Handle customer support queries with cache-first RAG using Redis, LangCache and OpenAI

An end-to-end Retrieval-Augmented Generation (RAG) customer support workflow for n8n, using a cache-first strategy (LangCache) combined with a Redis vector store powered by OpenAI embeddings. This template is designed for fast, accurate, and cost-efficient customer support chatbots, internal help desks, and knowledge-base assistants. --- Overview This workflow implements a production-ready RAG architecture optimized for customer support use cases. Incoming chat messages are processed through a structured pipeline that prioritizes cached answers, falls back to semantic vector search when needed, and validates response quality before returning a final answer. The workflow supports: - Multi-question user inputs - Intelligent query decomposition - Cache reuse to reduce latency and cost - High-precision retrieval from a Redis vector database - Quality evaluation and controlled retries - Final answer synthesis into a single, coherent response --- Key Features - Chat-based RAG pipeline using n8n’s Chat Trigger - Query decomposition for multi-topic questions - LangCache integration (search + save) - Redis Vector Store for semantic retrieval - OpenAI embeddings and chat models - Quality scoring with retry logic - Session memory buffers for contextual continuity - Fallback-safe behavior (no hallucinations) --- How the Workflow Works 1. Chat Trigger The workflow starts when a new chat message is received. 2. Configuration Setup A centralized configuration node defines: - LangCache base URL - Cache ID - Similarity threshold (default: 0.75) - Maximum retrieval iterations (default: 2) 3. Query Decomposition The user message is analyzed and decomposed into: - A single focused question, or - Multiple independent sub-questions This improves retrieval accuracy and cache reuse. 4. Cache-First Retrieval Each sub-question is processed independently: - The workflow first searches LangCache - If a high-similarity cached answer is found, it is reused immediately 5. Vector Retrieval (Cache Miss) If no cache hit exists: - The query is embedded using OpenAI embeddings - A semantic search is executed against the Redis vector index - Retrieved knowledge-base documents are passed to a research-only agent 6. Knowledge-Only Answering The research agent: - Answers strictly from the retrieved knowledge - Returns "no info found" if no relevant data exists 7. Quality Evaluation Each generated answer is evaluated by a dedicated quality-check node: - Outputs a numerical SCORE (0.0 – 1.0) - Provides textual feedback - Low scores can trigger limited retries 8. Cache Update High-quality answers are saved back to LangCache for future reuse. 9. Aggregation & Synthesis All sub-answers are aggregated and synthesized into: - One final, user-facing response, or - A polite fallback message if information is insufficient --- Main Nodes & Responsibilities - When Chat Message Received — Entry point for user messages - LangCache Config — Centralized configuration values - Decompose Query (LangChain Agent) — Splits complex queries - Structured Output Parser — Ensures valid JSON output - Search LangCache — Cache lookup via HTTP - Redis Vector Store — Semantic retrieval from Redis - Embeddings OpenAI — Vector generation - Research Agent — KB-only answering (no hallucinations) - Quality Evaluator — Scores answer relevance - Save to LangCache — Stores validated answers - Memory Buffers — Session context handling - Response Synthesizer — Final message generation --- Setup Instructions 1. Configure Credentials Create the following credentials in n8n: - OpenAI API - Redis - HTTP Bearer Auth (for LangCache) 2. Prepare the Knowledge Base - Embed your documents using OpenAI embeddings - Insert them into the configured Redis vector index - Ensure documents are concise and well-structured 3. Configure LangCache Update the configuration node with: - langcacheBaseUrl - langcacheCacheId - Optional tuning for similarity threshold and iterations 4. Test the Workflow - Use the example data loader