Resident-VLM RAG with Grounded Citations
Offline multimodal RAG built on one resident Qwen3.5-9B running on ROCm with plain transformers (no vLLM). The index pass starts loading the model in the background while it parses PDF, DOCX, XLSX, CSV, logs and source code; each file is guarded, so unreadable or encrypted files are skipped, never fatal. The VLM OCRs every image and image-heavy PDF page, and withdrawn or older-revision documents are flagged. Each query process talks to the resident daemon over a unix socket. Context size adapts to the GPU speed measured at warmup, and retrieval follows shared identifiers (ticket IDs, error codes) across files for multi-hop questions. Answers are forced into compact JSON with quoted evidence, checked against the cited text, and pruned to only the files that are strictly necessary. A YES/NO verifier turns unsupported answers into empty refusals. Outputs are written atomically, ROCm torch is pinned, and weights ship in the image.
.png&w=256&q=75)
.png&w=640&q=75)