AutoRAG vs manual RAG: the ingest loop is the difference
Manual RAG is a demo; AutoRAG is a pipeline that converts, embeds, and reindexes without a human in the loop each time the folder changes.
Same retrieval, different lifespan
Both shapes retrieve passages and put them in front of a model. Manual RAG stops after the first convert. AutoRAG keeps converting. The difference shows up on day thirty, when the policy PDF changed twice and only one team's answers still cite the old clause.
Who presses the button
In manual RAG a person chooses files, waits for chunks, downloads JSONL or watches a host build an index. In AutoRAG a schedule, a webhook or a folder watcher chooses. The person sets the path and the key once. That is the whole of 'auto': not smarter embeddings, fewer forgotten updates.
Cost profiles diverge fast
Manual rebuilds of an entire library after every small change waste embedding and OCR budget. AutoRAG done well is incremental - only new or changed files convert. Done badly it is a nightly full reindex that burns money for freshness you could have gotten with a checksum. The ingest loop is a design choice, not a checkbox on a model.
Failure modes look different
Manual RAG fails when nobody ran convert. AutoRAG fails when the watcher is down, the API key is empty, or a corrupt file aborts the batch. Good converters skip and name failures so ninety-seven files still land. Test that path before you trust the loop overnight.
When manual is still enough
A one-off diligence folder, a personal notes dump, a spike for a stakeholder demo - convert in the browser and stop. Promote to AutoRAG when the same path will receive files next week without you. Shipping AutoRAG for a static three-PDF experiment is ceremony.