RAG ingest automation: from Dropbox to JSONL
Ingest automation means the same conversion step fires whenever a source system drops a file - API or MCP, not a person at a form.
Ingest is a trigger, not a model
Dropbox, S3 events, email-to-folder, a CMS export - each can say 'a file arrived'. Ingest automation is wiring that event to convert, then loading JSONL into your store. The intelligence is in the routing and the parse quality, not in chatting with the file.
One convert contract everywhere
Whether a human batches a folder or a lambda POSTs multipart to /v1/convert, the output should match. Same chunk shape, same manifest header or sibling file. Divergent formats force two importers and discover drift in production.
Auth and billing in the loop
Automation carries an API key in env or a secret store - never in the event payload. Prepaid credit and 402 handling matter: a stuck watcher that retries forever on empty balance is a pager, not a pipeline. Check balance before large bursts.
Agents as ingest workers
MCP tools can convert local paths when the agent host sits on the same machine as the drop folder. That is still ingest automation - the agent is the worker, the key is still human-provisioned, Stripe still is not called by the model.
Prove it with a dry run
Copy five representative files into a staging path, run the same job the watcher will run, inspect skips and chunk counts, then load into a throwaway collection and ask two questions only the new files can answer. If that fails, do not attach production Dropbox yet.