Silent Key RAG: cited answers on AMD ROCm Silent Key RAG is a single container built on the mandated ROCm PyTorch base image, with its layers kept byte for byte so the base check passes. One resident process loads Qwen3-VL-8B once and keeps the index in memory, so each question pays only for retrieval and one model call.
Every file goes through a defensive parser. Empty folders, unreadable and encrypted files, and unknown formats are skipped and logged instead of crashing the run. Images are read by the vision model, so answers that only exist inside a picture are still found.
Retrieval is exact BM25 over chunks. The model proposes an answer and its sources; deterministic code then checks that the answer really occurs in a readable file, keeps exactly one answer file (current over superseded) plus a second only when it is a genuine bridge, and runs a short yes or no check that turns near misses into refusals, because citations are graded as an exact set. When nothing in the corpus supports an answer, the container refuses rather than guessing.
The image ships its own weights, needs no network at evaluation, and is sized for the 1 to 48 GiB VRAM window.
Match 4 (Intelligent Web Information Retrieval): the form has no Mini Challenge 4 Image field yet, so the image is listed here. Mini Challenge 4 Image: ghcr.io/silentkeytech/mc4-web:v1 (public). It answers each question with one exact value and its source page, runs on the same ROCm base image, and scored 10 of 10 on our own tasks on an AMD Radeon Pro W7900D.
.png&w=640&q=75)