Multimodal and document AI quiz

11 questions across 1 lesson. Choose an answer, check it, and read why each option is right or wrong.

Back to the module Flashcards

Checking answers needs JavaScript. You can still read every question and option below.

0 of 11 answered

Keyboard: arrow keys move between options. Space or Enter selects the focused option.

Multimodal GenAI and document AI11 questions

ObjectiveExplain at a high level how a vision-language model reads an image, and why image count and resolution drive cost and latency.

Question 1 of 11Your team switches a document assistant from sending page text to sending page images to a vision-language model. What happens to input cost and why?
Choose one answer.
- It stays the same, because the model reads the same words whether they arrive as text or as pixels.
- It falls, because an image file is smaller than a page of text and providers bill by bytes.
- It usually rises, because each image becomes patch tokens that grow with resolution, up to a cap.
Question 2 of 11To save tokens, a developer downscales every page image far below the provider's cap before sending it. What should they test for?
Choose one answer.
- Small print and table digits may blur, so extraction accuracy can drop even as cost drops.
- The provider upscales small images back to its cap, so the token cost does not actually drop at all.
- The vision encoder restores the lost resolution when it builds patch embeddings, so accuracy is unchanged.

ObjectiveChoose between text-layer extraction, OCR with layout analysis, VLM page reading and page-image retrieval for a document set, and justify the choice with accuracy, cost, latency and privacy trade-offs.

Question 3 of 11You must ingest 40,000 clean, born-digital PDFs of mostly prose on a tight budget. What is the best starting approach?
Choose one answer.
- Send every page to a VLM, since it handles layout, charts and forms better than the other three methods.
- Run OCR on every page, since reading pixels beats trusting the text embedded in the PDF file itself.
- Read the text layer, measure accuracy on a labelled sample, and escalate failing pages to OCR or a VLM.
Question 4 of 11A law firm wants to analyse scanned contracts that contain sensitive client information. Which consideration should weigh most in choosing the approach?
Choose one answer.
- Cost, because scanned pages must first be run through OCR before any model is able to read them.
- Accuracy, so send every page to the most capable hosted VLM and review the output afterwards.
- Data exposure: page images carry everything on the page, so check retention terms or self-host.
Question 10 of 11Which statement about page-image retrieval (ColPali-style) is correct?
Choose one answer.
- It keeps every visual detail on the page, so it needs no evaluation on your own documents.
- It embeds a page image after OCR turns it into text, so it is chunked text retrieval under another name.
- It retrieves whole pages from image embeddings, and a generator still has to read the page it gets.

ObjectiveGiven wrong answers from a document RAG system, separate parsing, retrieval and generation failures and name the likely parsing causes (reading order, tables, charts, scans).

Question 5 of 11Users say the assistant quotes the wrong revenue for a year. The retrieved chunk holds the table row as a string of numbers with no column headers. Where is the failure?
Choose one answer.
- Generation, because the model should have worked out for itself which column each number belongs to.
- Parsing or chunking, because the chunk reached the prompt without headers; later stages cannot recover them.
- Retrieval, because the ranker favoured a chunk of bare numbers over a neighbouring chunk of explanatory prose.
Question 6 of 11A two-column report is parsed with a simple left-to-right text extractor, and answers about the second column contain sentences that appear nowhere in the report. What is the likely cause?
Choose one answer.
- The embedding model has a token limit that cuts off the second column and leaves gaps in the text.
- Broken reading order: lines from the two columns were interleaved into sentences that never existed.
- The model hallucinated because its temperature was set too high for a factual question-answering task.
Question 11 of 11A document RAG system quotes a total that matches only the first of two sheets of a table. The text stored for that table contains the first sheet's rows only, and the second sheet's rows appear nowhere in the index. Where did the failure begin?
Choose one answer.
- Parsing, where the table was cut at the page break, so retrieval and generation never had the later rows to work with.
- Generation, because the model should have inferred the missing rows from the subtotal it could see on the first sheet.
- Retrieval, because the similarity search ranked the second sheet too low and a larger top-k would have recovered the rows.
- Embedding, because the model could not represent numbers and dropped the later rows while turning the chunk into a vector.

ObjectiveDesign chunks for parsed documents that keep tables whole, carry page and section metadata, and support page-level citations.

Question 7 of 11A 3-page financial table is split by a fixed 500-token chunker, and questions about the third page's rows now return wrong figures. What change fixes this?
Choose one answer.
- Increase the chunk overlap so that the header text appears twice in the index around each split.
- Use a larger embedding model so that row chunks without headers still match the question well.
- Split by row groups and repeat the column header and unit in every chunk, with page metadata on each.
Question 8 of 11An answer from a document assistant says "revenue grew 12%" with no source. Why put page and section metadata on every chunk?
Choose one answer.
- It makes the vectors more accurate, because page numbers carry meaning that the embedding captures.
- It lets reviewers open the cited page and check the claim, and lets you measure retrieval per page.
- It lets the generator skip the retrieved text and answer straight from the page number it was given.

ObjectiveDefine an evaluation for a document pipeline (field-level accuracy, page-level retrieval recall, programmatic number checks, human review for low confidence).

Question 9 of 11You extract 100 invoices with a VLM. Which evaluation design best shows whether the pipeline can be trusted in production?
Choose one answer.
- Ask a language-model judge to rate how fluent the extracted fields look, then average those ratings over the set.
- Compare fields to a labelled sample, check line items sum to the total, and route doubtful fields to a person.
- Measure the average time per invoice and the cost per page, since a fast and cheap run means that it works.