rag-pipeline-builder

Trigger

Use this skill when building a RAG system from scratch, creating document chatbots, implementing semantic search, or setting up knowledge base retrieval for AI applications.

Trigger phrases:

• "Build a RAG pipeline"
• "Document chatbot"
• "Semantic search over my docs"
• "Ingest and search"
• "Knowledge base retrieval"

Overview

Every production RAG pipeline follows the same flow: ingest documents → chunk intelligently → embed with quality models → store in vector database → retrieve relevant context → generate responses. This skill generates the complete pipeline with battle-tested patterns from real deployments (AA Companion, VAI workflows, Developer Day demos).

This is not about adding AI to an existing app (that's mongodb-ai-features). This is about building a RAG system from the ground up.

How to Use

Quick Start

1. Analyze your documents: python scripts/chunking_strategy_analyzer.py docs/
2. Estimate costs: python scripts/rag_cost_estimator.py --docs 1000 --queries-per-month 5000
3. Generate pipeline from templates in assets/
4. Validate: python scripts/pipeline_validator.py config.json

Python Tools

• scripts/chunking_strategy_analyzer.py — Analyze documents and recommend optimal chunking strategy
• scripts/rag_cost_estimator.py — Estimate embedding, storage, and query costs
• scripts/pipeline_validator.py — Validate pipeline config and test retrieval quality

Reference Docs

• references/chunking-strategies.md — Deep dive on chunking approaches
• references/rag-patterns.md — Production patterns from real deployments

Templates & Assets

• assets/ingest-pipeline.js — Complete ingestion with content hashing
• assets/vector-index-config.json — Atlas Vector Search index definition
• assets/retrieval-api.js — Retrieval API with score boosting
• assets/chat-endpoint.js — Streaming chat endpoint
• assets/sample-queries.json — Test queries for validation

Architecture Decisions

Why Voyage AI for Embeddings

• Quality: Voyage-3 outperforms OpenAI on retrieval benchmarks
• Cost: ~$0.02 per 1M tokens (10x cheaper than text-embedding-3-large)
• Speed: Fast batch processing (128 documents per batch)
• Integration: Seamless with MongoDB Atlas Vector Search

Content Hashing for Incremental Ingestion

Reprocessing entire corpus on every update is wasteful. Content hashing (SHA-256) enables incremental updates:

• Hash each document before embedding
• Store hash in metadata
• On re-ingestion, skip documents with matching hashes
• Only process new/changed documents

Result: 90%+ cost reduction after initial ingestion.

Chunking Strategy Selection

Category-Based Score Boosting

Not all results are equal. Boost by document category:

const boostedScore = baseScore * (doc.category === 'ui' ? 1.5 : 1.0);

Fire-and-Forget Usage Logging

Don't block responses to log usage:

logUsage(query, results).catch(err => console.error('Log failed:', err));
return results; // Return immediately

Generated Pipeline Structure

my-rag-app/
├── ingest/
│   ├── ingest.js              # Main ingestion pipeline
│   ├── chunkers/
│   │   ├── recursive.js       # Recursive text splitting
│   │   ├── semantic.js        # Paragraph-based
│   │   └── code.js            # Function-boundary
│   └── utils/
│       ├── content-hash.js    # SHA-256 hashing
│       └── batch-embed.js     # Voyage AI batching
├── index/
│   ├── create-index.js        # Atlas Vector Search setup
│   └── index-config.json      # Index definition
├── api/
│   ├── search.js              # Retrieval API
│   ├── chat.js                # Chat endpoint
│   └── middleware/
│       └── usage-logger.js    # Fire-and-forget logging
├── lib/
│   ├── voyage-client.js       # Voyage AI wrapper
│   ├── atlas-client.js        # MongoDB connection
│   └── score-booster.js       # Category boosting
├── tests/
│   ├── retrieval-quality.test.js
│   └── fixtures/
│       └── sample-queries.json
├── .env.example
└── README.md

When to Use vs. mongodb-ai-features

Environment Variables

# MongoDB Atlas
MONGODB_URI=mongodb+srv://user:pass@cluster.mongodb.net/rag-db

# Voyage AI
VOYAGE_API_KEY=pa-xxx...

# Optional: LLM for generation
OPENAI_API_KEY=sk-xxx...

# Configuration
VECTOR_INDEX_NAME=vector_index
COLLECTION_NAME=documents
EMBEDDING_DIMENSIONS=1024
CHUNK_SIZE=1000
CHUNK_OVERLAP=200

Production Checklist

• Chunking strategy validated with analyzer
• Cost estimation completed
• Vector index created (correct dimensions)
• Retrieval quality tested (MRR > 0.7)
• Incremental ingestion implemented
• Score boosting configured
• Usage logging implemented
• Error handling on embed failures
• Rate limiting configured
• Monitoring for latency/errors

References

• Voyage AI: https://voyage.ai/pricing
• Atlas Vector Search: https://mongodb.com/docs/atlas/atlas-vector-search/
• Chunking strategies: references/chunking-strategies.md
• RAG patterns: references/rag-patterns.md

Credits

Michael Lynn — mlynn.org · @mlynn · LinkedIn · GitHub