This submission implements a high-speed, offline multimodal Retrieval-Augmented Generation (RAG) pipeline powered by Qwen2.5-VL-7B-Instruct on AMD ROCm. To satisfy strict per-query latency limits without incurring repeated model initialization overhead, the system uses a persistent client-server architecture: a resident inference server (server.py) preloads the Vision-Language Model onto the AMD GPU at container startup, while a lightweight CLI entrypoint (app.py) handles indexing and query dispatch over local HTTP. The pipeline parses heterogeneous document formats (PDF, DOCX, XLSX, images, and text), fault-tolerantly skips corrupted or permission-restricted (chmod 000) files under dropped DAC_OVERRIDE privileges, and enforces strict grounding to return accurate answers with exact file citations and calibrated abstention on unanswerable queries.
.png&w=640&q=75)