This project delivers a Multimodal Retrieval-Augmented Generation (RAG) system that answers complex questions by grounding responses in visual, audio, and textual evidence. It converts PDFs into images for vision processing, encodes images with CLIP, transcribes audio with Whisper, and embeds text using SentenceTransformer; all embeddings are indexed in ChromaDB for fast approximate nearest-neighbor search. At query time, the pipeline performs multimodal retrieval—fetching relevant images and transcript segments—then composes a unified context bundle that Qwen-VL can “see” and reason over. The model generates an explanation that references the retrieved artifacts while the UI surfaces both the answer and the supporting images/snippets for auditability. Designed for real workloads such as document Q&A, compliance reviews, meeting/audio search, and visual inspections, the architecture favors modularity (you can swap models or the vector store), GPU acceleration where available, and pragmatic engineering touches like chunking, batching, and caching to keep latency low. Optional re-ranking improves cross-modal relevance; simple guardrails filter unsafe inputs. The result is a practical, extensible blueprint for multimodal assistants: transparent enough for stakeholders, flexible enough for rapid iteration, and robust enough to move from prototype to production without rebuilding the core pipeline.