🧹 chore: Optimize provisioning — single credential load, deferred DB writes

- loadCodeApiKey: load CODE_API_KEY once per request, pass to both
  checkSessionsAlive and provisionToCodeEnv (was N+1 lookups)
- provisionToCodeEnv/provisionToVectorDB now return fileUpdate objects
  instead of writing to DB immediately
- primeResources batches all DB updates via Promise.allSettled after
  provisioning completes
- Remove updateFile import from provision.js (no longer writes directly)
This commit is contained in:
Danny Avila 2026-03-22 12:54:26 -04:00
parent f3f10e9806
commit 4080f65f3c
4 changed files with 109 additions and 48 deletions

View file

@ -40,6 +40,7 @@ const {
enrichLoadedToolsWithAgentContext,
} = require('./skillDeps');
const {
loadCodeApiKey,
provisionToCodeEnv,
provisionToVectorDB,
checkSessionsAlive,
@ -409,6 +410,8 @@ const initializeClient = async ({ req, res, signal, endpointOption }) => {
provisionToCodeEnv,
provisionToVectorDB,
checkSessionsAlive,
loadCodeApiKey,
updateFile: db.updateFile,
},
);
@ -488,6 +491,8 @@ const initializeClient = async ({ req, res, signal, endpointOption }) => {
provisionToCodeEnv,
provisionToVectorDB,
checkSessionsAlive,
loadCodeApiKey,
updateFile: db.updateFile,
},
// The callback fires during BFS, before the helper prunes agents
// whose edges end up filtered. Don't populate `agentConfigs` here —

View file

@ -12,10 +12,21 @@ const { logger } = require('@librechat/data-schemas');
const { FileSources } = require('librechat-data-provider');
const { loadAuthValues } = require('~/server/services/Tools/credentials');
const { getStrategyFunctions } = require('./strategies');
const { updateFile } = require('~/models');
const axios = createAxiosInstance();
/**
* Loads the CODE_API_KEY for a user. Call once per request and pass the result
* to provisionToCodeEnv / checkSessionsAlive to avoid redundant lookups.
*
* @param {string} userId
* @returns {Promise<string>} The CODE_API_KEY
*/
async function loadCodeApiKey(userId) {
const result = await loadAuthValues({ userId, authFields: [EnvVar.CODE_API_KEY] });
return result[EnvVar.CODE_API_KEY];
}
/**
* Provisions a file to the code execution environment.
* Gets a read stream from our storage and uploads to the code env.
@ -24,9 +35,10 @@ const axios = createAxiosInstance();
* @param {object} params.req - Express request object (needs req.user.id)
* @param {import('librechat-data-provider').TFile} params.file - The file record from DB
* @param {string} [params.entity_id] - Optional entity ID (agent_id)
* @returns {Promise<string>} The fileIdentifier from the code env
* @param {string} [params.apiKey] - Pre-loaded CODE_API_KEY (avoids redundant loadAuthValues)
* @returns {Promise<{ fileIdentifier: string, fileUpdate: object }>} Result with deferred DB update
*/
async function provisionToCodeEnv({ req, file, entity_id = '' }) {
async function provisionToCodeEnv({ req, file, entity_id = '', apiKey }) {
const { getDownloadStream } = getStrategyFunctions(file.source);
if (!getDownloadStream) {
throw new Error(
@ -34,33 +46,26 @@ async function provisionToCodeEnv({ req, file, entity_id = '' }) {
);
}
const resolvedApiKey = apiKey ?? (await loadCodeApiKey(req.user.id));
const { handleFileUpload: uploadCodeEnvFile } = getStrategyFunctions(FileSources.execute_code);
const result = await loadAuthValues({ userId: req.user.id, authFields: [EnvVar.CODE_API_KEY] });
const stream = await getDownloadStream(req, file.filepath);
const fileIdentifier = await uploadCodeEnvFile({
req,
stream,
filename: file.filename,
apiKey: result[EnvVar.CODE_API_KEY],
apiKey: resolvedApiKey,
entity_id,
});
const updatedMetadata = {
...file.metadata,
fileIdentifier,
};
await updateFile({
file_id: file.file_id,
metadata: updatedMetadata,
});
logger.debug(
`[provisionToCodeEnv] Provisioned file "${file.filename}" (${file.file_id}) to code env`,
);
return fileIdentifier;
return {
fileIdentifier,
fileUpdate: { file_id: file.file_id, metadata: { ...file.metadata, fileIdentifier } },
};
}
/**
@ -71,27 +76,31 @@ async function provisionToCodeEnv({ req, file, entity_id = '' }) {
* @param {object} params.req - Express request object
* @param {import('librechat-data-provider').TFile} params.file - The file record from DB
* @param {string} [params.entity_id] - Optional entity ID (agent_id)
* @returns {Promise<{ embedded: boolean }>} Embedding result
* @param {import('stream').Readable} [params.existingStream] - Pre-fetched download stream (avoids duplicate storage fetch)
* @returns {Promise<{ embedded: boolean, fileUpdate: object | null }>} Result with deferred DB update
*/
async function provisionToVectorDB({ req, file, entity_id }) {
async function provisionToVectorDB({ req, file, entity_id, existingStream }) {
if (!process.env.RAG_API_URL) {
logger.warn('[provisionToVectorDB] RAG_API_URL not defined, skipping vector provisioning');
return { embedded: false };
return { embedded: false, fileUpdate: null };
}
const { getDownloadStream } = getStrategyFunctions(file.source);
if (!getDownloadStream) {
throw new Error(
`Cannot provision file "${file.filename}" to vector DB: storage source "${file.source}" does not support download streams`,
);
}
// The uploadVectors function expects a file-like object with a `path` property for fs.createReadStream.
// Since we're provisioning from storage (not a multer upload), we need to stream to a temp file first.
const tmpPath = path.join(os.tmpdir(), `provision-${file.file_id}${path.extname(file.filename)}`);
try {
const stream = await getDownloadStream(req, file.filepath);
let stream = existingStream;
if (!stream) {
const { getDownloadStream } = getStrategyFunctions(file.source);
if (!getDownloadStream) {
throw new Error(
`Cannot provision file "${file.filename}" to vector DB: storage source "${file.source}" does not support download streams`,
);
}
stream = await getDownloadStream(req, file.filepath);
}
// uploadVectors expects a file-like object with a `path` property for fs.createReadStream.
// Since we're provisioning from storage (not a multer upload), we stream to a temp file first.
await new Promise((resolve, reject) => {
const writeStream = fs.createWriteStream(tmpPath);
stream.pipe(writeStream);
@ -117,18 +126,15 @@ async function provisionToVectorDB({ req, file, entity_id }) {
const embedded = embeddingResult?.embedded ?? false;
await updateFile({
file_id: file.file_id,
embedded,
});
logger.debug(
`[provisionToVectorDB] Provisioned file "${file.filename}" (${file.file_id}) to vector DB, embedded=${embedded}`,
);
return { embedded };
return {
embedded,
fileUpdate: embedded ? { file_id: file.file_id, embedded } : null,
};
} finally {
// Clean up temp file
try {
fs.unlinkSync(tmpPath);
} catch {
@ -190,13 +196,11 @@ async function checkCodeEnvFileAlive({ file, apiKey }) {
*
* @param {object} params
* @param {import('librechat-data-provider').TFile[]} params.files - Files with metadata.fileIdentifier
* @param {string} params.userId - User ID for loading CODE_API_KEY
* @param {string} params.apiKey - Pre-loaded CODE_API_KEY
* @param {number} [params.staleSafeWindowMs=21600000] - Skip check if file updated within this window (default 6h)
* @returns {Promise<Set<string>>} Set of file_ids that are confirmed alive
*/
async function checkSessionsAlive({ files, userId, staleSafeWindowMs = 6 * 60 * 60 * 1000 }) {
const result = await loadAuthValues({ userId, authFields: [EnvVar.CODE_API_KEY] });
const apiKey = result[EnvVar.CODE_API_KEY];
async function checkSessionsAlive({ files, apiKey, staleSafeWindowMs = 6 * 60 * 60 * 1000 }) {
const aliveFileIds = new Set();
const now = Date.now();
@ -263,6 +267,7 @@ async function checkSessionsAlive({ files, userId, staleSafeWindowMs = 6 * 60 *
}
module.exports = {
loadCodeApiKey,
provisionToCodeEnv,
provisionToVectorDB,
checkCodeEnvFileAlive,

View file

@ -32,10 +32,12 @@ import type {
import type { LCAvailableTools, RequestScopedMCPConnectionStore } from '../mcp/types';
import type { ResolvedManualSkill, ResolvedAlwaysApplySkill } from './skills';
import type {
TFileUpdate,
TFilterFilesByAgentAccess,
TProvisionToCodeEnv,
TProvisionToVectorDB,
TCheckSessionsAlive,
TLoadCodeApiKey,
} from './resources';
import {
injectSkillCatalog,
@ -555,6 +557,10 @@ export interface InitializeAgentDbMethods extends EndpointDbMethods {
provisionToVectorDB?: TProvisionToVectorDB;
/** Optional: batch-check code env file liveness */
checkSessionsAlive?: TCheckSessionsAlive;
/** Optional: load CODE_API_KEY once per request */
loadCodeApiKey?: TLoadCodeApiKey;
/** Optional: persist file metadata updates after provisioning */
updateFile?: (data: TFileUpdate) => Promise<unknown>;
}
/**
@ -772,6 +778,8 @@ export async function initializeAgent(
provisionToCodeEnv: db.provisionToCodeEnv,
provisionToVectorDB: db.provisionToVectorDB,
checkSessionsAlive: db.checkSessionsAlive,
loadCodeApiKey: db.loadCodeApiKey,
updateFile: db.updateFile as ((data: TFileUpdate) => Promise<unknown>) | undefined,
});
/**

View file

@ -5,25 +5,34 @@ import type { IMongoFile, AppConfig, IUser } from '@librechat/data-schemas';
import type { FilterQuery, QueryOptions, ProjectionType } from 'mongoose';
import type { Request as ServerRequest } from 'express';
/** Deferred DB update from provisioning (batched after all files are provisioned) */
export type TFileUpdate = {
file_id: string;
metadata?: Record<string, unknown>;
embedded?: boolean;
};
/**
* Function type for provisioning a file to the code execution environment.
* @returns The fileIdentifier from the code env
* @returns The fileIdentifier and a deferred DB update object
*/
export type TProvisionToCodeEnv = (params: {
req: ServerRequest & { user?: IUser };
file: TFile;
entity_id?: string;
}) => Promise<string>;
apiKey?: string;
}) => Promise<{ fileIdentifier: string; fileUpdate: TFileUpdate }>;
/**
* Function type for provisioning a file to the vector DB for file_search.
* @returns Object with embedded status
* @returns Object with embedded status and a deferred DB update object
*/
export type TProvisionToVectorDB = (params: {
req: ServerRequest & { user?: IUser };
file: TFile;
entity_id?: string;
}) => Promise<{ embedded: boolean }>;
existingStream?: unknown;
}) => Promise<{ embedded: boolean; fileUpdate: TFileUpdate | null }>;
/**
* Function type for batch-checking code env file liveness.
@ -32,10 +41,13 @@ export type TProvisionToVectorDB = (params: {
*/
export type TCheckSessionsAlive = (params: {
files: TFile[];
userId: string;
apiKey: string;
staleSafeWindowMs?: number;
}) => Promise<Set<string>>;
/** Loads CODE_API_KEY for a user. Call once per request. */
export type TLoadCodeApiKey = (userId: string) => Promise<string>;
/**
* Function type for retrieving files from the database
* @param filter - MongoDB filter query for files
@ -196,6 +208,8 @@ export const primeResources = async ({
provisionToCodeEnv,
provisionToVectorDB,
checkSessionsAlive,
loadCodeApiKey,
updateFile,
}: {
req: ServerRequest & { user?: IUser };
appConfig?: AppConfig;
@ -213,6 +227,10 @@ export const primeResources = async ({
provisionToVectorDB?: TProvisionToVectorDB;
/** Optional callback to batch-check code env file liveness by session */
checkSessionsAlive?: TCheckSessionsAlive;
/** Optional callback to load CODE_API_KEY once per request */
loadCodeApiKey?: TLoadCodeApiKey;
/** Optional callback to persist file metadata updates after provisioning */
updateFile?: (data: TFileUpdate) => Promise<unknown>;
}): Promise<{
attachments: Array<TFile | undefined> | undefined;
requestAttachments: Array<TFile | undefined> | undefined;
@ -386,20 +404,34 @@ export const primeResources = async ({
enabledToolResources.has(EToolResources.file_search) && provisionToVectorDB != null;
if (needsCodeEnv || needsVectorDB) {
// Load CODE_API_KEY once for all code env operations
let codeApiKey: string | undefined;
if (needsCodeEnv && loadCodeApiKey && req.user?.id) {
try {
codeApiKey = await loadCodeApiKey(req.user.id);
} catch (error) {
logger.error('[primeResources] Failed to load CODE_API_KEY', error);
warnings.push('Code execution file provisioning unavailable');
}
}
// Batch staleness check: verify code env files are still alive
let aliveFileIds: Set<string> = new Set();
if (needsCodeEnv && checkSessionsAlive && req.user?.id) {
if (needsCodeEnv && codeApiKey && checkSessionsAlive) {
const filesWithIdentifiers = attachments.filter(
(f) => f?.metadata?.fileIdentifier && f.file_id,
);
if (filesWithIdentifiers.length > 0) {
aliveFileIds = await checkSessionsAlive({
files: filesWithIdentifiers as TFile[],
userId: req.user.id,
apiKey: codeApiKey,
});
}
}
// Collect deferred DB updates from provisioning
const pendingUpdates: TFileUpdate[] = [];
// Provision files in parallel
const provisionResults = await Promise.allSettled(
attachments.map(async (file) => {
@ -413,6 +445,7 @@ export const primeResources = async ({
// Code env provisioning (with staleness check)
if (
needsCodeEnv &&
codeApiKey &&
!processedResourceFiles.has(`${EToolResources.execute_code}:${file.file_id}`)
) {
const hasFileIdentifier = !!file.metadata?.fileIdentifier;
@ -428,12 +461,14 @@ export const primeResources = async ({
}
try {
const fileIdentifier = await provisionToCodeEnv({
const { fileIdentifier, fileUpdate } = await provisionToCodeEnv({
req: typedReq,
file,
entity_id: agentId,
apiKey: codeApiKey,
});
file.metadata = { ...file.metadata, fileIdentifier };
pendingUpdates.push(fileUpdate);
addFileToResource({
file,
resourceType: EToolResources.execute_code,
@ -471,6 +506,9 @@ export const primeResources = async ({
});
if (result.embedded) {
file.embedded = true;
if (result.fileUpdate) {
pendingUpdates.push(result.fileUpdate);
}
addFileToResource({
file,
resourceType: EToolResources.file_search,
@ -493,6 +531,11 @@ export const primeResources = async ({
logger.error('[primeResources] Unexpected provisioning rejection', result.reason);
}
}
// Batch DB updates after all provisioning completes
if (pendingUpdates.length > 0 && updateFile) {
await Promise.allSettled(pendingUpdates.map((update) => updateFile(update)));
}
}
}