retrievr

How it reads

Document intelligence

Every retrieval product eventually answers the same question: what did it do to my file before it searched it? If the answer is “split it into 500-character pieces”, the quotes that come back can be accurate to the letter and wrong about the argument, because a piece can start in the middle of a sentence and end in the next chapter.

Here, that answer is the engineering. A document is read the way it is built, cut where the document itself changes subject, and cited with the range it came from.

One file, three shapes

A PDF is a stack of pages, so it is read page by page — in the order the pages are numbered, not the order the file stores its objects. An EPUB is not a stack of anything: its spine defines the reading order, and the table of contents, the colophon, the index and the publisher boilerplate are all dropped before anything is embedded. Plain text and Markdown are already text, so they are not reparsed into a worse version of themselves.

That matters at the end, not the start. A citation that says “chapter 4, section 2” is only possible because the reader kept the chapters. A chunker that never saw a chapter cannot tell you which one it just quoted.

Passages are cut at sections, not at a character count

Each passage is packed to about 400 tokens, and never spans two sections. Four hundred is small on purpose: a passage is the unit an agent reads, so it has to be specific enough to be worth quoting and large enough to carry its own argument. A boundary at a heading means a quote cannot silently include the next topic — the most common way a citation looks right and is not.

Nothing is truncated to fit. A section longer than the budget becomes several passages, each with the offset it starts at, so the character range a citation returns points at the real bytes in your file. A search then returns a window of at most about 1,500 tokens — smaller than the file, larger than the passage — because the sentence before a claim is usually what makes the claim mean anything.

Every language, on the same scale

A document in one language and a question in another still have to meet somewhere. They meet in the vectors: text is embedded so that a Greek contract and an English question about it land close together when they mean the same thing, and apart when they do not, rather than only when they share a word.

Retrieval is hybrid: a vector leg for meaning, a lexical leg for the exact word — a clause number, a surname — and a phrase leg for the sentence you half-remember, fused by rank. Meaning alone misses the serial number; keywords alone miss the paraphrase. A citation needs both.

What a citation carries

Title, author, section path, the exact character range, the licence, and the score that put it there. Enough that a reader can open the page the agent quoted and find the sentence on it, and enough that a wrong answer can be taken apart: you can tell whether the search missed the passage or the model misread a passage it had.

Two things to know before you upload

Archives are refused rather than expanded. Re-uploading the same file creates a second document — there is no deduplication, so upload once and rename it.

Cost

Uploading and indexing are free; your first five documents need no payment, and after that one $5 top-up unlocks uploading. A query costs $0.01, charged when it is accepted.

See it on your own file

Paste this into the agent you already use, and let it work:

Fetch https://retrievr.io/install and follow it.

Then upload a PDF you know well and ask it a question whose answer you could point at. The docs cover uploads and the tools your agent holds, and the price list is one screen.