n8n RAG Agent with Vector Search for Knowledge Base Q&A

This workflow builds a Retrieval-Augmented Generation (RAG) agent in n8n 1.0+ that connects to Pinecone or Qdrant vector databases, retrieves relevant documents from your knowledge base, and answers user questions with grounded responses. The agentic reasoning step uses OpenAI GPT-4o or Claude Sonnet 4.6 to evaluate retrieved chunks against the query — it scores each chunk for relevance, discards low-scoring matches, and synthesizes an answer only from passages that pass the relevance threshold. This is not a simple keyword search or FAQ bot — the agent dynamically decides which chunks to use based on semantic fit, filters out contradictory information, and cites the source document for each claim. n8n provides built-in vector store nodes for Pinecone, Qdrant, Supabase Vector (pgvector), Chroma, and in-memory storage, with a visual workflow builder that requires no custom application code. Teams deploying n8n RAG agents report cutting support ticket resolution time from 15-20 minutes to 3-5 minutes for knowledge-base-answerable queries.

BUSINESS PROBLEM

A customer support team at a B2B SaaS company with 4,000+ knowledge base articles receives 250+ tickets per week that could be answered from existing documentation. Support agents spend 6-9 minutes per ticket searching for the right article, reading it, and composing a response. The current search tool finds articles by keyword match only — agents waste time clicking through irrelevant results. The company tried a basic chatbot built on intent classification, but it hallucinated answers when it could not find a match, eroding customer trust. According to a 2025 Gartner survey, 67% of customers prefer self-service over speaking to an agent for known-issue resolution, and 80% expect a company to know their history across interactions. A simple FAQ bot fails here — customers ask complex multi-part questions that require synthesizing information from 2-3 different articles. The annual cost of manual knowledge-base lookup: $49,000-74,000 per support team of 5 agents at $35/hr fully loaded.

WHO BENEFITS

Customer support teams at SaaS companies (50-500 employees) handling 200+ tickets per week who need to reduce time-to-resolution for documentation-answerable queries — the RAG agent cuts lookup and response time from 6-9 minutes to under 2 minutes. Internal IT helpdesks managing employee support for policies, benefits, and software access who field the same 30-40 questions repeatedly — the agent answers from internal wikis and runs 24/7 without escalation. Product documentation teams who want to embed a "search your docs" widget into their SaaS product — n8n exposes the RAG agent as a webhook API that the product frontend calls directly with zero additional infrastructure.

HOW IT WORKS

1. Document ingestion: You connect n8n to your knowledge base source (Google Drive, Notion, Confluence, or a local folder). The workflow loads documents using n8n's document loader nodes. Output: raw document text.

2. Text chunking: A Recursive Character Text Splitter node breaks documents into chunks of 1,000-1,500 characters with 200-300 character overlap. Chunk size determines retrieval granularity. Output: array of text chunks with source metadata.

3. Embedding generation: Each chunk passes through an Embeddings OpenAI node configured with text-embedding-3-small (1,536 dimensions). This converts text into vector representations. Output: embedding vectors paired with original text.

4. Vector indexing: The chunk-embedding pairs are inserted into a Pinecone or Qdrant collection. Qdrant can be self-hosted or cloud — n8n provides a dedicated Vecto Store node for each. Output: searchable vector index.

5. Query processing: A user's question arrives via webhook or chat widget. The workflow generates an embedding for the query using the same embedding model. This is critical: mismatched embedding models produce garbage results. Output: query vector.

6. AI reasoning checkpoint: The AI Agent node receives the query and the top 4-6 retrieved chunks. It scores each chunk for relevance to the specific question, discards chunks scoring below 0.7 similarity, and checks retrieved chunks against each other for contradictions. Output: filtered relevant context JSON.

7. Response generation: The LLM (GPT-4o or Claude Sonnet 4.6) synthesizes an answer using only the filtered context. The response includes inline citations to the source document. Output: natural language answer with source references.

8. Human review (optional): High-stakes answers (refund policy, compliance questions) are routed to a human approval step before the response is sent. The agent flags queries containing keywords like "refund", "cancel", or "legal" for manual review.

TOOL INTEGRATION

n8n 1.0+: The orchestration layer. All AI nodes (Vector Store, AI Agent, Embeddings) are built into n8n 1.0+ and work in both cloud and self-hosted instances. Gotcha: The self-hosted Docker image requires additional configuration to enable AI nodes — set N8N_AI_ENABLED=true and ensure your n8n version is 1.0+, not the 0.x LTS release which lacks these nodes entirely.

Pinecone or Qdrant: The vector database. Pinecone is fully managed (starts at $70/month for the standard tier). Qdrant Cloud has a free tier with 1GB storage. Gotcha: When creating a Qdrant collection for n8n, use the simple "vectors" key, not "named_vectors" — n8n expects an unnamed vector structure and returns a 400 error if the collection uses named vectors. This is the most common Qdrant + n8n integration failure.

OpenAI text-embedding-3-small: The embedding model. Fixed at 1,536 dimensions. If you switch to text-embedding-3-large (3,072 dimensions), you must re-index all documents — the two models produce incompatible vector dimensions. Gotcha: The official OpenAI docs show embedding cost as per-token, but RAG ingestion at scale (10,000+ documents) can cost $20-50 in one-time embedding generation — budget this as a setup cost, not a runtime cost.

Window Buffer Memory (n8n): Provides conversation memory. Set the window size to 10 turns to maintain context across multi-step conversations. Gotcha: Memory stores raw conversation text, so if your vector store indexes customer-specific data, access control is not enforced by the memory node — sensitive data may be included in LLM context across sessions if you do not clear memory per session.

ROI METRICS

1. Ticket resolution time (knowledge-base types): 15-20 minutes to 3-5 minutes. Source: n8n RAG case studies, 2025-2026.
2. First-contact resolution rate: 55-65% with keyword search to 80-85% with RAG-based answers. Measurable from week 1 via ticket tags.
3. Support agent capacity: One agent handling 35-40 tickets/day to 60-70 tickets/day with RAG-assisted responses.
4. Hallucination rate: 15-25% with prompt-only chatbot to under 3% with retrieval-grounded generation. Source: Internal QA audit of chatbot responses.
5. Monthly infrastructure cost: $200-400/month (n8n cloud + vector DB + embedding API) versus $3,500-6,000/month in equivalent support agent time.

CAVEATS

1. Embedding model mismatch: If you change the embedding model, all previously indexed vectors become unusable. The entire collection must be re-indexed, which can take hours for large knowledge bases.
2. Chunk boundary issues: Answers that span two document chunks (e.g., a policy described across two PDF pages) may lose context. The chunk overlap of 200-300 characters helps but does not eliminate this risk entirely.
3. Vector database costs at scale: Pinecone's free tier is insufficient for production workloads above 50K vectors. Qdrant Cloud's $25/month tier is more cost-effective for small teams, but requires self-hosting for full data control.
4. Access control gaps: The RAG agent retrieves chunks by semantic similarity, not by user permissions. It cannot restrict access to documents based on the end user's role without a separate authorization step before the query.