What is AutoRAG?
AutoRAG is retrieval that stays current because conversion runs on a schedule or on file change - not because someone re-uploads a PDF into a chat.
The name is doing real work
AutoRAG is not a brand for a smarter chatbot. It names the missing half of most RAG demos: the ingest loop. Retrieval-augmented generation only answers from what you indexed. If the index is a manual dump from last Tuesday, the answers are last Tuesday's answers dressed as live knowledge. AutoRAG means conversion - parse, chunk, embed - fires without a person dragging files into a form whenever the source library moves.
Manual RAG is a screenshot of the truth
A common path looks finished: upload three PDFs, ask a question, get a grounded reply. That is a prototype of retrieval, not a system. The next version of the handbook, the new policy deck, the weekly export from the share drive - none of them enter the corpus until someone remembers to upload again. AutoRAG treats that upload as a job, not a habit.
What has to be automated
Detect new or changed files. Convert them into the same JSONL shape your store already loads. Upsert vectors and metadata. Optionally drop or re-chunk files that disappeared. The language model is not in that loop. It reads retrieved passages after the fact. Confusing those jobs is how teams pay frontier prices to re-translate a folder they already own.
Where conversion sits
Conversion is the expensive step you run least often relative to questions - but relative to file churn, you may run it daily. Browser convert suits a one-off folder. An API or agent tool (MCP) suits a watcher that POSTs each arrival. Either way the artifact is the same: chunks with embeddings and a manifest, ready for SpaceDB, pgvector or Qdrant.
What AutoRAG is not
It is not auto-tuning prompts. It is not a model that invents a knowledge graph unaided. It is not 'chat with your Drive' as a hosted product that keeps your files. Those can be useful products. AutoRAG, in the sense worth searching for, is automated conversion into a corpus you control so retrieval stays honest as the files change.