feat(rag): language-aware EUR-Lex source links on regulation citations (#424 phase 1)

backend/controllers/ragController.js

@@ -12,7 +12,7 @@ import logger from '../config/logger.js';
  * This controller only handles HTTP request/response shaping.
  */
 export const askQuestion = catchAsync(async (req, res) => {
-  const { question, conversationId, filters, useIntentAware = true, forceIntent } = req.body;
+  const { question, conversationId, filters, useIntentAware = true, forceIntent, lang } = req.body;
 
   if (!conversationId) {
     return sendError(res, 400, 'conversationId is required');
@@ -29,6 +29,7 @@ export const askQuestion = catchAsync(async (req, res) => {
       authorizedWorkspaceIds: req.authorizedWorkspaces?.map((w) => w.workspaceId) || [],
       forceIntent: forceIntent || null,
       useIntentAware,
+      lang,
     });
 
     // B1: honor the workspace canViewSources permission before returning sources.
@@ -54,7 +55,7 @@ export const askQuestion = catchAsync(async (req, res) => {
  * progressively.
  */
 export const askQuestionStream = catchAsync(async (req, res) => {
-  const { question, conversationId, filters, useIntentAware = true, forceIntent } = req.body;
+  const { question, conversationId, filters, useIntentAware = true, forceIntent, lang } = req.body;
 
   if (!conversationId) {
     return sendError(res, 400, 'conversationId is required');
@@ -92,6 +93,7 @@ export const askQuestionStream = catchAsync(async (req, res) => {
       authorizedWorkspaceIds: req.authorizedWorkspaces?.map((w) => w.workspaceId) || [],
       forceIntent: forceIntent || null,
       useIntentAware,
+      lang,
       onEvent: send,
     });
   } catch (error) {

backend/services/rag.js

@@ -586,7 +586,7 @@ class RAGService {
     // unseeded compliance_kb) doesn't take down the other.
     const [vendorResult, regulationResult] = await Promise.allSettled([
       this.vectorStore.similaritySearch(searchQuery, 15, qdrantFilter),
-      this.retrieveRegulationDocs(searchQuery, 5),
+      this.retrieveRegulationDocs(searchQuery, 5, options.lang),
     ]);
 
     const vendorDocsRaw = vendorResult.status === 'fulfilled' ? vendorResult.value : [];

backend/services/rag/complianceKbRetriever.js

@@ -59,11 +59,33 @@ async function getComplianceKbStore(embeddings = defaultEmbeddings) {
   return cachedStorePromise;
 }
 
+// EUR-Lex CELEX identifiers for the regulations in the KB. DORA-RTS spans
+// several delegated acts with no single reliable CELEX, so it gets no direct
+// link (better no link than a misleading one).
+const CELEX_BY_REGULATION = { DORA: '32022R2554' };
+
+/**
+ * Official EUR-Lex source URL for a regulation, in the user's language (#424).
+ * EUR-Lex serves the verbatim text in all EU languages, so a French user gets a
+ * link to the official French DORA text. Returns undefined when we have no
+ * reliable CELEX for that regulation.
+ */
+export function eurLexUrl(regulation, lang) {
+  const celex = CELEX_BY_REGULATION[regulation];
+  if (!celex) return undefined;
+  const langCode = String(lang || 'en')
+    .toLowerCase()
+    .startsWith('fr')
+    ? 'FR'
+    : 'EN';
+  return `https://eur-lex.europa.eu/legal-content/${langCode}/TXT/?uri=CELEX:${celex}`;
+}
+
 /**
  * Map a raw compliance_kb document to the shape the RAG context formatter
  * expects, so citations render as the article number rather than "Untitled".
  */
-function adaptRegulationDoc(doc) {
+function adaptRegulationDoc(doc, lang) {
   const meta = doc.metadata || {};
   const regulation = meta.regulation || 'Regulation';
   const article = meta.article || '';
@@ -72,6 +94,8 @@ function adaptRegulationDoc(doc) {
   const documentTitle =
     [regulation, article].filter(Boolean).join(' ') + (title ? `: ${title}` : '');
 
+  const url = eurLexUrl(regulation, lang);
+
   return {
     pageContent: doc.pageContent,
     metadata: {
@@ -80,6 +104,7 @@ function adaptRegulationDoc(doc) {
       documentTitle: documentTitle || regulation,
       heading_path: [regulation, article].filter(Boolean),
       documentType: 'regulation',
+      ...(url ? { url } : {}),
     },
   };
 }
@@ -93,15 +118,15 @@ function adaptRegulationDoc(doc) {
  * @param {number} [k=5]
  * @returns {Promise<Array<{pageContent: string, metadata: object}>>}
  */
-export async function retrieveRegulationDocs(query, k = 5) {
+export async function retrieveRegulationDocs(query, k = 5, lang = 'en') {
   if (!query || typeof query !== 'string') return [];
 
   const store = await getComplianceKbStore();
   if (!store) return [];
 
   try {
     const docs = await store.similaritySearch(query, k);
-    return docs.map(adaptRegulationDoc);
+    return docs.map((doc) => adaptRegulationDoc(doc, lang));
   } catch (error) {
     logger.warn('compliance_kb similarity search failed', {
       service: 'compliance-kb-retriever',

backend/services/ragExecutor.js

@@ -33,6 +33,7 @@ export async function executeRAG({
   onEvent = null,
   userId = null,
   authorizedWorkspaceIds = null,
+  lang = 'en',
 }) {
   logger.info('Executing RAG query', {
     service: 'rag-executor',
@@ -47,6 +48,7 @@ export async function executeRAG({
     onEvent,
     userId,
     authorizedWorkspaceIds,
+    lang,
   });
 
   logger.info('RAG query completed', {

backend/tests/unittest/complianceKbSourceUrl.test.js

@@ -0,0 +1,35 @@
+import { describe, it, expect, vi } from 'vitest';
+
+// Avoid pulling the real LangChain Qdrant store at import time.
+vi.mock('@langchain/qdrant', () => ({
+  QdrantVectorStore: { fromExistingCollection: vi.fn() },
+}));
+
+const { eurLexUrl } = await import('../../services/rag/complianceKbRetriever.js');
+
+/**
+ * #424 — regulation citations link to the OFFICIAL EUR-Lex text, in the user's
+ * language (the regulation is published verbatim in all EU languages).
+ */
+describe('compliance KB source URLs (#424)', () => {
+  it('links DORA to the English EUR-Lex text by default', () => {
+    expect(eurLexUrl('DORA', 'en')).toBe(
+      'https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554'
+    );
+  });
+
+  it('links DORA to the French EUR-Lex text for fr / fr-FR', () => {
+    expect(eurLexUrl('DORA', 'fr')).toContain('/FR/TXT/');
+    expect(eurLexUrl('DORA', 'fr-FR')).toContain('/FR/TXT/');
+  });
+
+  it('defaults to EN for an unknown or missing locale', () => {
+    expect(eurLexUrl('DORA', undefined)).toContain('/EN/TXT/');
+    expect(eurLexUrl('DORA', 'de')).toContain('/EN/TXT/');
+  });
+
+  it('returns no URL where we have no reliable CELEX (DORA-RTS, unknown)', () => {
+    expect(eurLexUrl('DORA-RTS', 'fr')).toBeUndefined();
+    expect(eurLexUrl('Unknown', 'en')).toBeUndefined();
+  });
+});

backend/tests/unittest/ragService.test.js

@@ -505,7 +505,8 @@ describe('askWithConversation', () => {
 
     expect(vendorSearch).toHaveBeenCalledTimes(1);
     expect(retrieveRegulationDocs).toHaveBeenCalledTimes(1);
-    expect(retrieveRegulationDocs).toHaveBeenCalledWith(expect.any(String), 5);
+    // 3rd arg = the UI locale, threaded through for language-aware source links (#424).
+    expect(retrieveRegulationDocs).toHaveBeenCalledWith(expect.any(String), 5, undefined);
 
     // Reranker should see merged set (vendor first, regulation second).
     const mergedArg = rerankDocuments.mock.calls[0][0];

backend/validators/schemas.js

@@ -34,6 +34,8 @@ export const askQuestionSchema = z
     // workspaceId may arrive in the body (requireWorkspaceAccess also reads it from
     // the X-Workspace-Id header); declare it so .strict() doesn't reject it.
     workspaceId: z.string().max(64).optional(),
+    // UI locale (e.g. 'fr', 'en-US') — drives language-aware regulation source links.
+    lang: z.string().max(10).optional(),
     filters: z
       .object({
         page: z.number().int().positive().optional(),
@@ -58,6 +60,8 @@ export const streamQuestionSchema = z
     conversationId: z.string().regex(/^[0-9a-fA-F]{24}$/, 'Invalid conversation ID'),
     // See askQuestionSchema: workspaceId may be sent in the body.
     workspaceId: z.string().max(64).optional(),
+    // UI locale (e.g. 'fr', 'en-US') — drives language-aware regulation source links.
+    lang: z.string().max(10).optional(),
     filters: z
       .object({
         page: z.number().int().positive().optional(),

frontend/src/features/chat/hooks/use-streaming.ts

@@ -3,6 +3,7 @@
 import { useState, useCallback, useRef } from 'react';
 import type { Source } from '@/types';
 import { getActiveWorkspaceContextId } from '@/shared/lib/workspace-context';
+import i18n from '@/shared/i18n/config';
 
 /**
  * ISSUE #41 FIX: Streaming timeout configuration
@@ -111,7 +112,7 @@ export function useStreaming(options: UseStreamingOptions = {}) {
             ...(resolvedWorkspaceId ? { 'X-Workspace-Id': resolvedWorkspaceId } : {}),
           },
           credentials: 'include',
-          body: JSON.stringify({ question, conversationId }),
+          body: JSON.stringify({ question, conversationId, lang: i18n.language }),
           signal: controller.signal,
         });
 

frontend/src/tests/use-streaming.test.ts

@@ -169,15 +169,12 @@ describe('useStreaming', () => {
         result.current.startStreaming('test question', 'conv-123');
       });
 
-      expect(mockFetch).toHaveBeenCalledWith(
-        expect.any(String),
-        expect.objectContaining({
-          body: JSON.stringify({
-            question: 'test question',
-            conversationId: 'conv-123',
-          }),
-        })
-      );
+      // Parse the body so the assertion ignores the added `lang` field (#424).
+      const body = JSON.parse((mockFetch.mock.calls[0][1] as { body: string }).body);
+      expect(body).toMatchObject({
+        question: 'test question',
+        conversationId: 'conv-123',
+      });
     });
 
     it('should use POST method', () => {