Sidenote

Hallucinating AI summaries plague the current generation of document tools, where plausible-sounding claims bypass verification until the reader discovers they were never in the source material. Sidenote tackles this through a forced citation model: every answer the AI generates must quote directly from the document being analyzed, and the server validates each citation against the actual retrieved text. Wrong claims get dropped. Citations become clickable, scrolling the document to highlight the exact passage. This transforms summarization from something requiring skeptical re-reading into a tool that actually compresses reading time.

The product works as a browser extension across Chrome, Edge, and Firefox, sitting in a side panel next to whatever document the user is reading. It handles the document ecosystem comprehensively: Confluence pages, PDFs with OCR support, Google Docs, Notion, SharePoint, GitHub READMEs, arXiv papers, Slack canvases, and generic web articles. The core features cover typical summarization needs: text summaries, chat interfaces, explanations, glossary generation, and semantic search within documents. Users can highlight passages, attach notes, export outputs as Markdown, HTML, or PDF, and organize their document library with tags and starred collections.

The product positions itself for internal knowledge work first, targeting users who spend time with runbooks, specifications, wikis, and technical documentation where accuracy matters more than polish. This focus is reflected in the data architecture: embeddings live in a UK region with row-level security policies isolating documents per user account. Users choose between two retention modes for each document. Store mode keeps documents indexed for the account lifetime, while Discard purges everything within 24 hours. The commitment to data privacy is explicit: documents are never used to train models.

What distinguishes Sidenote is its acceptance of friction as a feature. The citation requirement means summaries trade some fluency for verifiability, a calculation that resolves a real reliability problem other summarization tools avoid through opacity. The product launches free, remains a solo project built in the UK, and emphasizes shipped features over roadmap promises. For document-heavy work where verification time often exceeds reading time, this tradeoff addresses a genuine gap in the current landscape.