RAG from First Principles

RAG from First Principles An architecture diagram generated by Archify. Learner · asks a question · Architecture component · query time Learner asks a question query time CLI / Web Client · src/ask.ts · Architecture component CLI / Web Client src/ask.ts Source Documents · sample-data/*.txt · Architecture component · source knowledge Source Documents sample-data/*.txt source knowledge Ingestion · src/ingest.ts · Local learning environment Ingestion src/ingest.ts Chunking · src/chunk.ts · Local learning environment Chunking src/chunk.ts Ollama Embedder · nomic-embed-text · Local learning environment · 768 dimensions Ollama Embedder nomic-embed-text 768 dimensions Postgres + pgvector · documents + chunks · Local learning environment · cosine search Postgres + pgvector documents + chunks cosine search Retrieve Context · embed + top-k search · Local learning environment · retrieval Retrieve Context embed + top-k search retrieval Ollama Chat Model · llama3.2:1b · Local learning environment · local generation Ollama Chat Model llama3.2:1b local generation Grounded Answer · answer + [Source N] · Architecture component Grounded Answer answer + [Source N] asks read text split embed chunks store vectors question embed question retrieve top-k context + question generate with sources return answer Local learning environment

Retrieve

  • • A question becomes an embedding
  • • pgvector finds nearby chunks

Augment

  • • Retrieved text becomes model context
  • • The model receives evidence before answering

Generate

  • • Ollama writes the response locally
  • • Source labels make the answer traceable