visual-retrieval

🐊 Snappy's unique approach unifies vision-language late interaction with structured OCR for region-level knowledge retrieval. Like the project? Drop a star! ⭐

Vision-first document RAG for PDF, DOCX, and image QA/extraction with ColQwen2, Qwen2.5-VL, Qdrant, and FastAPI.

RAG Lab — Local visual document analysis workbench powered by ColPali embeddings and Ollama vision models

Drop any document corpus in, get a grounded enterprise agent out. Visual retrieval over page images, a navigable 3D universe of files, page-exact cited verdicts. Built solo at the RAISE Summit hackathon (Vultr track), powered by NVIDIA Nemotron.

A Faster R-CNN object detection model for recognizing handwritten mathematical symbols on whiteboards. Built as the core detection engine for an end-to-end math expression visual retrieval pipeline.

Structure-aware post-NMS refinement resolving geometric containment for visual retrieval.

Multimodal RAG with per-query routing to text / visual / hybrid retrieval paths. Vision-LLM answers and regression-gated evals.

Local-first visual memory prototype: pixel tiles, hypergraph evidence, image retrieval, reports, and demo video.

Scoped ColPali reproduction with a tested, tuned Triton MaxSim kernel on NVIDIA L4

Open-source PR proposal workflow skill｜公关人专用开源提案技能：Brief → 洞察 → 策略 → Campaign → Pitch Deck