FDE-AI Enterprise RAG & Hybrid Retrieval MEDIUM
PRODUCTION SCENARIO
An SOP retriever splits 4,000 maintenance procedures on a fixed 800-token window. A numbered step and the prerequisite table that governs it land in different chunks, so retrieved steps arrive without the warning that applies to them.

Which ingestion experiment best targets this defect?

Answering here is anonymous. Nothing is saved unless you sign in.

Show answer and explanation

Answer: Split on the document's own heading structure instead of a fixed window

The defect is a boundary placed without regard to the document's structure, so the fix is to let the headings decide where chunks end, which keeps each procedure with its prerequisites. A 200-token window cuts the same relationships more often. Fifty percent overlap copies neighbouring text, which does not reach a warning several sections away. Section metadata narrows which chunks are eligible and never puts the warning text into the step's chunk.
Free

Keep practicing FDE-AI

24 original FDE-AI practice questions, each with an explanation and a source link. No account needed.

Start free practice set → Timed, explained, free