Multimodal RAG chunking is a mess if you don't have a strict

I'm currently wrestling with a pipeline that ingests PDFs containing diagrams and figures interspersed with technical text. The goal is a multimodal RAG system, but the standard text-splitting approach is failing because it's completely blind to visual elements.

My current setup:

Extracting images from the PDF stream as separate objects.

Using semantic chunking for the text to keep the prose coherent.

Planning to use Google’s multimodal embeddings to map both text and images into a shared vector space.

The problem is the structural disconnect. Standard utilities like LangChain's splitters treat the document as a flat string. They have zero awareness of where an image was physically located relative to the text. If I retrieve a chunk of text that refers to "the figure below," but the image is stored in a completely separate pipeline without a spatial pointer, the context is dead on arrival.

I’m stuck on the implementation details for three specific points:

1. Context Preservation: How do I prevent an image from becoming a "homeless" chunk? If the retrieval engine pulls a diagram, it needs the surrounding text to make sense, but if I bundle them, I risk bloating the context window.
2. Chunk Unit Logic: Should I be treating an image + its caption as a single atomic unit, or should the image exist as its own entity with metadata pointers to the text block above it?
3. Database Schema: How are people actually structuring the vector DB to maintain this association? I'm trying to avoid a situation where the similarity search returns a high-scoring image but the metadata link to the explanatory text is broken or misaligned.

Is anyone actually running production systems using placeholder markers or hierarchical chunking to solve this? I don't want to over-engineer a complex metadata mapping system if a simpler layout-aware parsing approach exists.