LIBRAIN: A Multi-Agent RAG System for Scientific Discovery

LIBRAIN is a multi-agent retrieval-augmented generation (RAG) system for scientific discovery, built on .NET 10 with Claude, OpenAI embeddings, and Qdrant. A Reader agent ingests papers, chunks them, and embeds them into a vector store. Two pipelines then run over that retrieval layer. A grounded query pipeline (Synthesis and Evaluator agents) answers using only the retrieved chunks, and a citation-validation contract guarantees that every cited chunk resolves to a real retrieved document, giving zero fabricated references by construction. A separate Discovery pipeline is allowed to reason beyond the sources, but isolates the speculative portion into an explicit novelClaim field that is labeled and scored apart from the grounded scaffolding. A cross-cutting audit layer assigns one correlation identifier per request, so any output can be traced back to the exact chunks it used. The paper evaluates LIBRAIN against Single-LLM and Naive-RAG baselines over ten cross-domain topic pairs, together with a human study (n=30 across two independent three-rater panels). The citation contract removed fabricated citations entirely, and on a cross-provider fabrication-delta harness the system caught 67 fabrication attempts with none leaking. Speculation fencing raised novelty at a measured cost to plausibility, and the reported results include that trade-off rather than only the favorable axes. Source code and evaluation material: https://github.com/erennmutlu1/librain