Dalberg RAG: Building a Knowledge Base with 95% Recall

Ashish took on the Dalberg project at a point where the client had already attempted to build their own RAG system and was unsatisfied with the results. Dalberg, the nonprofit consulting arm of BCG, had a large Airtable knowledge base containing documents in every major format: PDFs, DOCX files, and PowerPoint presentations. They needed a system that could retrieve accurate answers from that corpus.

His first task was architecture selection. He evaluated trade-offs across database options, deployment platforms, and the approach for connecting the chatbot interface to the RAG backend via an MCP server. Each decision was documented against its alternatives before a path was chosen.

The hardest technical problem was multimodal extraction. Early attempts to extract content from images and PPT slides by passing them to a large language model produced inaccurate text descriptions. Ashish identified that vision-language models were the right tool for image-heavy content, and that PPTs should be converted to images before processing, since they can contain flow diagrams, embedded images, and other visual elements that text extraction misses.

With the architecture settled, he implemented the full RAG pipeline: spawning open-source VLLM models to process the document corpus faster and at lower cost, generating embeddings, ingesting everything into a vector database, and fine-tuning retrieval based on query patterns to maximize recall.

The system was deployed to the cloud and tested against a gold-standard dataset Dalberg provided. The client was satisfied with the results and indicated they were ready to bring Tailored AI another project.