How Intric searches your documents

Most tools treat a document as a pile of text and match words against it. Intric works through the document first, so it knows what it’s holding before anyone asks — and when you do ask, it blends several ways of looking depending on what the question needs.

What happens when you add a document

Reading the page as a page

Intric reads a page as a page, not as a stream of words. A heading belongs above its table, a caption belongs below its chart, and a table keeps each value under its column header. A scanned page is read with OCR, so it counts as much as one that was typed. Where a page carries a diagram or an infographic rather than text, Intric describes what the image shows and keeps that description in the right place on the page — nothing is quietly dropped.

Summarising the document

Before any question is asked, Intric already has a sense of what each file is about. That’s how it knows which file to open when you ask something — much like a colleague who remembers roughly what’s in each report before walking to the shelf.

This matters most when an assistant has many files. Searching all fifty at once turns up weak matches from documents that have nothing to do with the question, and those can crowd out the passage you need. Picking the right file first keeps the answer focused.

Indexing for navigation

The document is split into short passages, and each one remembers which page it came from and where it sits in the document. That lets Intric move through it in order, page by page, the way you’d flip through a book — and it’s what lets every answer point back to a page.

Spreadsheets are the exception. They aren’t cut into passages at all, because a table in fragments can’t be added up or filtered. Instead, Intric works with the actual rows, calculating the answer to a question like “how many cases were closed in March” rather than estimating it.

How a question is answered

When you ask a question, Intric weighs up several ways of looking and blends whichever suit the question. You don’t choose — it decides. Each method is good at something the others miss, which is why it has all of them.

Keyword search

For exact matches — a case number, a name, a specific regulation — Intric prioritizes word-for-word matching.

Why it matters: a search by meaning treats “Article 12” and “Article 21” as nearly the same thing, and it can drift toward a paragraph that discusses the topic in general. When the exact string is the whole point, only a word-for-word match is reliable.

Semantic search

For meaning rather than wording. Ask about “the risks of working night shifts” and it also finds a section on “health effects of irregular schedules,” because it’s matching concepts, not terms.

Why it matters: people rarely ask a question in the words the document happens to use. Without semantic search, you’d need to know the document’s vocabulary before you could find anything in it.

Reading the whole document

For tasks that need the full picture, such as summarizing a short report or comparing two agreements clause by clause.

Why it matters: some questions can’t be answered from passages, because the answer depends on the whole. For long documents, though, reading everything would bury the relevant parts, so Intric only reads a document in full when it’s short enough. For anything longer, it searches and reads selectively instead.

Navigational reading

For finding a place and then following it. Ask about chapter eight, and Intric locates where that chapter starts, then reads forward page by page until it reaches chapter nine.

Why it matters: a search returns the few passages that match best, which for a chapter means its opening and some scattered paragraphs. Reading forward is how Intric gets everything in between, the way a person would, instead of summarizing a chapter it has only half seen.

How they work together

Keyword and semantic search aren’t two separate choices. Every search blends them, and Intric sets the balance for each one: a precise factual question might lean mostly on keyword matching with some semantic search alongside it, while a broad, conceptual question might be answered almost entirely semantically.

The other methods build on that. A typical answer starts by picking the right file from its summary, uses a blended search to find the right place, and then reads forward from there if the question needs more than a single passage. Intric may repeat these steps several times for one question before it answers.

Why not read everything

It can seem simpler to hand the model every document in full and let it find the answer. In practice, models given very large amounts of text at once start to overlook what’s in the middle and lose track of details. Working with a small amount of relevant text, and widening it deliberately when needed, gives more accurate answers — and keeps every part of the answer traceable to its source.

For the technical detail behind each of these steps — how pages are read, how figures are described, and how the index is built — see OCR pipeline.

Why answers carry references

Every answer is traceable back to its source. You can see which document, which page, and which passage the assistant drew on, and open it to check. That’s what makes it possible to rely on an answer rather than take it on faith — the reference is the point, not a footnote.

Why the original file matters

This is also why it matters that the assistant has the original file. References point into the document you uploaded, so the closer that file is to the real thing, the more useful the reference is. An exported or converted version of a document — plain text pulled from the legacy knowledge system, for instance — won’t give you references as reliable as the original PDF or Word file would.