acaprino/rag-development

RAG Development

Comprehensive knowledge base for building production-grade Retrieval-Augmented Generation systems.

When to Use

• Building a new RAG pipeline from scratch
• Choosing chunking strategy, embedding model, or vector database
• Implementing hybrid search, re-ranking, or contextual retrieval
• Evaluating RAG quality with RAGAS or DeepEval
• Optimizing production RAG for cost, latency, or accuracy
• Designing multi-tenant RAG with access control
• Upgrading from naive RAG to advanced patterns

Quick Start Recommendation

For 80% of use cases, start with:

1. Chunking: Recursive character splitting at 512 tokens, 10-15% overlap
2. Embedding: OpenAI text-embedding-3-small (best value) or Cohere embed-v4 (best accuracy)
3. Vector DB: Qdrant with scalar INT8 quantization
4. Retrieval: Hybrid search (dense + sparse + RRF)
5. Evaluation: RAGAS metrics from day one

Then upgrade incrementally based on measured failures:

• Keyword misses -> add sparse vectors (SPLADE/BM25)
• Ambiguous chunks -> add contextual retrieval (Anthropic pattern)
• Irrelevant results -> add cross-encoder re-ranking
• Multi-hop failures -> upgrade to agentic RAG

Reference Materials

Detailed reference documents are in the references/ directory:

• chunking-strategies.md -- all chunking approaches with code, benchmarks, and selection guide
• embedding-models.md -- model comparison, Matryoshka embeddings, fine-tuning, sparse/dense/multi-vector
• retrieval-patterns.md -- hybrid search, HyDE, contextual retrieval, re-ranking, MMR
• advanced-rag-patterns.md -- Graph RAG, RAPTOR, CRAG, Self-RAG, Agentic RAG, multi-modal RAG
• vector-databases.md -- Qdrant deep dive, database comparison, scaling strategies
• production-guide.md -- evaluation, observability, caching, security, cost optimization

Pipeline Architecture

Document Ingestion:
  Raw Docs -> Preprocessing (Unstructured.io) -> Chunking -> Context Enrichment -> Embedding -> Vector DB

Query Pipeline:
  User Query -> Query Transform -> Encode (Dense + Sparse) -> Hybrid Search -> Re-rank -> LLM Generation

Evaluation Loop:
  Ground Truth + Predictions -> RAGAS/DeepEval -> Faithfulness, Relevancy, Precision, Recall

Key Decision Points

| Decision | Default | Upgrade When | |----------|---------|-------------| | Chunking | Recursive 512 tok | Structured docs -> markdown-aware; cross-refs -> late chunking | | Embedding | text-embedding-3-small | Need accuracy -> embed-v4; self-hosted -> NV-Embed-v2 | | Vector DB | Qdrant + INT8 | Already on Postgres -> pgvector; need managed -> Pinecone | | Search | Dense only | Keyword misses -> add sparse hybrid; poor diversity -> add MMR | | Re-ranking | None | Top-k results contain irrelevant items -> add Cohere Rerank | | Caching | None | Production latency/cost concerns -> semantic cache | | Evaluation | Manual spot checks | Any production use -> RAGAS automated metrics |