obsidian-rag-email-sender

This project turns a folder of Obsidian Markdown notes into a searchable knowledge base, then allows users to ask questions grounded in those notes and draft and send emails grounded in them, with source citations to prevent hallucination. It ingests an Obsidian vault—a directory of .md notes with YAML frontmatter—and chunks them in a heading-aware way so citations point at the right section. The system includes an Ask mode for question answering over the vault using only retrieved context, with the exact source chunks shown and scored, and a Compose Email mode where users describe the email they want, the app retrieves relevant notes, drafts a subject and body grounded in them, lets the user edit it, and sends it over SMTP. A Preview-only toggle (on by default) makes the public demo safe without exposing mail credentials. The generation step is instructed to use only the retrieved context and to cite its sources. Retrieval uses asymmetric embeddings (retrieval_document for notes, retrieval_query for the question), overlapping chunks so facts near a boundary stay retrievable, and the retrieved chunks are surfaced in the UI so a human can verify grounding. The app ships with a sample vault, and users can add a free Gemini key, click Build index, and ask questions. The tech stack includes Streamlit for the UI, Google Gemini 1.5 Flash for the LLM, Gemini text-embedding-004 for embeddings, NumPy cosine similarity for vector search, and stdlib smtplib for email. The index is in-memory and rebuilt per session, suitable for an MVP over a personal vault, with a persistent store as the next step for scale. Obsidian [[wikilinks]] are embedded as plain text, not resolved into a graph, and there is no auth—it is intended as a single-user demo.