← All questions
RAGMedium· Multimodal RAG· Document Processing· Information Retrieval

A teammate proposes captioning every table in your scan...

The risk is that the captioning process acts as a lossy natural-language compression through a vocabulary chosen by the model. Because the model summarizes the table, specific data points—such as an exact accounts-payable balance in a specific quarter—may be excluded from the generated caption. As a result, even if the number sits unambiguously on the page, the retrieval system will fail to surface the document because the embedding is based on the summary rather than the raw data.

Code playground (Python)