How to Build an Enterprise‑Grade Vector Search Q&A Bot with Milvus and n8n

Background and Challenges

When building a domain‑specific Q&A bot for a framework, platform, or library, several core challenges arise:

Knowledge lag : Technical documentation updates frequently; relying on a generic LLM or manually updated knowledge bases quickly becomes outdated.

Hallucinations and factual errors : General LLMs may fabricate answers, leading developers to waste time debugging.

Domain terminology : Specific terms have different meanings across fields, and generic LLMs often miss these nuances.

High maintenance cost : Manually tracking documentation changes and updating the knowledge base is error‑prone and hard to scale.

This guide focuses on Alibaba Cloud Milvus and the low‑code workflow orchestration platform n8n to address these issues.

Milvus Basics and Architecture

Milvus is a distributed database designed for vector similarity search. Its core technologies include:

Approximate Nearest Neighbor (ANN) search using HNSW, IVF, PQ, etc., balancing accuracy and speed.

Separate vector index and query layers, supporting dynamic construction of indexes such as FLAT, IVF_FLAT, IVF_PQ, HNSW.

Vector data sharding and distributed computation for high throughput and low latency.

Milvus adopts a cloud‑native, compute‑storage‑separated micro‑service architecture with four layers: access, coordination, execution, and storage. Components can be scaled independently, ensuring high performance, availability, and elasticity.

Milvus is suitable for any scenario requiring similarity matching, such as image/video search, semantic text search, personalized recommendation, scientific research, and autonomous driving data mining.

n8n Platform Overview

n8n (NodeNation) is an open‑source, visual workflow automation platform often described as “the programmer’s Zapier.” It lets users connect nodes representing actions (e.g., database read, API call, email send) on a canvas, enabling complex data pipelines without writing extensive code. Core concepts include:

Node : The smallest functional unit that performs a specific operation.

Workflow : A directed graph of nodes supporting branching, loops, and error handling.

Engine : Executes the workflow, manages task queues, and handles retries.

Volume : Persistent storage for n8n configuration and logs, typically mounted as /home/n8n/.n8n.

Solution Overview

Two n8n workflows are created:

A scheduled workflow pulls Milvus product documentation from GitHub, vectorizes the content, and stores it in a Milvus collection.

An AI Agent workflow receives chat input, decides whether the query relates to Milvus, retrieves relevant chunks from Milvus, and combines them with a large language model (LLM) response.

Milvus Access Configuration

Create a Milvus account and fill in the Base URL (including port 19530), Username, and Password in n8n’s Milvus credential.

GitHub Access Configuration

Create a GitHub account, generate an OpenAPI token, and configure the GitHub credential with the repository owner, name, and the path of markdown files to ingest.

Workflow Configuration – Building the Knowledge Base

Add a GitHub List Files node to list documentation files, then a Split Out node to extract the download_url. Use an If node to filter out null or non‑markdown links.

Configure an HTTP Request node to fetch the raw markdown content, then a Milvus Vector Store node to embed the text (using text-embedding-v1) and write vectors to a collection (e.g., n8n_test) with the “Clear Collection” option enabled.

n8n Installation and Verification

docker volume create n8n_data</code>
<code>docker run --name n8n -d \
    -e N8N_SECURE_COOKIE=false \
    -p 5678:5678 \
    -v n8n_data:/home/node/.n8n docker.n8n.io/n8nio/n8n

After starting, access http://127.0.0.1:5678/ to set the admin credentials and open the dashboard.

Workflow Configuration – LLM Retrieval Augmentation

Create a second workflow triggered by the When chat message received node. Add an AI Agent node with the following system prompt:

你是主控 AI，负责统筹不同的专业子代理，以执行复杂的用户任务。
你有以下的Agent/Tool
1. MilvusDocumentAgent  Agent负责搜索所有Milvus相关的文档，当涉及到Milvus相关搜索，运行这个Agent
其他情况下，直接在该节点执行

Select the qwen-plus model as the chat model, use Simple Memory (max 5 turns), and add the Milvus Vector Store as a Tool (collection n8n_test, top‑4 results, no rerank).

Finally, add a Set node to extract the text output for the chat response.

Verification

Run the chat workflow and ask a Milvus‑related question such as “What are the data classification strategies in Milvus?” The AI Agent retrieves relevant chunks from Milvus, lets the LLM refine them, and returns a readable answer.

Execution logs show each node’s input and output, confirming successful vector retrieval and LLM integration.

Observability and OpenTelemetry Integration

Complex workflows benefit from tracing. n8n does not provide native tracing, but it can be extended with OpenTelemetry. By adding OpenTelemetry environment variables to the Docker run command, traces can be exported to platforms such as Alibaba Cloud SLS, Jaeger, or SkyWalking.

docker run --name n8n -d \
    -e N8N_SECURE_COOKIE=false \
    -e OTEL_SDK_DISABLED="false" \
    -e OTEL_SERVICE_NAME="n8n" \
    -e OTEL_LOG_LEVEL="info" \
    -e OTEL_EXPORTER_OTLP_ENDPOINT="https://taihao-server-cn-hongkong-prod.cn-hongkong-intranet.log.aliyuncs.com:10010" \
    -e OTEL_EXPORTER_OTLP_HEADERS="x-sls-otel-project=taihao-server-cn-hongkong-prod,x-sls-otel-instance-id=taihao-cn-hongkong,x-sls-otel-ak-id=xxx,x-sls-otel-ak-secret=xxx" \
    -p 5678:5678 \
    -v n8n_data:/home/node/.n8n n8n-otel:latest

After restarting, the workflow logs include trace data that can be visualized in the chosen observability platform.