What it does
Give the agent the text of up to 10 documents (60,000 characters in total) and get a concept map: the people, organizations, products and ideas they mention and how they connect. The agent takes document text, not PDF files: ask Claude to read your PDF and pass its text, keeping page breaks as [[page N]] lines. Scanned PDFs with no text layer cannot be read directly; ask Claude to transcribe the pages first. Negation is checked in English and Turkish only, with a fixed word list (such as not, never, cannot and refuses); other ways of denying something may pass as affirmed, Turkish negation by verb suffix (such as "-maz", "-medi") is not checked by the word list, and scripts without spaces between words, such as Japanese or Chinese, are not supported. Your own model proposes the links; the server then checks them as far as text can show. A link marked stated carries a quote that the server found word for word in the cited sentence, with the document and page. The wording of each relation is the model's; only the quote and the concept names are checked. A link marked inferred lists the 2 to 5 sentences it rests on; it is checked only for the concept names and the sentence ids, and the model's reasoning is labelled as reasoning, not a quote. Links that fail a check are dropped and counted by reason, so every map says how many links were proposed, verified, merged as extra evidence and rejected. Nothing is silently cut: a document that is too long or has no text is refused by name. Ask how two concepts connect and get a path where every step shows its quote, or what one concept is linked to. When the same relation is both stated and denied, both quotes are shown together and the agent does not say which side is right; differing numbers, dates or amounts are not detected as contradictions. Text inside a document that looks like an instruction cannot steer the result and is reported by sentence. Shows what your documents say and how things connect. It gives no judgement in law, medicine, money or employment. Links marked inferred are not written in the documents. Your documents are not stored and go only to your own model. The agent uses only the text you give it and never looks people up elsewhere; it cannot verify that you may use the documents, and it does not ask you to declare it. Idea credit: https://github.com/Graphify-Labs/graphify (Apache-2.0); built from scratch. Not affiliated with Graphify Labs.
