RAG / Vector DB
Retrieval-Augmented Generation, the pattern that enriches LLMs with your private knowledge via vector embeddings and semantic search. Not a tool: an architecture.
Strengths
- Semantic similarity search, finds meaning, not just keywords
- No fine-tuning needed, update data without retraining the model
- pgvector in Supabase: zero additional infra for small projects
- RAG foundation: enriches LLMs with private company knowledge
Limitations
- Quality depends on chunking, poor splitting = mediocre results, persistent hallucinations
- Embedding cost scales with document volume, each update re-vectorizes
- Architectural complexity: ingestion pipeline, chunking, embedding, retrieval, reranking
Best for
- Chatbots on private documentation: FAQs, wikis, knowledge bases
- AI agents with long-term memory and context retrieval
- Semantic search engines on structured or unstructured corpora