Haay's picture
Upload 926 files
519a20c verified
import path from 'node:path';
import fs from 'node:fs';
import vectra from 'vectra';
import express from 'express';
import sanitize from 'sanitize-filename';
import { getConfigValue } from '../util.js';
import { getNomicAIBatchVector, getNomicAIVector } from '../vectors/nomicai-vectors.js';
import { getOpenAIVector, getOpenAIBatchVector } from '../vectors/openai-vectors.js';
import { getTransformersVector, getTransformersBatchVector } from '../vectors/embedding.js';
import { getExtrasVector, getExtrasBatchVector } from '../vectors/extras-vectors.js';
import { getMakerSuiteVector, getMakerSuiteBatchVector } from '../vectors/makersuite-vectors.js';
import { getCohereVector, getCohereBatchVector } from '../vectors/cohere-vectors.js';
import { getLlamaCppVector, getLlamaCppBatchVector } from '../vectors/llamacpp-vectors.js';
import { getVllmVector, getVllmBatchVector } from '../vectors/vllm-vectors.js';
import { getOllamaVector, getOllamaBatchVector } from '../vectors/ollama-vectors.js';
// Don't forget to add new sources to the SOURCES array
const SOURCES = [
'transformers',
'mistral',
'openai',
'extras',
'palm',
'togetherai',
'nomicai',
'cohere',
'ollama',
'llamacpp',
'vllm',
'webllm',
'koboldcpp',
];
/**
* Gets the vector for the given text from the given source.
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {string} text - The text to get the vector for
* @param {boolean} isQuery - If the text is a query for embedding search
* @param {import('../users.js').UserDirectoryList} directories - The directories object for the user
* @returns {Promise<number[]>} - The vector for the text
*/
async function getVector(source, sourceSettings, text, isQuery, directories) {
switch (source) {
case 'nomicai':
return getNomicAIVector(text, source, directories);
case 'togetherai':
case 'mistral':
case 'openai':
return getOpenAIVector(text, source, directories, sourceSettings.model);
case 'transformers':
return getTransformersVector(text);
case 'extras':
return getExtrasVector(text, sourceSettings.extrasUrl, sourceSettings.extrasKey);
case 'palm':
return getMakerSuiteVector(text, directories);
case 'cohere':
return getCohereVector(text, isQuery, directories, sourceSettings.model);
case 'llamacpp':
return getLlamaCppVector(text, sourceSettings.apiUrl, directories);
case 'vllm':
return getVllmVector(text, sourceSettings.apiUrl, sourceSettings.model, directories);
case 'ollama':
return getOllamaVector(text, sourceSettings.apiUrl, sourceSettings.model, sourceSettings.keep, directories);
case 'webllm':
return sourceSettings.embeddings[text];
case 'koboldcpp':
return sourceSettings.embeddings[text];
}
throw new Error(`Unknown vector source ${source}`);
}
/**
* Gets the vector for the given text batch from the given source.
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {string[]} texts - The array of texts to get the vector for
* @param {boolean} isQuery - If the text is a query for embedding search
* @param {import('../users.js').UserDirectoryList} directories - The directories object for the user
* @returns {Promise<number[][]>} - The array of vectors for the texts
*/
async function getBatchVector(source, sourceSettings, texts, isQuery, directories) {
const batchSize = 10;
const batches = Array(Math.ceil(texts.length / batchSize)).fill(undefined).map((_, i) => texts.slice(i * batchSize, i * batchSize + batchSize));
let results = [];
for (let batch of batches) {
switch (source) {
case 'nomicai':
results.push(...await getNomicAIBatchVector(batch, source, directories));
break;
case 'togetherai':
case 'mistral':
case 'openai':
results.push(...await getOpenAIBatchVector(batch, source, directories, sourceSettings.model));
break;
case 'transformers':
results.push(...await getTransformersBatchVector(batch));
break;
case 'extras':
results.push(...await getExtrasBatchVector(batch, sourceSettings.extrasUrl, sourceSettings.extrasKey));
break;
case 'palm':
results.push(...await getMakerSuiteBatchVector(batch, directories));
break;
case 'cohere':
results.push(...await getCohereBatchVector(batch, isQuery, directories, sourceSettings.model));
break;
case 'llamacpp':
results.push(...await getLlamaCppBatchVector(batch, sourceSettings.apiUrl, directories));
break;
case 'vllm':
results.push(...await getVllmBatchVector(batch, sourceSettings.apiUrl, sourceSettings.model, directories));
break;
case 'ollama':
results.push(...await getOllamaBatchVector(batch, sourceSettings.apiUrl, sourceSettings.model, sourceSettings.keep, directories));
break;
case 'webllm':
results.push(...texts.map(x => sourceSettings.embeddings[x]));
break;
case 'koboldcpp':
results.push(...texts.map(x => sourceSettings.embeddings[x]));
break;
default:
throw new Error(`Unknown vector source ${source}`);
}
}
return results;
}
/**
* Extracts settings for the vectorization sources from the HTTP request headers.
* @param {string} source - Which source to extract settings for.
* @param {object} request - The HTTP request object.
* @returns {object} - An object that can be used as `sourceSettings` in functions that take that parameter.
*/
function getSourceSettings(source, request) {
switch (source) {
case 'togetherai':
return {
model: String(request.body.model),
};
case 'openai':
return {
model: String(request.body.model),
};
case 'cohere':
return {
model: String(request.body.model),
};
case 'llamacpp':
return {
apiUrl: String(request.body.apiUrl),
};
case 'vllm':
return {
apiUrl: String(request.body.apiUrl),
model: String(request.body.model),
};
case 'ollama':
return {
apiUrl: String(request.body.apiUrl),
model: String(request.body.model),
keep: Boolean(request.body.keep),
};
case 'extras':
return {
extrasUrl: String(request.body.extrasUrl),
extrasKey: String(request.body.extrasKey),
};
case 'transformers':
return {
model: getConfigValue('extensions.models.embedding', ''),
};
case 'palm':
return {
model: String(request.body.model || 'text-embedding-004'),
};
case 'mistral':
return {
model: 'mistral-embed',
};
case 'nomicai':
return {
model: 'nomic-embed-text-v1.5',
};
case 'webllm':
return {
model: String(request.body.model),
embeddings: request.body.embeddings ?? {},
};
case 'koboldcpp':
return {
model: String(request.body.model),
embeddings: request.body.embeddings ?? {},
};
default:
return {};
}
}
/**
* Gets the model scope for the source.
* @param {object} sourceSettings - The settings for the source
* @returns {string} The model scope for the source
*/
function getModelScope(sourceSettings) {
return (sourceSettings?.model || '');
}
/**
* Gets the index for the vector collection
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string} collectionId - The collection ID
* @param {string} source - The source of the vector
* @param {object} sourceSettings - The model for the source
* @returns {Promise<vectra.LocalIndex>} - The index for the collection
*/
async function getIndex(directories, collectionId, source, sourceSettings) {
const model = getModelScope(sourceSettings);
const pathToFile = path.join(directories.vectors, sanitize(source), sanitize(collectionId), sanitize(model));
const store = new vectra.LocalIndex(pathToFile);
if (!await store.isIndexCreated()) {
await store.createIndex();
}
return store;
}
/**
* Inserts items into the vector collection
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string} collectionId - The collection ID
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {{ hash: number; text: string; index: number; }[]} items - The items to insert
*/
async function insertVectorItems(directories, collectionId, source, sourceSettings, items) {
const store = await getIndex(directories, collectionId, source, sourceSettings);
await store.beginUpdate();
const vectors = await getBatchVector(source, sourceSettings, items.map(x => x.text), false, directories);
for (let i = 0; i < items.length; i++) {
const item = items[i];
const vector = vectors[i];
await store.upsertItem({ vector: vector, metadata: { hash: item.hash, text: item.text, index: item.index } });
}
await store.endUpdate();
}
/**
* Gets the hashes of the items in the vector collection
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string} collectionId - The collection ID
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @returns {Promise<number[]>} - The hashes of the items in the collection
*/
async function getSavedHashes(directories, collectionId, source, sourceSettings) {
const store = await getIndex(directories, collectionId, source, sourceSettings);
const items = await store.listItems();
const hashes = items.map(x => Number(x.metadata.hash));
return hashes;
}
/**
* Deletes items from the vector collection by hash
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string} collectionId - The collection ID
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {number[]} hashes - The hashes of the items to delete
*/
async function deleteVectorItems(directories, collectionId, source, sourceSettings, hashes) {
const store = await getIndex(directories, collectionId, source, sourceSettings);
const items = await store.listItemsByMetadata({ hash: { '$in': hashes } });
await store.beginUpdate();
for (const item of items) {
await store.deleteItem(item.id);
}
await store.endUpdate();
}
/**
* Gets the hashes of the items in the vector collection that match the search text
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string} collectionId - The collection ID
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {string} searchText - The text to search for
* @param {number} topK - The number of results to return
* @param {number} threshold - The threshold for the search
* @returns {Promise<{hashes: number[], metadata: object[]}>} - The metadata of the items that match the search text
*/
async function queryCollection(directories, collectionId, source, sourceSettings, searchText, topK, threshold) {
const store = await getIndex(directories, collectionId, source, sourceSettings);
const vector = await getVector(source, sourceSettings, searchText, true, directories);
const result = await store.queryItems(vector, topK);
const metadata = result.filter(x => x.score >= threshold).map(x => x.item.metadata);
const hashes = result.map(x => Number(x.item.metadata.hash));
return { metadata, hashes };
}
/**
* Queries multiple collections for the given search queries. Returns the overall top K results.
* @param {import('../users.js').UserDirectoryList} directories - User directories
* @param {string[]} collectionIds - The collection IDs to query
* @param {string} source - The source of the vector
* @param {Object} sourceSettings - Settings for the source, if it needs any
* @param {string} searchText - The text to search for
* @param {number} topK - The number of results to return
* @param {number} threshold - The threshold for the search
*
* @returns {Promise<Record<string, { hashes: number[], metadata: object[] }>>} - The top K results from each collection
*/
async function multiQueryCollection(directories, collectionIds, source, sourceSettings, searchText, topK, threshold) {
const vector = await getVector(source, sourceSettings, searchText, true, directories);
const results = [];
for (const collectionId of collectionIds) {
const store = await getIndex(directories, collectionId, source, sourceSettings);
const result = await store.queryItems(vector, topK);
results.push(...result.map(result => ({ collectionId, result })));
}
// Sort results by descending similarity, apply threshold, and take top K
const sortedResults = results
.sort((a, b) => b.result.score - a.result.score)
.filter(x => x.result.score >= threshold)
.slice(0, topK);
/**
* Group the results by collection ID
* @type {Record<string, { hashes: number[], metadata: object[] }>}
*/
const groupedResults = {};
for (const result of sortedResults) {
if (!groupedResults[result.collectionId]) {
groupedResults[result.collectionId] = { hashes: [], metadata: [] };
}
groupedResults[result.collectionId].hashes.push(Number(result.result.item.metadata.hash));
groupedResults[result.collectionId].metadata.push(result.result.item.metadata);
}
return groupedResults;
}
/**
* Performs a request to regenerate the index if it is corrupted.
* @param {import('express').Request} req Express request object
* @param {import('express').Response} res Express response object
* @param {Error} error Error object
* @returns {Promise<any>} Promise
*/
async function regenerateCorruptedIndexErrorHandler(req, res, error) {
if (error instanceof SyntaxError && !req.query.regenerated) {
const collectionId = String(req.body.collectionId);
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
if (collectionId && source) {
const index = await getIndex(req.user.directories, collectionId, source, sourceSettings);
const exists = await index.isIndexCreated();
if (exists) {
const path = index.folderPath;
console.warn(`Corrupted index detected at ${path}, regenerating...`);
await index.deleteIndex();
return res.redirect(307, req.originalUrl + '?regenerated=true');
}
}
}
console.error(error);
return res.sendStatus(500);
}
export const router = express.Router();
router.post('/query', async (req, res) => {
try {
if (!req.body.collectionId || !req.body.searchText) {
return res.sendStatus(400);
}
const collectionId = String(req.body.collectionId);
const searchText = String(req.body.searchText);
const topK = Number(req.body.topK) || 10;
const threshold = Number(req.body.threshold) || 0.0;
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
const results = await queryCollection(req.user.directories, collectionId, source, sourceSettings, searchText, topK, threshold);
return res.json(results);
} catch (error) {
return regenerateCorruptedIndexErrorHandler(req, res, error);
}
});
router.post('/query-multi', async (req, res) => {
try {
if (!Array.isArray(req.body.collectionIds) || !req.body.searchText) {
return res.sendStatus(400);
}
const collectionIds = req.body.collectionIds.map(x => String(x));
const searchText = String(req.body.searchText);
const topK = Number(req.body.topK) || 10;
const threshold = Number(req.body.threshold) || 0.0;
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
const results = await multiQueryCollection(req.user.directories, collectionIds, source, sourceSettings, searchText, topK, threshold);
return res.json(results);
} catch (error) {
return regenerateCorruptedIndexErrorHandler(req, res, error);
}
});
router.post('/insert', async (req, res) => {
try {
if (!Array.isArray(req.body.items) || !req.body.collectionId) {
return res.sendStatus(400);
}
const collectionId = String(req.body.collectionId);
const items = req.body.items.map(x => ({ hash: x.hash, text: x.text, index: x.index }));
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
await insertVectorItems(req.user.directories, collectionId, source, sourceSettings, items);
return res.sendStatus(200);
} catch (error) {
return regenerateCorruptedIndexErrorHandler(req, res, error);
}
});
router.post('/list', async (req, res) => {
try {
if (!req.body.collectionId) {
return res.sendStatus(400);
}
const collectionId = String(req.body.collectionId);
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
const hashes = await getSavedHashes(req.user.directories, collectionId, source, sourceSettings);
return res.json(hashes);
} catch (error) {
return regenerateCorruptedIndexErrorHandler(req, res, error);
}
});
router.post('/delete', async (req, res) => {
try {
if (!Array.isArray(req.body.hashes) || !req.body.collectionId) {
return res.sendStatus(400);
}
const collectionId = String(req.body.collectionId);
const hashes = req.body.hashes.map(x => Number(x));
const source = String(req.body.source) || 'transformers';
const sourceSettings = getSourceSettings(source, req);
await deleteVectorItems(req.user.directories, collectionId, source, sourceSettings, hashes);
return res.sendStatus(200);
} catch (error) {
return regenerateCorruptedIndexErrorHandler(req, res, error);
}
});
router.post('/purge-all', async (req, res) => {
try {
for (const source of SOURCES) {
const sourcePath = path.join(req.user.directories.vectors, sanitize(source));
if (!fs.existsSync(sourcePath)) {
continue;
}
await fs.promises.rm(sourcePath, { recursive: true });
console.info(`Deleted vector source store at ${sourcePath}`);
}
return res.sendStatus(200);
} catch (error) {
console.error(error);
return res.sendStatus(500);
}
});
router.post('/purge', async (req, res) => {
try {
if (!req.body.collectionId) {
return res.sendStatus(400);
}
const collectionId = String(req.body.collectionId);
for (const source of SOURCES) {
const sourcePath = path.join(req.user.directories.vectors, sanitize(source), sanitize(collectionId));
if (!fs.existsSync(sourcePath)) {
continue;
}
await fs.promises.rm(sourcePath, { recursive: true });
console.info(`Deleted vector index at ${sourcePath}`);
}
return res.sendStatus(200);
} catch (error) {
console.error(error);
return res.sendStatus(500);
}
});