NanoSage
A local research assistant that runs on a laptop using small models, recursively exploring a query with a Monte Carlo search strategy and compiling what it finds into a structured, cited report.
NanoSage refines a user's query, builds a knowledge base from local and web sources, and dynamically explores follow-up subqueries, tracking its own progress through a running table of contents that also serves as a record of the search. Monte Carlo-based exploration balances how deeply it follows one line of inquiry against how broadly it branches into new ones, ranking each branch's relevance to keep the search from wandering into unrelated territory, then compiles the most useful findings into a report using retrieval-augmented generation.
Web search runs through the Tavily API, giving access to academic sources such as PubMed and other scholarly databases, and a hybrid embedding setup uses SigLIP or CLIP for images and PDFs and a small text embedding model for ordinary text, avoiding a dimension mismatch between the two. Getting started means installing Python packages from a requirements file and, optionally, a CUDA build of PyTorch for GPU acceleration; an example generated report is linked from the README.