1// LLM context windows (approximate):
2// GPT-4o-mini: 128K tokens
3// GPT-4o: 128K tokens
4// Claude 3.5 Sonnet: 200K tokens
5
6// Rule of thumb: keep context under 50% of window
7// Reserve rest for system prompt + response
8
9function estimateTokens(text) {
10  // Rough: 1 token per 4 chars for English
11  // 1 token per 2-3 chars for Vietnamese
12  return Math.ceil(text.length / 3);
13}
14
15function trimContext(chunks, maxTokens = 3000) {
16  let totalTokens = 0;
17  const selected = [];
18  
19  for (const chunk of chunks) {
20    const tokens = estimateTokens(chunk.content);
21    if (totalTokens + tokens > maxTokens) break;
22    selected.push(chunk);
23    totalTokens += tokens;
24  }
25  
26  return { chunks: selected, totalTokens };
27}

1// n8n Window Buffer Memory node
2// Configuration:
3// - Session Key: {{ $json.sessionId }}
4// - Context Window Length: 10 (messages)
5// - Memory Type: Window Buffer
6
7// This automatically:
8// 1. Stores conversation messages
9// 2. Retrieves last N messages
10// 3. Includes in AI context

1// For persistent memory across sessions
2// n8n Redis Chat Memory node
3
4// Configuration:
5// - Session Key: {{ $json.userId }}_{{ $json.conversationId }}
6// - Session TTL: 3600 (1 hour)
7// - Context Window Length: 20
8
9// Redis automatically manages:
10// - Session isolation per user
11// - TTL-based cleanup
12// - Resumable conversations

1// Problem: "What about the pricing?"
2// Without history, this is ambiguous
3// With history: user was asking about Product X
4
5// OpenAI node: Rewrite query with context
6const rewritePrompt = `
7Given the conversation history, rewrite the latest query 
8to be self-contained and specific.
9
10History:
11${$json.history.map(m => `${m.role}: ${m.content}`).join('\n')}
12
13Latest query: "${$json.query}"
14
15Rewritten query (self-contained):`;

1// Code node: Combine retrieval + history context
2const retrievedChunks = $json.chunks;
3const conversationHistory = $json.history.slice(-5);
4
5const systemPrompt = `You are a helpful Q&A assistant for our knowledge base.
6
7Conversation history for context:
8${conversationHistory.map(m => `${m.role}: ${m.content}`).join('\n')}
9
10Retrieved information:
11${retrievedChunks.map((c, i) => `[${i + 1}] ${c.content}`).join('\n\n')}
12
13Instructions:
14- Use retrieved information to answer
15- Consider conversation context
16- If follow-up question, connect to previous answers
17- Cite sources: [Source N]`;
18
19return { json: { systemPrompt } };

1// When context is too long, compress it
2const compressPrompt = `
3Compress the following text while preserving all key information 
4relevant to answering: "${$json.query}"
5
6Text to compress:
7${$json.longContext}
8
9Compressed version (keep essential facts, remove filler):`;
10
11// This can reduce context by 50-70%

1// Code node: Manage chat sessions
2const sessionStore = {
3  // Create new session
4  createSession(userId) {
5    return {
6      sessionId: `${userId}_${Date.now()}`,
7      userId,
8      startedAt: new Date().toISOString(),
9      messageCount: 0,
10      topics: []
11    };
12  },
13  
14  // Check if session should expire
15  shouldExpire(session, maxIdleMinutes = 30) {
16    const lastActivity = new Date(session.lastActivity);
17    const now = new Date();
18    const idle = (now - lastActivity) / 60000;
19    return idle > maxIdleMinutes;
20  }
21};