papertrail
Turns a corpus of documents that cite each other into a cited wiki with a searchable web app and chat: Marker JSON → section tree → LLM classification → cited wiki pages → SQLite/FTS index, plus an MCP server so a coding agent gets the same search/read/cite access.
Kind
GitHub repo
Author
Jae-Joon Lee (leejjoon)
Added
August 26, 2026
papertrail builds a browsable,
citable knowledge base out of any corpus of cross-referencing papers.
Every claim the system produces — a wiki page, a chat answer, an
agent-published answer — carries the same contract: prose with inline
[n] markers plus a citations[] list mapping each one to an exact
(doc_id, node_id), validated by a single validate_and_enrich()
function. A citation isn’t a similarity score bolted on after the fact —
it’s a pointer into the exact paragraph, table, figure, or equation a
reader (or another agent) can jump straight to.
Three ways to read a built corpus: a web app with inline-cited streamed chat answers, an MCP server (9 tools) giving a coding agent the same search/read/cite/publish access, or the generated wiki markdown itself — no server required.
Proven on a real corpus of ~210 papers (~38,500 indexed nodes) about the exoplanet WASP-12b, plus a second, deliberately differently-shaped demo corpus, both running end-to-end through the full pipeline.