Lesson 06: Prepare information for questions and answers
Before adding files to a knowledge base, inspect them. A handbook, screenshots, and a chat excerpt can contain old rules, missing words, and unattributed retellings. A careful model may still repeat faulty evidence.
Preparation begins by deciding whether each document can serve as evidence.
Work backward from questions
Questions about limits, loan periods, and weekend closing need corresponding policies. Event news cannot replace borrowing guidance. Write common questions and identify supporting sources rather than maximize file counts.
Leave gaps for unsupported questions: the pack lacks fine and deposit policies. Gaps expose coverage limits; filling them with familiar amounts turns guesses into supposed evidence.
Give numbers and negation another look
Images, PDFs, and tables may need text extraction. Check numbers, units, and negation against the original. Fourteen becoming forty is noticeable; dropping not from cannot borrow can hide in fluent prose. Check table row associations too.
Manual transcription is sufficient for short passages. Reliable text matters more than building an automated parser for eight cards.
Keep dates and status with the text
B says three books and E says two. Text alone looks contradictory; titles, dates, and E’s superseded status explain which applies now and which supports historical comparison.
These fields are metadata. Start with title, origin, publication or effective dates, scope, and status. Publication and effective dates can differ. Retained archives need their archive labels.
Make a source list for A–H or three to five public documents. Describe what each answers, check five numbers or conditions, and identify two remaining gaps. This shows what needs preparation next.
Distinctions to look for
B is current, E archived, and D date-specific. Fees and deposits remain unsupported. Combining two and three into a flexible two-to-three limit invents a rule neither source states.
Further reading: Microsoft: RAG preparation phase
← Previous · Learning plan · Fictional library practice sources · Next →