Most real documents are not plain prose: they carry tables, charts, diagrams and screenshots.
A text-only pipeline silently drops all of it, and the answer is missing whatever lived in the figures.
One approach describes each image or table with a vision model and embeds that description.
Another embeds the page image directly, so the original picture is what gets retrieved.
Tables deserve special handling, since a flattened table loses the row and column relationships.
The documents people actually want to search are rarely walls of clean prose.
Financial reports are mostly tables, technical manuals are full of diagrams, and slide decks are almost all figures.
A conventional pipeline extracts text and quietly discards everything that was not text.
The result is a system that answers confidently while never having seen the chart holding the answer.
Tables suffer a subtler version of the same problem, since extraction usually flattens them.
Once rows and columns become a run of words, the association between a value and its header is gone.
The first family of solutions converts non-text content into text at indexing time.
A vision model writes a description of each chart or figure, and that description is embedded.
Tables are serialised deliberately, often to markdown or to one sentence per row, preserving headers.
The original element is stored alongside so it can be shown or passed to the model when retrieved.
The second family skips the conversion and embeds the page image itself into a shared space.
Models in this line index page screenshots directly and retrieve them against a text query.
That keeps layout, which matters for documents where position carries meaning.
Generation then needs a vision-capable model, since the retrieved context is an image.
In practice a hybrid works best: text for prose, structured extraction for tables, and image embeddings for figures.
One question on this card. Every option explains itself once you pick, right or wrong.
19 concepts in this module, then a graded test with every answer explained.