diff --git a/app/api/chat/message/route.ts b/app/api/chat/message/route.ts
index b9f2909..2e31593 100644
--- a/app/api/chat/message/route.ts
+++ b/app/api/chat/message/route.ts
@@ -1,4 +1,5 @@
 import { createServerClientSSR } from '@/lib/supabase/server'
+import { getRAGAgent } from '@/lib/rag/agent'
 import { requireAuth } from '@/lib/utils/auth'
 import { ChatMessageSchema } from '@/lib/types'
 import { successResponse, errorResponse, handleError } from '@/lib/utils/api-response'
@@ -24,56 +25,47 @@ export async function POST(request: NextRequest) {
       return errorResponse('Session not found', 404)
     }
 
-    // Save user message
-    const { data: userMessage, error: userMessageError } = await supabase
-      .from('chat_messages')
-      .insert({
-        session_id: validatedData.sessionId,
-        role: 'user' as const,
-        content: validatedData.content,
-      } as any)
-      .select()
-      .single()
-
-    if (userMessageError) {
-      throw userMessageError
-    }
-
-    // TODO: Implement RAG pipeline
-    // 1. Retrieve relevant chunks from paper_chunks using vector similarity
-    // 2. Build context with citations
-    // 3. Generate response using GPT-4o-mini
-    // 4. Save assistant message with citations
+    const ragAgent = getRAGAgent()
+    const result = await ragAgent.query(
+      validatedData.content,
+      user.id,
+      validatedData.sessionId,
+      session.paper_id || undefined
+    )
 
-    // For now, return a placeholder response
-    const placeholderResponse = "I'm ready to help you understand this paper! However, the RAG (Retrieval Augmented Generation) system is not yet implemented. Once it's ready, I'll be able to:\n\n1. Search through the paper content\n2. Find relevant sections to answer your question\n3. Provide accurate responses with citations\n\nPlease check back soon!"
-
-    // Save assistant message (placeholder)
-    const { data: assistantMessage, error: assistantMessageError } = await supabase
+    const { data: latestMessages, error: latestMessagesError } = await supabase
       .from('chat_messages')
-      .insert({
-        session_id: validatedData.sessionId,
-        role: 'assistant' as const,
-        content: placeholderResponse,
-      } as any)
-      .select()
-      .single()
+      .select('*')
+      .eq('session_id', validatedData.sessionId)
+      .in('role', ['user', 'assistant'])
+      .order('created_at', { ascending: false })
+      .limit(4)
 
-    if (assistantMessageError) {
-      throw assistantMessageError
+    if (latestMessagesError) {
+      throw latestMessagesError
     }
 
-    // Update session's updated_at timestamp
-    await (supabase
-      .from('chat_sessions') as any)
-      .update({ updated_at: new Date().toISOString() })
-      .eq('id', validatedData.sessionId)
+    const userMessage =
+      latestMessages?.find((message: any) => (
+        message.role === 'user' && message.content === validatedData.content
+      )) ||
+      latestMessages?.find((message: any) => message.role === 'user') ||
+      null
+
+    const assistantMessage =
+      latestMessages?.find((message: any) => message.role === 'assistant') ||
+      null
 
     return successResponse({
       userMessage,
       assistantMessage,
+      answer: result.answer,
+      sources: result.sources,
+      citations: result.citations,
+      queryTime: result.queryTime,
+      sessionId: result.sessionId,
     })
   } catch (error) {
     return handleError(error)
   }
-}
\ No newline at end of file
+}
diff --git a/app/api/rag/query/route.ts b/app/api/rag/query/route.ts
index 24c099d..6810da2 100644
--- a/app/api/rag/query/route.ts
+++ b/app/api/rag/query/route.ts
@@ -1,20 +1,9 @@
 import { NextRequest, NextResponse } from 'next/server';
-import { RAGAgent } from '@/lib/rag/rag_agent';
+import { getRAGAgent } from '@/lib/rag/agent';
 import { ConfigManager, loadConfigFromEnv } from '@/lib/rag/config';
 
-const configManager = new ConfigManager(loadConfigFromEnv());
-const config = configManager.get();
-
-const ragAgent = new RAGAgent({
-  openaiApiKey: config.openai.apiKey,
-  supabaseUrl: process.env.NEXT_PUBLIC_SUPABASE_URL!,
-  supabaseKey: process.env.SUPABASE_SERVICE_ROLE_KEY!,
-  chromaUrl: config.vectorStore.chromaUrl,
-  vectorStoreType: config.vectorStore.type,
-  modelName: config.openai.model,
-  chunkSize: config.chunking.chunkSize,
-  chunkOverlap: config.chunking.chunkOverlap,
-});
+const config = new ConfigManager(loadConfigFromEnv()).get();
+const ragAgent = getRAGAgent();
 
 export async function POST(request: NextRequest) {
   try {
@@ -106,4 +95,4 @@ export async function POST(request: NextRequest) {
       { status: 500 }
     );
   }
-}
\ No newline at end of file
+}
diff --git a/app/api/rag/session/route.ts b/app/api/rag/session/route.ts
index 0312fc2..0845e92 100644
--- a/app/api/rag/session/route.ts
+++ b/app/api/rag/session/route.ts
@@ -1,20 +1,7 @@
 import { NextRequest, NextResponse } from 'next/server';
-import { RAGAgent } from '@/lib/rag/rag_agent';
-import { ConfigManager, loadConfigFromEnv } from '@/lib/rag/config';
+import { getRAGAgent } from '@/lib/rag/agent';
 
-const configManager = new ConfigManager(loadConfigFromEnv());
-const config = configManager.get();
-
-const ragAgent = new RAGAgent({
-  openaiApiKey: config.openai.apiKey,
-  supabaseUrl: process.env.NEXT_PUBLIC_SUPABASE_URL!,
-  supabaseKey: process.env.SUPABASE_SERVICE_ROLE_KEY!,
-  chromaUrl: config.vectorStore.chromaUrl,
-  vectorStoreType: config.vectorStore.type,
-  modelName: config.openai.model,
-  chunkSize: config.chunking.chunkSize,
-  chunkOverlap: config.chunking.chunkOverlap,
-});
+const ragAgent = getRAGAgent();
 
 export async function POST(request: NextRequest) {
   try {
@@ -78,4 +65,4 @@ export async function GET(request: NextRequest) {
       { status: 500 }
     );
   }
-}
\ No newline at end of file
+}
diff --git a/lib/rag/agent.ts b/lib/rag/agent.ts
new file mode 100644
index 0000000..e58539b
--- /dev/null
+++ b/lib/rag/agent.ts
@@ -0,0 +1,26 @@
+import { ConfigManager, loadConfigFromEnv } from './config'
+import { RAGAgent } from './rag_agent'
+
+let ragAgent: RAGAgent | null = null
+
+export function getRAGAgent(): RAGAgent {
+  if (ragAgent) {
+    return ragAgent
+  }
+
+  const configManager = new ConfigManager(loadConfigFromEnv())
+  const config = configManager.get()
+
+  ragAgent = new RAGAgent({
+    openaiApiKey: config.openai.apiKey,
+    supabaseUrl: process.env.NEXT_PUBLIC_SUPABASE_URL!,
+    supabaseKey: process.env.SUPABASE_SERVICE_ROLE_KEY!,
+    chromaUrl: config.vectorStore.chromaUrl,
+    vectorStoreType: config.vectorStore.type,
+    modelName: config.openai.model,
+    chunkSize: config.chunking.chunkSize,
+    chunkOverlap: config.chunking.chunkOverlap,
+  })
+
+  return ragAgent
+}