Modality relevance is not utility: post-hoc scaling for multimodal RAG

In the current ecosystem of artificial intelligence applied to business, multimodal retrieval-augmented generation (multimodal RAG) systems have become centrally relevant. However, traditional architecture often makes early, binary decisions about which modalities to process —text, tables, or images— based solely on the apparent relevance of a modality to a query. This creates inefficiencies: high computational costs are incurred by activating expensive vision models even when the question could be answered with lighter sources. Recent research shows that the relevance of a modality does not equate to its actual utility. A detailed analysis reveals that a significant fraction of questions whose logical support includes an image can be answered correctly using only text and tables. The optimal decision is not to ask which modality seems necessary, but rather which modality truly adds value to the answer. This gives rise to the concept of selective post-hoc modality scaling: first resolve the question using the cheapest channel (text and tables), run a verifier on the triplet (query, draft answer, evidence) to pinpoint exactly where visual information is missing, and only then activate the vision-language model (VLM) on those specific pieces. This approach, supported by a router calibrated by the value of scaling, recovers the accuracy of an always-on VLM pipeline while drastically reducing visual calls. For companies developing custom applications with AI capabilities, this architecture represents a qualitative leap in efficiency and cost. Instead of over-provisioning infrastructure with universal multimodal models, an intelligent scaling system can be integrated that dynamically decides when it is worth paying for visual processing. This fits perfectly with the philosophy of custom software that we offer at Q2BSTUDIO: solutions that adapt not only to data, but to the resources and business objectives of each client. Our AI for business projects incorporate AI agents capable of managing multimodal flows without wasting computing capacity. Furthermore, we combine this logic with AWS and Azure cloud services to deploy scalable pipelines, and with business intelligence services such as Power BI to visualize results. Cybersecurity also plays a key role: by minimizing calls to external models and centralizing processing, attack surfaces are reduced. Ultimately, post-hoc scaling is not just a technical improvement, but a strategy that aligns innovation in artificial intelligence with the operational and economic sustainability that every organization needs today.