1// Recursive Character Text Splitter
2// Configuration in n8n:
3// - Chunk Size: 1000 (characters)
4// - Chunk Overlap: 200 (characters)
5// - Separators: ["\n\n", "\n", " ", ""]
6
7// Best practices:
8// Chunk size 500-1500 chars for most docs
9// Overlap 10-20% of chunk size
10// Larger chunks = more context, less precision
11// Smaller chunks = more precision, less context

1// Code node: Add metadata before indexing
2const chunks = $input.all();
3
4return chunks.map((chunk, index) => ({
5  json: {
6    content: chunk.json.text,
7    metadata: {
8      source: chunk.json.source || "unknown",
9      page: chunk.json.page || 0,
10      chunkIndex: index,
11      documentTitle: chunk.json.title || "",
12      indexedAt: new Date().toISOString()
13    }
14  }
15}));

1// Vector Store Search node
2// Configuration:
3// - Top K: 5 (number of results)
4// - Score Threshold: 0.7 (minimum similarity)
5// - Filter: metadata-based filtering
6
7// Search with metadata filter
8// Filter by source document:
9// metadata.source == "company-handbook"
10
11// Filter by date:
12// metadata.indexedAt > "2025-01-01"

1// Code node: Build context from search results
2const results = $input.all();
3
4const context = results.map((r, i) => {
5  const sim = r.json.score ? r.json.score.toFixed(2) : "N/A";
6  return `[Source ${i + 1}] (similarity: ${sim})
7${r.json.document.pageContent}
8---`;
9}).join("\n\n");
10
11const prompt = `
12Based on the following context, answer the user's question.
13If the context doesn't contain the answer, say "I don't have enough information."
14
15Context:
16${context}
17
18Question: ${$json.query}
19
20Answer:`;
21
22return { json: { prompt, resultCount: results.length } };

1// AI Agent node configuration:
2// - Agent Type: Tools Agent
3// - LLM: OpenAI GPT-4o-mini
4// - Tools: Vector Store Retriever
5// - System Message:
6
7const systemMessage = `
8You are a helpful assistant that answers questions 
9based on the company knowledge base.
10
11Rules:
121. Only answer based on the retrieved context
132. If you are not sure, say so
143. Cite the source when possible
154. Be concise but complete
16`;

1// Workflow: Update documents when source changes
2// Google Drive Trigger (on file change) → Re-index
3
4// Code node: Upsert logic
5const documentId = $json.fileId;
6
7// Step 1: Delete old chunks for this document
8// Vector Store: Delete where metadata.sourceId == documentId
9
10// Step 2: Re-process and insert new chunks
11// Load → Split → Embed → Insert with metadata.sourceId = documentId

1// Code node: Track indexing metrics
2const metrics = {
3  timestamp: new Date().toISOString(),
4  operation: $json.operation, // insert, search, delete
5  documentCount: $json.count,
6  processingTime: $json.duration,
7  vectorCount: $json.totalVectors
8};
9
10// Save to Google Sheets for tracking
11return { json: metrics };