Rag Dev

Butterbase RAG (Retrieval-Augmented Generation)

Two tools cover the entire RAG surface:

• manage_rag_content — collections, document ingestion, status polling, deletion
• rag_query — semantic search, optional LLM synthesis

Documents are ingested asynchronously: text or files become embeddings stored in pgvector, and queries do a similarity search at runtime.

---

1. The mental model

Collection                        Documents                     Chunks
──────────                       ──────────                     ──────
"product-faq" ──────────────►   doc_1 (PDF) ───────────►       chunk 1, 2, 3...
                                doc_2 (text) ──────────►       chunk 4, 5...
                                doc_3 (markdown) ──────►       chunk 6...

A collection holds documents; a document is split into chunks and embedded; rag_query searches by cosine similarity across chunks within a collection.

chunk_size and chunk_overlap are set once at collection creation and immutable — to change them, delete and recreate the collection.

---

2. End-to-end workflow