RAG & Retrieval · Ground answers in evidence
How do you handle tables, images and scanned PDFs?
Short answer
Preserve tables as structured text with headers repeated per row, describe images or use a multimodal model, and OCR scans with confidence checks. Naive text extraction destroys all three.
A table flattened into a stream of numbers loses its column meaning. An image has no text layer at all. A scan may be OCR'd into plausible-looking garbage. Each format needs a representation that keeps enough structure for the model to interpret it correctly.
- Tables: convert to Markdown or JSON, repeat headers, keep units and captions.
- Charts and diagrams: store a description plus the surrounding text, or index page images for a multimodal model.
- Scans: OCR, flag low-confidence regions, and keep the page image for verification.
- Large tables: index summaries and row groups rather than one giant chunk.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you make a large financial table retrievable?
- 2When would you use a multimodal embedding instead of OCR?
- 3How do you evaluate table extraction quality?
Say it in the interview
I preserve structure. Tables become Markdown or JSON with headers repeated, charts get descriptions alongside the surrounding text, and scans go through OCR with confidence checks. For very large tables I index summaries and row groups instead of one giant chunk. If the user needs a calculation, I would query the data with a tool rather than ask the model to read numbers out of retrieved text.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live RAG & Retrieval mock interview and get scored feedback on depth, clarity and confidence.