Enterprise AI: Blind to Graphs, a Challenge to Overcome

Enterprise AI Facing a Blind Spot

Enterprise artificial intelligence systems, despite their promises, remain largely unable to process non-textual data. This shortcoming is particularly problematic in a context where financial dashboards, architecture diagrams, support ticket captures, and other operational diagrams play a crucial role. Currently, these elements escape the analysis of AIs, which continue to prioritize textual data.

The perception that AI can access all internal knowledge of an organization is therefore misleading. In reality, it misses a significant portion of information, which limits its effectiveness and reliability. The majority of strategic data still eludes current systems, posing a major challenge for businesses.

The Promises and Limits of RAG

Retrieval-Augmented Generation (RAG) has brought significant advancements by allowing AI models to rely on relevant and up-to-date documents. This has reduced hallucinations and anchored responses in verifiable facts. LLMs are finally becoming operational in real professional contexts. However, this approach remains limited by its dependence on textual data.

In an audit report, for example, graphs are as important as the text. Similarly, a technical runbook often relies on annotated screenshots. A market analysis often consists of curves before it reaches conclusions. Traditional RAG, by focusing on text, therefore misses an essential part of the information.

Towards Hybrid and Multimodal Search

To address these limitations, two evolutions are necessary:

• Multimodal RAG, which allows for the processing of text, images, and tables in a common vector space.
• Hybrid search, which is too often overlooked, combines vector and lexical search for increased efficiency.

Vector search is effective for identifying similar concepts, but it cannot replace lexical search for precise tasks like retrieving a contract number or a specific acronym. The integration of these two approaches is therefore essential for a reliable RAG system.

Rethinking AI Architecture

Integrating a multimodal and hybrid RAG pipeline requires a complete overhaul of AI architecture, from data ingestion to response generation. This involves normalizing varied content, merging different relevance scores, and maintaining the coherence of the context conveyed to the generative model.

Solutions like LlamaIndex and LangChain are making progress in this area, while models such as GPT-4o and Gemini are making joint text-image interpretation more viable in production. The next crucial step will be multimodal re-ranking, which will evaluate the coherence of results before submitting them to the generator.

In conclusion, the shift to multimodal RAG and hybrid search is not merely an optimization. It is a necessity for enterprise AI to finally understand and analyze all available data, thus laying the groundwork for truly comprehensive and reliable artificial intelligence. The technical debate thus becomes a strategic discussion, essential for the future of organizations.