RAG Pipelines

Retrieval-Augmented Generation grounds language model responses in your actual data instead of relying on training knowledge alone.

Pipeline stages

1. Ingest — load documents from sources
2. Chunk — split into semantic segments
3. Embed — convert chunks to vector representations
4. Index — store embeddings in a vector database
5. Retrieve — find relevant chunks for a query
6. Generate — pass retrieved context + query to the language model

Chunking strategies

Strategy	When to use
Fixed size (500 tokens)	Simple, predictable, fast
Semantic (by heading/section)	Structured documents
Recursive splitting	General purpose with overlap
Sentence-level	When precision matters

Retrieval methods

• Semantic search — cosine similarity on embeddings
• Keyword search — BM25 or full-text search
• Hybrid — combine semantic + keyword with reciprocal rank fusion
• Re-ranking — second pass with a cross-encoder for precision