179 questions
No questions match those filters.
How do you build an ingestion pipeline for a RAG knowle...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe bug that surfaces most often in production RAG systems is stale answers after a document was clearly updated, and it’s almost always the missing delete step. Naively re-processing an updated document and upserting new chunks leaves the old chunks orphaned in the vector store under their previous IDs; if the old chunks happen to score higher on similarity for a given query — often because they were phrased more literally close to how users ask — the system serves outdated information with full confidence, and nothing in the logs looks wrong because retrieval technically “succeeded.” The fix is cheap once diagnosed: always delete every chunk carrying a document’s ID before inserting its replacement chunks, never insert-then-hope-the-new-one-wins.
Change detection should be event-driven rather than polling wherever the source system supports it — a webhook from cloud storage, a CDC stream off a database, or an API’s native change-notification feature — because polling on a fixed interval either wastes compute re-scanning unchanged documents or introduces lag before a change is picked up. Each event handler branches on the operation: a new file goes through the full parse-chunk-embed-upsert path; a modified file goes through delete-then-reprocess; a deleted file goes straight to a targeted delete by document ID with no reprocessing at all.
The quality gate closes the loop: after each ingestion batch, re-run a small fixed set of evaluation queries against the freshly updated index and compare faithfulness or hit-rate against a stored baseline. If it drops meaningfully, halt further ingestion and page a human rather than letting a bad parse (a mis-OCR’d page, a malformed table) silently corrupt the knowledge base that every subsequent query relies on.