// the find
HKUDS/VideoRAG
[KDD'2026] "VideoRAG: Chat with Your Videos"
VideoRAG is a research system for querying long videos using retrieval-augmented generation, built around a graph-driven knowledge index that combines ASR transcripts, visual features, and entity graphs. The paper targets the niche of genuinely long-context video (lecture series, documentary collections, hundreds of hours), not just 5-minute clips. There's also a desktop wrapper (Vimo) built in Electron + Python backend.
The dual-channel architecture — graph-based knowledge indexing alongside a vector store — is more principled than naive frame-sampling or transcript chunking for very long content. The LongerVideos benchmark (134+ hours, 602 queries across lectures/docs/entertainment) is a real contribution; most video QA benchmarks top out at a few minutes. Running on a single RTX 3090 at this scale is a meaningful engineering constraint to have actually tested against. The codebase is reasonably structured: the videorag core is separate from the desktop app, storage backends are swappable (networkx vs neo4j, hnswlib vs nanovectordb).
The Vimo desktop app is still vaporware — Mac beta is 'coming soon' and the README tells you to run from source, which means installing a Python backend plus an Electron frontend with no packaged release. The Python codebase duplicates wholesale into Vimo-desktop/python_backend rather than importing from VideoRAG-algorithm, which is a maintenance problem from day one. Benchmark numbers compare against models without VideoRAG applied (MiniCPM variants), not against other RAG-over-video approaches — that makes the 60.2% vs 56.3% improvement less convincing. Last push is March 2026 and the desktop release is still pending, which suggests the project may have stalled after the paper deadline.