Skip to content

Enhanced ChatGPT Clone: Features Agents, MCP, Skills, DeepSeek, Anthropic, AWS, OpenAI, Responses API, Azure, Groq, o1, GPT-5, Mistral, OpenRouter, Vertex AI, Gemini, Artifacts, AI model switching, message search, Code Interpreter, langchain, DALL-E-3, OpenAPI Actions, Functions, Secure Multi-User Auth, Presets, open-source for self-hosting. Active

🛼 fix: Bound Document Name Collision Processing (#17386)

lia-by-librechat[bot]68826ec343d7
diff --git a/api/server/utils/import/importers.spec.js b/api/server/utils/import/importers.spec.js
index ca6e0686fa..21dfe36ad9 100644
--- a/api/server/utils/import/importers.spec.js
+++ b/api/server/utils/import/importers.spec.js
@@ -1,4 +1,6 @@
const fs = require('fs');
+const { formatMessage } = require('@librechat/agents');
+const { dedupeDocumentNames, ContentTraversalLimitError } = require('@librechat/api');
const path = require('path');
const {
EModelEndpoint,
@@ -60,6 +62,41 @@ it('strips server-private owner metadata from untrusted LibreChat imports', asyn
expect(importBatchBuilder.messages[0]).not.toHaveProperty('privacyRevision');
});
+it.each([2000, 8000])('bounds replay of %i imported inline document blocks', async (count) => {
+ const content = Array.from({ length: count / 2 }, (_, index) => {
+ const name = 'a'.repeat(196) + index.toString(36).padStart(4, '0');
+ const document = { type: 'document', document: { name } };
+ return [document, structuredClone(document)];
+ }).flat();
+ const jsonData = {
+ conversationId: 'imported-documents',
+ title: 'Imported documents',
+ endpoint: EModelEndpoint.agents,
+ messages: [
+ {
+ messageId: 'source',
+ parentMessageId: Constants.NO_PARENT,
+ isCreatedByUser: false,
+ content,
+ },
+ ],
+ };
+ const builder = new ImportBatchBuilder('owner');
+ await getImporter(jsonData)(jsonData, 'owner', () => builder);
+ const imported = builder.messages[0];
+ expect(imported.content).toEqual(content);
+ const payload = [formatMessage({ message: imported })];
+
+ if (count === 8000) {
+ expect(() => dedupeDocumentNames(payload)).toThrow(ContentTraversalLimitError);
+ } else {
+ dedupeDocumentNames(payload);
+ const names = payload[0].content.map((part) => part.document.name.toLowerCase());
+ expect(new Set(names).size).toBe(count);
+ }
+ expect(imported.content).toEqual(content);
+});
+
describe('importChatGptConvo', () => {
it('should import conversation correctly', async () => {
const expectedNumberOfMessages = 19;
diff --git a/packages/api/src/files/encode/document.spec.ts b/packages/api/src/files/encode/document.spec.ts
index c73423a3ae..66a86bf27e 100644
--- a/packages/api/src/files/encode/document.spec.ts
+++ b/packages/api/src/files/encode/document.spec.ts
@@ -3,6 +3,11 @@ import { mbToBytes } from 'librechat-data-provider';
import type { AppConfig, IMongoFile } from '@librechat/data-schemas';
import type { BedrockDocumentBlock, ServerRequest } from '~/types';
import type { PayloadContentPart } from './document';
+import {
+ CONTENT_TRAVERSAL_MAX_DEPTH,
+ CONTENT_TRAVERSAL_MAX_NODES,
+ ContentTraversalLimitError,
+} from '~/protection/adapters/nested';
import { encodeAndFormatDocuments, dedupeDocumentNames } from './document';
/** Mock the validation module */
@@ -1282,6 +1287,104 @@ describe('dedupeDocumentNames', () => {
expect(second[2].length).toBeLessThanOrEqual(200);
});
+ it.each([1000, 2000])('bounds collision lookups for %i shared-prefix documents', (count) => {
+ const content = Array.from({ length: count / 2 }, (_, index) => {
+ const name = 'a'.repeat(196) + index.toString(36).padStart(4, '0');
+ return [bedrock(name), bedrock(name)];
+ }).flat();
+ const messages: Message[] = [{ content }];
+ const has = jest.spyOn(Set.prototype, 'has');
+
+ dedupeDocumentNames(messages);
+
+ const lookups = has.mock.calls.length;
+ has.mockRestore();
+ const result = names(contentOf(messages[0]));
+ expect(lookups).toBeLessThan(count * 6);
+ expect(new Set(result.map((name) => name.toLowerCase())).size).toBe(count);
+ expect(result.every((name) => name.length <= 200)).toBe(true);
+ expect(names(content)).toEqual(content.map((part) => part.document.name));
+ });
+
+ it('shares progress after suffix digit widths change and skips occupied suffixes', () => {
+ const prefix = 'A'.repeat(196);
+ const occupied = Array.from({ length: 1000 }, (_, index) => {
+ const suffix = ` (${index + 2})`;
+ return bedrock(prefix.slice(0, 200 - suffix.length) + suffix);
+ });
+ const originals = [prefix + '0000', prefix.toLowerCase() + '0001', prefix + ' '];
+ const content = [...occupied, ...originals.flatMap((name) => [bedrock(name), bedrock(name)])];
+ const messages: Message[] = [{ content }];
+
+ dedupeDocumentNames(messages);
+
+ const result = names(contentOf(messages[0]));
+ expect(result.slice(-6)).toEqual([
+ originals[0],
+ `${prefix.slice(0, 193)} (1002)`,
+ originals[1],
+ `${prefix.toLowerCase().slice(0, 193)} (1003)`,
+ originals[2],
+ `${prefix.slice(0, 193)} (1004)`,
+ ]);
+ expect(new Set(result.map((name) => name.toLowerCase())).size).toBe(content.length);
+ });
+
+ it('bounds imported names before case folding and resolves truncation collisions', () => {
+ const prefix = 'a'.repeat(200);
+ const messages: Message[] = [
+ { content: [bedrock(prefix + 'x'.repeat(100000)), bedrock(prefix)] },
+ ];
+
+ dedupeDocumentNames(messages);
+
+ expect(names(contentOf(messages[0]))).toEqual([prefix, `${prefix.slice(0, 196)} (2)`]);
+ });
+
+ it('rejects oversized replay arrays before copying or naming their parts', () => {
+ const content = [bedrock('notes'), bedrock('notes')];
+ content.length = CONTENT_TRAVERSAL_MAX_NODES;
+ const messages: Message[] = [{ content }];
+
+ expect(() => dedupeDocumentNames(messages)).toThrow(ContentTraversalLimitError);
+ expect(messages[0].content).toBe(content);
+ expect(content[1].document.name).toBe('notes');
+ });
+
+ it('shares the part budget across messages and nested steer media', () => {
+ const media: PayloadContentPart[] = new Array(CONTENT_TRAVERSAL_MAX_NODES - 3);
+ const messages: Message[] = [
+ { content: [{ type: 'steer', media }] },
+ { content: [bedrock('x')] },
+ ];
+
+ expect(() => dedupeDocumentNames(messages)).toThrow(ContentTraversalLimitError);
+ });
+
+ it('accepts a replay exactly at the node budget and bounds the message array', () => {
+ const messages: Message[] = [{ content: new Array(CONTENT_TRAVERSAL_MAX_NODES - 1) }];
+
+ expect(() => dedupeDocumentNames(messages)).not.toThrow();
+ expect(() => dedupeDocumentNames(new Array(CONTENT_TRAVERSAL_MAX_NODES + 1))).toThrow(
+ ContentTraversalLimitError,
+ );
+ });
+
+ it('bounds recursive media and cycles without overflowing the stack', () => {
+ const root: PayloadContentPart = { type: 'steer', media: [] };
+ let part = root;
+ for (let depth = 0; depth < CONTENT_TRAVERSAL_MAX_DEPTH; depth++) {
+ const child: PayloadContentPart = { type: 'steer', media: [] };
+ part.media = [child];
+ part = child;
+ }
+ const messages: Message[] = [{ content: [root] }];
+
+ expect(() => dedupeDocumentNames(messages)).toThrow(ContentTraversalLimitError);
+ root.media = [root];
+ expect(() => dedupeDocumentNames(messages)).toThrow(ContentTraversalLimitError);
+ });
+
it('leaves blocks without a Bedrock document name alone', () => {
const anthropic = {
type: 'document',
diff --git a/packages/api/src/files/encode/document.ts b/packages/api/src/files/encode/document.ts
index 7fa7323499..27444f6446 100644
--- a/packages/api/src/files/encode/document.ts
+++ b/packages/api/src/files/encode/document.ts
@@ -15,6 +15,11 @@ import type {
DocumentResult,
ServerRequest,
} from '~/types';
+import {
+ CONTENT_TRAVERSAL_MAX_DEPTH,
+ CONTENT_TRAVERSAL_MAX_NODES,
+ ContentTraversalLimitError,
+} from '~/protection/adapters/nested';
import {
getFileStream,
getConfiguredFileSizeLimit,
@@ -34,25 +39,26 @@ export interface PayloadContentPart {
media?: PayloadContentPart[];
}
-/** `name (n)`, trimmed so the suffix fits the limit and never follows a space. */
-function numberedDocumentName(name: string, index: number): string {
- const suffix = ` (${index})`;
- return `${name.slice(0, BEDROCK_DOCUMENT_NAME_LIMIT - suffix.length).trimEnd()}${suffix}`;
-}
-
-/** Hands out each name once, numbering repeats in the order they are asked for. */
+/** Hands out each name once, sharing suffix progress between colliding truncated prefixes. */
function createDocumentNamer(): (name: string) => string {
const used = new Set<string>();
const nextIndex = new Map<string, number>();
- return (original) => {
- const baseKey = original.toLowerCase();
+ return (input) => {
+ const original = input.slice(0, BEDROCK_DOCUMENT_NAME_LIMIT);
let name = original;
- let index = nextIndex.get(baseKey) ?? 2;
+ let index = 2;
while (used.has(name.toLowerCase())) {
- name = numberedDocumentName(original, index);
- index++;
+ const suffix = ` (${index})`;
+ const prefix = original.slice(0, BEDROCK_DOCUMENT_NAME_LIMIT - suffix.length).trimEnd();
+ const key = `${suffix.length}:${prefix.toLowerCase()}`;
+ const next = nextIndex.get(key);
+ if (next != null && next > index) {
+ index = next;
+ continue;
+ }
+ name = `${prefix}${suffix}`;
+ nextIndex.set(key, ++index);
}
- nextIndex.set(baseKey, index);
used.add(name.toLowerCase());
return name;
};
@@ -62,7 +68,13 @@ function createDocumentNamer(): (name: string) => string {
function renameDocumentParts(
parts: PayloadContentPart[],
nameOf: (name: string) => string,
+ budget: { remaining: number },
+ depth = 0,
): PayloadContentPart[] {
+ if (depth > CONTENT_TRAVERSAL_MAX_DEPTH || parts.length > budget.remaining) {
+ throw new ContentTraversalLimitError();
+ }
+ budget.remaining -= parts.length;
let renamed: PayloadContentPart[] | undefined;
for (let i = 0; i < parts.length; i++) {
const part = parts[i];
@@ -72,7 +84,7 @@ function renameDocumentParts(
const name = nameOf(original);
next = name === original ? part : { ...part, document: { ...part.document, name } };
} else if (Array.isArray(part?.media)) {
- const media = renameDocumentParts(part.media, nameOf);
+ const media = renameDocumentParts(part.media, nameOf, budget, depth + 1);
next = media === part.media ? part : { ...part, media };
}
if (next !== part) {
@@ -90,17 +102,21 @@ function renameDocumentParts(
* formatted payload in conversation order (`report`, `report (2)`), leaving the first occurrence,
* every other kind of part, and the parts themselves untouched: a message whose documents change
* gets a new content array. Names compare case-insensitively, as the restriction is not
- * documented to be case-sensitive.
+ * documented to be case-sensitive. Replay always enforces traversal limits, even without filters.
*/
export function dedupeDocumentNames(
messages: Array<{ content?: string | PayloadContentPart[] | null }>,
): void {
+ const budget = { remaining: CONTENT_TRAVERSAL_MAX_NODES - messages.length };
+ if (budget.remaining < 0) {
+ throw new ContentTraversalLimitError();
+ }
const nameOf = createDocumentNamer();
for (const message of messages) {
if (!Array.isArray(message?.content)) {
continue;
}
- const content = renameDocumentParts(message.content, nameOf);
+ const content = renameDocumentParts(message.content, nameOf, budget);
if (content !== message.content) {
message.content = content;
}