CiteLens

Inspiration

AI assistants answer fluently — including when they are wrong. For research, the cost of a confident hallucination is high, yet most tools bury their sources in a link dump and never check their own claims. We built CiteLens around a simple rule: if a sentence can’t be traced to retrieved evidence, the user should see that immediately.

What it does

CiteLens plans a research question, retrieves evidence with Tavily, generates a structured report with Nebius Token Factory (NVIDIA Nemotron), and runs an independent verifier pass that re-reads every sentence against the retrieved snippets. Verified sentences keep visible [n] citations; sentences the verifier cannot support are explicitly flagged in the report instead of slipping through. A grounding score (verified sentences / total sentences) summarizes trustworthiness at a glance. With no API keys it runs fully offline on a bundled sample corpus; with keys it runs live against fresh web sources.

How we built it

A single-file, stdlib-only Python service (no framework, no build step) keeps deployment trivial: python3 app.py. ReAct-style loop: plan → search → read → draft → verify → revise. Sentence-level verification uses token-overlap grounding against retrieved evidence plus an LLM verifier in live mode. The frontend is plain HTML/CSS/JS with a LIVE/DEMO mode badge so judges can tell at a glance which mode a deployment is running.

Built With

• nebius-token-factory
• nemotron
• python
• tavily