Cognoir reads each document by its actual structure.
Read the document by its actual shape
A 10-K is not the same shape as a Quranic surah. A clinical paper is not the same shape as a Supreme Court decision. Cognoir extracts each format with a parser suited to it — PDFs, Word, slides, spreadsheets, images, web pages and video transcripts — and keeps the structure it finds: headings, chapters, slides, sheets, volumes, pages.
Scanned pages and images go through OCR, with a vision model as the fallback for pages the fast path cannot read. Structure survives extraction, because everything downstream depends on it.
Split on meaning, not on length
Text is divided at the boundaries the document itself provides — a heading, a chapter, a section — rather than every N characters. A passage that is split mid-argument retrieves badly and cites worse: the citation points at a fragment that no longer says what the author said.
Each passage carries where it came from. That is what lets a citation name a volume and a page instead of a file and an offset.
Retrieve twice, not once
A question is embedded and matched against your library by meaning, which returns a broad set of candidates. Those candidates are then re-scored by a reranking model against the actual question, and only the strongest survive.
The second pass is what keeps an answer from being assembled out of passages that merely share vocabulary with the question. When nothing scores well enough, Cognoir says it found nothing rather than answering from the weakest match it has.
Cite the way the field cites
Citation format follows the source type: surah and ayah for Quran, collection and number for hadith, volume and page for a classical work, paragraph and jurisdiction for case law, timestamp for video, section and page for a paper. Bibliographies export in APA, MLA, Chicago, Vancouver, Bluebook or OSCOLA.
Before a quoted passage is shown as the words of a source, it is checked against the stored text of that source. A quote that cannot be found is dropped rather than displayed, because a citation that cannot be verified is worse than no citation at all.
What that buys you
Answers are bounded by your library
Cognoir answers from the documents you gave it. When they do not cover a question, it says so — and offers a web search you have to click, never one it ran on its own.
Disagreements are surfaced, not smoothed
When two of your sources contradict each other, the answer names the point of disagreement instead of silently choosing one.
Your library is not training data
Nothing you upload trains a model. Deleting a document removes the file, its text, and its vectors.
Every document type Cognoir parses, and the citation each one produces, is listed on source types.