Deep Research Agent for Scientific Literature with Cited Reports

Deep-research assistants search the literature, read papers and write reports with citations - but small language models misread evidence and can cite sources that do not say what the report claims. This project builds a deep research agent that runs completely on a laptop with RTX 2050 4 GB with local models in Ollama (qwen2.5:3b). A plan- and-execute multi-agent system built with LangGraph - planner, searcher, reader, writer and verifier - plans search queries, retrieves abstracts (LlamaIndex hybrid search over a local Qdrant vector store and BM25), reads them, writes a verdict and a short report in which every sentence cites a retrieved abstract, and checks every cited sentence against its source. What cannot be verified is removed; if nothing is left, the agent refuses instead of citing unverified text. It is measured honestly on the 300 SciFact dev claims against a classical BM25 + TF-IDF baseline, a closed-book model, single-shot RAG and four ablations. Every step is traced (JSONL, five databases, Arize Phoenix) and a human approves every report before it can be exported.

Category: Department of Computer Science and Engineering / Agent Development

Agent Development Project AG02 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.