Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo cya RAG Ikibazo

Gerageza ibyo uzi kuri multimodal RAG: kwandika amashusho mumyandiko, gusangira amashusho-inyandiko, hamwe na page ya ColPali-page-yerekana amashusho.

Inzira zijyanye no kuyoboraRAG
Ikibazo 1 Bya 8

Ni ubuhe buryo butagenda neza iyo umuyoboro wa RAG wanditse gusa utunganya PDF yuzuye imbonerahamwe n'imbonerahamwe?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni ubuhe buryo butagenda neza iyo umuyoboro wa RAG wanditse gusa utunganya PDF yuzuye imbonerahamwe n'imbonerahamwe?
    • 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. Nubuhe ntege nyamukuru yibisobanuro-byose-muburyo-bwanditse?
    • 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. Niki cyerekana imiterere ya CLIP ituma bishoboka muri RAG ya multimodal?
    • 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. Ni iki gitandukanya uburyo ColPali yakoresheje?
    • 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. Mugutinda gutinda, ni gute ikibazo cyatsinzwe kurupapuro?
    • 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