RAG Strategy

Retrieval-Augmented Generation strategy for memory search and context injection.

Plan for implementing Retrieval-Augmented Generation using Qdrant vector database.

Overview

RAG will enable QUOX to:

Reference 7000+ n8n workflow templates
Remember conversation history
Access infrastructure documentation
Learn from code patterns and solutions

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Query Flow                              │
│                                                              │
│  User Query ─────────────────────────────────────────────┐   │
│       │                                                  │   │
│       ▼                                                  │   │
│  ┌─────────────┐     ┌─────────────┐     ┌───────────┐  │   │
│  │   Embed     │────▶│   Qdrant    │────▶│  Results  │  │   │
│  │   Query     │     │   Search    │     │  (Top K)  │  │   │
│  └─────────────┘     └─────────────┘     └───────────┘  │   │
│                                                  │       │   │
│                                                  ▼       │   │
│                                          ┌───────────┐  │   │
│                                          │  Claude   │◀─┘   │
│                                          │  Context  │      │
│                                          └───────────┘      │
└─────────────────────────────────────────────────────────────┘

Collections

1. n8n_workflows

Purpose: Index 7000+ workflow templates for reference and inspiration.

json

{
  "collection": "n8n_workflows",
  "vector_size": 1536,
  "distance": "Cosine",
  "schema": {
    "name": "string",
    "description": "string",
    "nodes": "string[]",
    "triggers": "string[]",
    "category": "string",
    "source_url": "string"
  }
}

Data sources:

n8n official templates
Community workflows from inspiration repos
Custom QUOX workflows

2. infrastructure_docs

Purpose: Store host configurations, runbooks, and operational knowledge.

json

{
  "collection": "infrastructure_docs",
  "vector_size": 1536,
  "distance": "Cosine",
  "schema": {
    "hostname": "string",
    "ip_address": "string",
    "role": "string",
    "services": "string[]",
    "config": "string",
    "runbook": "string"
  }
}

Data sources:

Host inventory (50+ hosts)
Service configurations
Operational runbooks
Troubleshooting guides

3. conversations

Purpose: Long-term memory of user interactions and solutions.

json

{
  "collection": "conversations",
  "vector_size": 1536,
  "distance": "Cosine",
  "schema": {
    "session_id": "string",
    "timestamp": "datetime",
    "user_query": "string",
    "assistant_response": "string",
    "actions_taken": "string[]",
    "outcome": "string"
  }
}

Use cases:

"How did I fix this before?"
"What was the solution for prometheus01 last time?"
Learning from successful interactions

4. code_patterns

Purpose: Reusable code snippets and patterns.

json

{
  "collection": "code_patterns",
  "vector_size": 1536,
  "distance": "Cosine",
  "schema": {
    "name": "string",
    "language": "string",
    "pattern_type": "string",
    "code": "string",
    "use_case": "string",
    "dependencies": "string[]"
  }
}

Examples:

n8n Code node patterns
SSH command templates
API integration patterns
Error handling snippets

Deployment

Qdrant Container

Add to /opt/n8n/docker-compose.yml:

yaml

  qdrant:
    image: qdrant/qdrant:latest
    container_name: qdrant
    restart: unless-stopped
    ports:
      - "6333:6333"
      - "6334:6334"
    volumes:
      - /opt/qdrant/storage:/qdrant/storage
    environment:
      - QDRANT__SERVICE__GRPC_PORT=6334

n8n RAG Workflow

Create workflow for RAG queries:

Webhook (POST /webhook/rag-query)
    │
    ▼
Code Node (Generate Embedding)
    │
    ▼
HTTP Request (Qdrant Search)
    │
    ▼
Code Node (Format Results)
    │
    ▼
Respond to Webhook

Embedding Strategy

Model Selection

Recommended: OpenAI text-embedding-3-small

Cost: $0.00002/1K tokens
Dimensions: 1536
Good balance of quality and cost

Alternative: OpenAI text-embedding-3-large

Better quality, higher cost
Dimensions: 3072

Chunking Strategy

Content Type	Chunk Size	Overlap
Workflows	Full workflow	N/A
Documentation	500 tokens	50 tokens
Conversations	Full exchange	N/A
Code patterns	Full snippet	N/A

Query Patterns

Workflow Search

javascript

// User: "How do I send a Slack message when a GitHub issue is created?"
const query = "github issue slack notification workflow";
const results = await qdrant.search("n8n_workflows", embed(query), { limit: 5 });
// Returns relevant workflow templates

Infrastructure Lookup

javascript

// User: "What services run on prometheus01?"
const query = "prometheus01 services configuration";
const results = await qdrant.search("infrastructure_docs", embed(query), { limit: 3 });
// Returns host documentation

Conversation History

javascript

// User: "How did I fix the docker networking issue?"
const query = "docker networking fix solution";
const results = await qdrant.search("conversations", embed(query), { limit: 5 });
// Returns past interactions with solutions

Implementation Phases

Phase 1: Deploy Qdrant

Add Qdrant to docker-compose
Verify connectivity
Create collections

Phase 2: Seed Data

Import n8n workflow templates
Index infrastructure inventory
Add existing runbooks

Phase 3: Query Integration

Create RAG query workflow
Integrate with quox-chat
Add context injection

Phase 4: Continuous Learning

Store conversations automatically
Index successful solutions
Periodic re-embedding for updates

Estimated Resources

Storage

7000 workflows × 10KB = ~70MB vectors
50 hosts × 5KB = ~250KB vectors
Conversations: ~10MB/month
Total: ~100MB initially, growing

API Costs (OpenAI Embeddings)

Initial indexing: ~$5-10
Queries: ~$0.01/day typical use
Monthly: ~$1-5

Success Metrics

Query response time < 500ms
Relevance score > 0.7 for top results
Context injection improves Claude responses
Reduced repetitive questions

Next Steps

Add Qdrant to docker-compose
Create collections via API
Build embedding workflow in n8n
Import n8n workflow templates
Index infrastructure inventory
Integrate with quox-chat
Test and iterate on relevance