The engine room
The library is the demo
Falsafa reads as a quiet library: 949 works, 9,124 chapters, free and public. Underneath it is a working argument about how machines should read: a librarian, with deterministic tools and honest citations, instead of a vector database. This wing of the site is that argument, with the receipts.
The historical case comes first: Carried Across, a short book on how ideas travel, argues that knowledge survives only when carriers keep the chains legible, and the atlas behind it traces fourteen books across a thousand years of named hands. Falsafa is that practice continued for the newest readers. The pages below are the engineering that follows from it.
/try
Try it live
Ask the librarian a question with your own API key. Any model, streaming, every citation resolvable to a paragraph. Install instructions for Claude, Cursor and Codex included.
/eval
The eval explorer
Over a thousand adversarial retrieval cases, browsable one by one: the question, each model arm, the tool trace, and the cited paragraphs deep-linked into the corpus. Audit any row.
/thesis
The thesis
Why this corpus ships with no vector database: markdown, stable paragraph IDs and ten deterministic tools, with the ablation charts that test the claim instead of asserting it.
/numbers
By the numbers
Corpus totals, eval pass rates, build metrics. Every figure deep-links to the data that produced it.
Run the librarian yourself
The whole corpus and all ten tools ship as one npm package, no API key, no server, no telemetry. The first run downloads about 48 MB of corpus.
claude mcp add falsafa npx -y @falsafa/mcpMore install targets on the try page. Source, pipeline and eval suite on GitHub.