appropriate tag`, + }, + subset_docs: { + model: 'gpt-4-turbo', + maxTokens: 1024, + temp: 0, + prompt: `I'm going to give you a list of descriptions. + Each one is separated by '${DescriptionSeperator}' on either side. + Descriptions will vary in length, so make sure to only separate when you see '${DescriptionSeperator}'. + Based on the question the user asks, provide a subset of the given descriptions that best matches the user's specifications. + Make sure each description is only in the list once. Each item should be separated by '${DescriptionSeperator}'. + Immediately afterward, surrounded by '${DocSeperator}' on BOTH SIDES, provide some insight into your reasoning in the 2nd person (and mention nothing about the formatting details given in this description). + It is VERY important that you format it exactly as described, ensuring the proper number of '${DescriptionSeperator[0]}' and '${DocSeperator[0]}' (${DescriptionSeperator.length} of each) and NO commas`, + }, + + doc_info: { + model: 'gpt-4-turbo', + maxTokens: 1024, + temp: 0, + prompt: `Answer the user's question with a short (<100 word) response. + If a particular document is selected I will provide that information (which may help with your response)`, + }, + make_rubric: { + model: 'gpt-4-turbo', + maxTokens: 1024, + temp: 0, + prompt: `BRIEFLY (<25 words) provide a definition for the following term. + It will be used as a rubric to evaluate the user's understanding of the topic`, + }, }; let lastCall = ''; let lastResp = ''; @@ -131,13 +187,15 @@ let lastResp = ''; * @returns AI Output */ const gptAPICall = async (inputTextIn: string, callType: GPTCallType, prompt?: string, dontCache?: boolean) => { - const inputText = [GPTCallType.SUMMARY, GPTCallType.FLASHCARD, GPTCallType.QUIZ, GPTCallType.STACK].includes(callType) ? inputTextIn + '.' : inputTextIn; + const inputText = inputTextIn + ([GPTCallType.SUMMARY, GPTCallType.FLASHCARD, GPTCallType.QUIZDOC, GPTCallType.STACK].includes(callType) ? '.' : ''); const opts = callTypeMap[callType]; - if (lastCall === inputText && dontCache !== true) return lastResp; + if (!opts) { + console.log('The query type:' + callType + ' requires a configuration.'); + return 'Error connecting with API.'; + } + if (lastCall === inputText && dontCache !== true && lastResp) return lastResp; try { - lastCall = inputText; - - const usePrompt = prompt ? prompt + opts.prompt : opts.prompt; + const usePrompt = prompt ? prompt + '.' + opts.prompt : opts.prompt; const messages: ChatCompletionMessageParam[] = [ { role: 'system', content: usePrompt }, { role: 'user', content: inputText }, @@ -149,9 +207,12 @@ const gptAPICall = async (inputTextIn: string, callType: GPTCallType, prompt?: s temperature: opts.temp, max_tokens: opts.maxTokens, }); - lastResp = response.choices[0].message.content ?? ''; - console.log('RESP:' + lastResp); - return lastResp; + const result = response.choices[0].message.content ?? ''; + if (!dontCache) { + lastResp = result; + lastCall = inputText; + } + return result; } catch (err) { console.log(err); return 'Error connecting with API.'; @@ -249,6 +310,41 @@ const gptHandwriting = async (src: string): Promise<string> => { } }; +const gptDescribeImage = async (image: string): Promise<string> => { + try { + const response = await openai.chat.completions.create({ + model: 'gpt-4o', + temperature: 0, + messages: [ + { + role: 'user', + content: [ + { + type: 'text', + text: `Very briefly identify what this drawing is and list all the drawing elements and their location within the image. Do not include anything about the drawing style.`, + }, + { + type: 'image_url', + image_url: { + url: `${image}`, + detail: 'low', + }, + }, + ], + }, + ], + }); + if (response.choices[0].message.content) { + console.log('GPT DESCRIPTION', response.choices[0].message.content); + return response.choices[0].message.content; + } + return 'Unknown drawing'; + } catch (err) { + console.log(err); + return 'Error connecting with API'; + } +}; + const gptDrawingColor = async (image: string, coords: string[]): Promise<string> => { try { const response = await openai.chat.completions.create({ @@ -276,11 +372,11 @@ const gptDrawingColor = async (image: string, coords: string[]): Promise<string> if (response.choices[0].message.content) { return response.choices[0].message.content; } - return 'Missing labels'; + return 'Unknown drawing'; } catch (err) { console.log(err); return 'Error connecting with API'; } }; -export { gptAPICall, gptImageCall, GPTCallType, gptImageLabel, gptGetEmbedding, gptHandwriting, gptDrawingColor }; +export { gptAPICall, gptImageCall, GPTCallType, gptImageLabel, gptGetEmbedding, gptHandwriting, gptDescribeImage, gptDrawingColor }; |