Feature Description

A node reranker that can handle multimodal data. Today, AFAIK, this would only affect images and text since I do not know of a case (except for perhaps custom retriever implementations) when any retrievers return audio/video data. This allows information that is in say a powerpoint image to be ranked according to a search query.

Reason

In order to achieve the goal of Multimodal Pipelines/Engines, we need node postprocessors that can support multimodal data.

Value of Feature

In certain document types (especially pptx, but also some pdfs), considerable amounts of information may be stored in images. Today, node postprocessors cannot handle ImageNodes.