Cuestionario vinculado a una guía · Medio Nivel

Prueba RAG multimodal

Pruebe lo que sabe sobre RAG multimodal: subtítulos de imágenes en texto, incrustaciones de imágenes y texto compartidas y recuperación de capturas de pantalla de páginas al estilo ColPali.

Rutas de guía relacionadasRAG multimodal
Pregunta 1 de 8

What typically goes wrong when a text-only RAG pipeline processes a PDF full of tables and charts?

Vista previa de cada pregunta8 preguntas

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

  1. What typically goes wrong when a text-only RAG pipeline processes a PDF full of tables and charts?
    • The PDF becomes too small to index
    • Tables become jumbled and chart values or diagrams are lost
    • The embeddings become images
    • The model refuses to answer
  2. What is the main weakness of the caption-everything-into-text approach?
    • It cannot use existing text embeddings
    • It requires a vector database for images
    • It only works for scanned forms
    • Anything the captioner fails to describe is lost to retrieval
  3. What do CLIP-style models make possible in multimodal RAG?
    • Embedding images and text in a shared vector space so text queries can retrieve images
    • Reading dense small text on scanned pages with high accuracy
    • Compressing PDFs before storage
    • Translating captions into other languages
  4. What is distinctive about ColPali's approach?
    • It converts every page to Markdown with OCR first
    • It only indexes image captions
    • It embeds page screenshots with a vision-language model and skips OCR and layout parsing
    • It stores one vector per document
  5. In late interaction, how is a query scored against a page?
    • One query vector is compared with one page vector
    • Pages are ranked by file size
    • Keyword counts are compared
    • Many query-token vectors are compared with many image-patch vectors
  6. Why are ColPali-style indexes much larger than text-chunk indexes?
    • They store the original PDF files inside each vector
    • Each page gets roughly one vector per image patch instead of a single vector
    • They duplicate every page in several languages
    • They require storing the full model weights per page
  7. Why pass the page image, not only extracted text, to the generating model?
    • Questions about chart values or visual layout need the model to see the visual itself
    • Images are always cheaper than text
    • Text cannot be cited
    • Generators cannot read text
  8. Which statement about multimodal RAG in practice is accurate?
    • It always replaces text pipelines entirely
    • It only works on photographs
    • Hybrid systems combining parsed text, table extraction and page-image retrieval are common
    • It requires no evaluation