LibreChat/api/app/clients/tools/dynamic/OpenAPIPlugin.js
Danny Avila 317a1bd8da
feat: ConversationSummaryBufferMemory (#973)
* refactor: pass model in message edit payload, use encoder in standalone util function

* feat: add summaryBuffer helper

* refactor(api/messages): use new countTokens helper and add auth middleware at top

* wip: ConversationSummaryBufferMemory

* refactor: move pre-generation helpers to prompts dir

* chore: remove console log

* chore: remove test as payload will no longer carry tokenCount

* chore: update getMessagesWithinTokenLimit JSDoc

* refactor: optimize getMessagesForConversation and also break on summary, feat(ci): getMessagesForConversation tests

* refactor(getMessagesForConvo): count '00000000-0000-0000-0000-000000000000' as root message

* chore: add newer model to token map

* fix: condition was point to prop of array instead of message prop

* refactor(BaseClient): use object for refineMessages param, rename 'summary' to 'summaryMessage', add previous_summary
refactor(getMessagesWithinTokenLimit): replace text and tokenCount if should summarize, summary, and summaryTokenCount are present
fix/refactor(handleContextStrategy): use the right comparison length for context diff, and replace payload first message when a summary is present

* chore: log previous_summary if debugging

* refactor(formatMessage): assume if role is defined that it's a valid value

* refactor(getMessagesWithinTokenLimit): remove summary logic
refactor(handleContextStrategy): add usePrevSummary logic in case only summary was pruned
refactor(loadHistory): initial message query will return all ordered messages but keep track of the latest summary
refactor(getMessagesForConversation): use object for single param, edit jsdoc, edit all files using the method
refactor(ChatGPTClient): order messages before buildPrompt is called, TODO: add convoSumBuffMemory logic

* fix: undefined handling and summarizing only when shouldRefineContext is true

* chore(BaseClient): fix test results omitting system role for summaries and test edge case

* chore: export summaryBuffer from index file

* refactor(OpenAIClient/BaseClient): move refineMessages to subclass, implement LLM initialization for summaryBuffer

* feat: add OPENAI_SUMMARIZE to enable summarizing, refactor: rename client prop 'shouldRefineContext' to 'shouldSummarize', change contextStrategy value to 'summarize' from 'refine'

* refactor: rename refineMessages method to summarizeMessages for clarity

* chore: clarify summary future intent in .env.example

* refactor(initializeLLM): handle case for either 'model' or 'modelName' being passed

* feat(gptPlugins): enable summarization for plugins

* refactor(gptPlugins): utilize new initializeLLM method and formatting methods for messages, use payload array for currentMessages and assign pastMessages sooner

* refactor(agents): use ConversationSummaryBufferMemory for both agent types

* refactor(formatMessage): optimize original method for langchain, add helper function for langchain messages, add JSDocs and tests

* refactor(summaryBuffer): add helper to createSummaryBufferMemory, and use new formatting helpers

* fix: forgot to spread formatMessages also took opportunity to pluralize filename

* refactor: pass memory to tools, namely openapi specs. not used and may never be used by new method but added for testing

* ci(formatMessages): add more exhaustive checks for langchain messages

* feat: add debug env var for OpenAI

* chore: delete unnecessary comments

* chore: add extra note about summary feature

* fix: remove tokenCount from payload instructions

* fix: test fail

* fix: only pass instructions to payload when defined or not empty object

* refactor: fromPromptMessages is deprecated, use renamed method fromMessages

* refactor: use 'includes' instead of 'startsWith' for extended OpenRouter compatibility

* fix(PluginsClient.buildPromptBody): handle undefined message strings

* chore: log langchain titling error

* feat: getModelMaxTokens helper

* feat: tokenSplit helper

* feat: summary prompts updated

* fix: optimize _CUT_OFF_SUMMARIZER prompt

* refactor(summaryBuffer): use custom summary prompt, allow prompt to be passed, pass humanPrefix and aiPrefix to memory, along with any future variables, rename messagesToRefine to context

* fix(summaryBuffer): handle edge case where messagesToRefine exceeds summary context,
refactor(BaseClient): allow custom maxContextTokens to be passed to getMessagesWithinTokenLimit, add defined check before unshifting summaryMessage, update shouldSummarize based on this
refactor(OpenAIClient): use getModelMaxTokens, use cut-off message method for summary if no messages were left after pruning

* fix(handleContextStrategy): handle case where incoming prompt is bigger than model context

* chore: rename refinedContent to splitText

* chore: remove unnecessary debug log
2023-09-26 21:02:28 -04:00

174 lines
4.6 KiB
JavaScript

require('dotenv').config();
const { z } = require('zod');
const fs = require('fs');
const yaml = require('js-yaml');
const path = require('path');
const { DynamicStructuredTool } = require('langchain/tools');
const { createOpenAPIChain } = require('langchain/chains');
const { ChatPromptTemplate, HumanMessagePromptTemplate } = require('langchain/prompts');
function addLinePrefix(text, prefix = '// ') {
return text
.split('\n')
.map((line) => prefix + line)
.join('\n');
}
function createPrompt(name, functions) {
const prefix = `// The ${name} tool has the following functions. Determine the desired or most optimal function for the user's query:`;
const functionDescriptions = functions
.map((func) => `// - ${func.name}: ${func.description}`)
.join('\n');
return `${prefix}\n${functionDescriptions}
// The user's message will be passed as the function's query.
// Always provide the function name as such: {{"func": "function_name"}}`;
}
const AuthBearer = z
.object({
type: z.string().includes('service_http'),
authorization_type: z.string().includes('bearer'),
verification_tokens: z.object({
openai: z.string(),
}),
})
.catch(() => false);
const AuthDefinition = z
.object({
type: z.string(),
authorization_type: z.string(),
verification_tokens: z.object({
openai: z.string(),
}),
})
.catch(() => false);
async function readSpecFile(filePath) {
try {
const fileContents = await fs.promises.readFile(filePath, 'utf8');
if (path.extname(filePath) === '.json') {
return JSON.parse(fileContents);
}
return yaml.load(fileContents);
} catch (e) {
console.error(e);
return false;
}
}
async function getSpec(url) {
const RegularUrl = z
.string()
.url()
.catch(() => false);
if (RegularUrl.parse(url) && path.extname(url) === '.json') {
const response = await fetch(url);
return await response.json();
}
const ValidSpecPath = z
.string()
.url()
.catch(async () => {
const spec = path.join(__dirname, '..', '.well-known', 'openapi', url);
if (!fs.existsSync(spec)) {
return false;
}
return await readSpecFile(spec);
});
return ValidSpecPath.parse(url);
}
async function createOpenAPIPlugin({ data, llm, user, message, memory, verbose = false }) {
let spec;
try {
spec = await getSpec(data.api.url, verbose);
} catch (error) {
verbose && console.debug('getSpec error', error);
return null;
}
if (!spec) {
verbose && console.debug('No spec found');
return null;
}
const headers = {};
const { auth, name_for_model, description_for_model, description_for_human } = data;
if (auth && AuthDefinition.parse(auth)) {
verbose && console.debug('auth detected', auth);
const { openai } = auth.verification_tokens;
if (AuthBearer.parse(auth)) {
headers.authorization = `Bearer ${openai}`;
verbose && console.debug('added auth bearer', headers);
}
}
const chainOptions = {
llm,
verbose,
};
if (memory) {
verbose && console.debug('openAPI chain: memory detected', memory);
chainOptions.memory = memory;
}
if (data.headers && data.headers['librechat_user_id']) {
verbose && console.debug('id detected', headers);
headers[data.headers['librechat_user_id']] = user;
}
if (Object.keys(headers).length > 0) {
verbose && console.debug('headers detected', headers);
chainOptions.headers = headers;
}
if (data.params) {
verbose && console.debug('params detected', data.params);
chainOptions.params = data.params;
}
chainOptions.prompt = ChatPromptTemplate.fromMessages([
HumanMessagePromptTemplate.fromTemplate(
`# Use the provided API's to respond to this query:\n\n{query}\n\n## Instructions:\n${addLinePrefix(
description_for_model,
)}`,
),
]);
const chain = await createOpenAPIChain(spec, chainOptions);
const { functions } = chain.chains[0].lc_kwargs.llmKwargs;
return new DynamicStructuredTool({
name: name_for_model,
description_for_model: `${addLinePrefix(description_for_human)}${createPrompt(
name_for_model,
functions,
)}`,
description: `${description_for_human}`,
schema: z.object({
func: z
.string()
.describe(
`The function to invoke. The functions available are: ${functions
.map((func) => func.name)
.join(', ')}`,
),
}),
func: async ({ func = '' }) => {
const result = await chain.run(`${message}${func?.length > 0 ? `\nUse ${func}` : ''}`);
return result;
},
});
}
module.exports = {
getSpec,
readSpecFile,
createOpenAPIPlugin,
};