soms3r: Private Document Intelligence on AMD ROCm

soms3r is a high-throughput, privacy-first Retrieval-Augmented Generation (RAG) system engineered specifically for AMD Instinct GPUs and the open-source ROCm software stack. In modern enterprise environments, sensitive documentation, technical schematics, and internal codebases cannot be exposed to third-party proprietary APIs due to compliance and confidentiality mandates. soms3r solves this challenge by delivering an end-to-end, on-premise/cloud-isolated architecture. Powered by vLLM running natively on AMD ROCm, soms3r utilizes the massive High Bandwidth Memory (HBM) of AMD Instinct hardware to serve both heavy dense embedding models (BAAI/bge-large-en-v1.5) and instruction-tuned reasoning models (Qwen 2.5 / Llama 3.1) simultaneously on a single unified instance. Key Architecture Features: - AMD ROCm Acceleration: Native GPU compute optimization with vLLM's paged attention backend. - Zero Cloud Leakage: Completely self-contained document parsing, vector indexing with ChromaDB, and inference. - Accurate Source Attribution: Multi-chunk semantic retrieval with direct source citations to eliminate hallucinations. - Modular Deployment: Designed with a lightweight FastAPI backend and an interactive Streamlit analyst dashboard. Built for the Lablab x AMD AI Academy Challenge, soms3r showcases the efficiency and scalability of building production AI agents on AMD's open AI compute ecosystem.