AI Research Assistant
Multi-agent RAG over ~200 arXiv papers on retrieval-augmented generation. Every answer cites the exact sections it came from, and is fact-checked before you see it.
agent-pipeline
How an answer is made
- 1A router resolves follow-ups and screens out off-topic or unsafe requests
- 2Hybrid search and an LLM reranker find the right paper sections
- 3A grader checks the evidence; a verifier fact-checks every sentence
Verifier
4/4 sentences supportedAnswer policy
Only sentences the sources state are shown. When the papers don't cover a question, the assistant says so and names the closest papers it found.
Answered correctly
100%
Unanswerables declined
83%
Attacks succeeded
0%