How it works
SigPhi answers from public-domain primary sources and preserves the link between each answer and the passages it retrieved.
Where the texts come from
The corpus brings together public-domain editions from digital libraries such as Project Gutenberg, Wikisource and the Internet Archive, as well as texts from other public repositories. The Library displays the author, work, language and editorial notes available. Biographies provide context only and are not citable sources. The texts themselves are never altered: cleanup removes only the print artifacts of scanning — page numbers, running headers, words split across line breaks — never a single word of the text, and different editions are never stitched together silently.
Splitting and indexing
Texts are normalized and split into overlapping passages so context is not lost at fragment boundaries. Each passage retains its author, work, language and authorship or completeness warnings. Search combines semantic similarity with word matching before selecting the passages sent to the model.
What “verifiable citation” means
A citation in the chat is published only if it passes checks that do not depend on the model: the sentence supporting it must appear verbatim in the retrieved passage, the author and work in the parenthesis must be those of that passage, and whatever the reader sees in quotation marks must be found in some passage. A second, automatic reading checks that the claim follows from it, and it can be wrong. The chapter or section comes from the citation's actual position in the edition and is not shown when it is not recorded. When part of an answer cannot be supported, its citation is withdrawn and a note says so. Opening a source shows the fragment with the evidence highlighted, the context before and after, and a link to the passage in the reader. In the quotation checker, the author's own passage and passages by other authors who quote the phrase are shown separately, each with its own link; “not found” does not mean “false”.
Languages and translations
The answer is generated in the language of the question and may translate or paraphrase a passage's meaning. In the quotation index, SigPhi provides a generated translation in the page language and shows the included edition's text immediately underneath. The generated translation is not a published edition; historical translations are identified in the catalogue.
Editions and repeated works
The same work may appear in several languages, translations, volumes or editions. SigPhi keeps them as separate records to preserve traceability. When they share a translated title, the index also displays the original title and language so they can be distinguished. The citation always points to the edition actually used, original or translation, and says which: it is never swapped for another edition without an explicit alignment of the two texts. Giving priority to the original is a rule about what enters the corpus, not an automatic redirection of citations.
Limits and corrections
Digitized texts may contain OCR errors, incomplete metadata or imperfect divisions. Generated answers can also misread a passage. Responses can be rated with the chat controls, and under each answer — or from a citation's passage — a specific problem can be reported: the reason, the affected citation and a comment; before sending, you see exactly what will be sent. Each work page says whether the state of its text comes from an automatic measure — which detects scanning debris but not misread words that look like words — or from a check by hand. The study notebook keeps passages and notes only in this browser: notes are the reader's own and SigPhi never uses them as a source.
Found a specific error?