wtf is second-brain-rag?

Second-brain-rag is an experiment that asks whether an AI agent can be given a structured memory similar to the note-taking app Obsidian, using linked markdown files and a knowledge graph, instead of just storing everything as raw text chunks in a vector database. The project tests this idea on a small set of insurance documents, covering things like underwriting rules and claims procedures. Each source document is written as plain markdown with a YAML header describing its title, type, tags, and links to related documents, using the same double bracket wikilink style Obsidian uses. When the documents are loaded, a small routing model decides whether a question can be answered by looking directly in the existing set of documents, or whether it needs a combined search using both keyword matching and semantic similarity. Whichever path is chosen, the top matching documents are then expanded by following their wikilinks up to two steps away, pulling in related concepts before an answer is generated. The project includes three web pages: one explaining the underlying idea, one showing the document ingestion process building the knowledge graph live, and one for browsing the resulting graph and clicking on individual concept nodes. The final answers are generated by an AI model, with the documents the answer was based on shown alongside it. Under the hood it combines several pieces: cloud hosted AI models for routing questions and generating answers, a semantic search model, a classic keyword search algorithm, a small custom scoring method that blends both search results, a Python web server with live streaming updates, and a graph visualization library for the browser. To try it, you install the Python dependencies, add your own cloud API keys to a config file, and run the server locally. The project is licensed under the MIT License.