Engineering ·
How VOL AI cites every claim
A look inside the retrieval-augmented pipeline that turns a question into a grounded answer with citations and map coordinates.
VOLEngineering at VOLEngineering
When you ask VOL AI a question, a lot happens in under two seconds. This post walks through the pipeline that makes every answer verifiable.
1. Retrieval
We run a hybrid search that combines dense embeddings with keyword matching over the selected corpus. Hybrid retrieval is more robust to rare terms, acronyms and project codes, which are everywhere in research.
2. Reranking
A cross-encoder reranker scores the top candidates against the question. Only the passages that actually answer it survive; the rest are dropped before the language model ever sees them.
3. Grounded generation
The model writes a concise answer only from the retained passages and marks each claim with the passage it came from. If the passages do not contain an answer, VOL AI says so.
4. From citations to coordinates
Each passage belongs to a document, and each document is a node on the map. We return the node indices with the answer, plus clusters that group related citations. The interface uses them to highlight nodes and fly the camera to the most relevant cluster.
| Step | Typical latency |
|---|---|
| Retrieval | 120 ms |
| Reranking | 180 ms |
| Generation (first token) | 450 ms |
Try it
Open the map, pick a corpus and ask anything. Then click a citation and watch the map fly.