Guides · measured July 2026
How to search inside PDFs and Word docs, not just their names.
You remember a phrase (“ground coffee beans”) but not the file. It's on page 7 of a PDF your scanner named scan_0032.pdf. Type it into Windows Search on a real 4.47-million-file drive and you wait about 67 seconds, if it finds it at all. The reason is simple: most tools search the file's name, and your phrase lives in its text. Here is how document-content search actually works, why the OS is slow at it, and how to search inside every document in milliseconds.
TL;DR
- →
Most tools match a file's name, but the phrase you remember lives in its text. A byte scanner like grep or ripgrep can't even read a PDF or Word file, because the words are stored compressed and encoded.
- →
The OS proves it can be done: Spotlight indexes a Mac's whole volume, while Windows Search covers only a slice and throttles the rest: a median near 67 s on a 4.47M-file drive, about 480,000× slower.
- →
Interlinked decodes and indexes the text of every PDF, Word, Excel and PowerPoint on the drive: content queries land in 7 to 9 ms, filenames in about 85 microseconds, and a file you just saved is searchable roughly 1 ms later, all from about 44 MB at rest.
- →
Your agents get the same index over MCP: a document hunt that took Claude Code 6m57s and 71 tool calls comes back in 16 ms and one call, about 7,200,000× faster, with the context it burns dropping from ~58% to almost nothing.
- →
Everything local is free forever, no card. Hosting starts at $5.99/mo.
Two different searches
The answer is in the text. You keep searching the name.
There are two completely different things people call “search.” One matches the filename: fast, and what your fingers reach for by habit. The other matches the words inside the document. When you can't remember what a file is called (a scanned receipt, an exported report, a contract someone emailed you) only the second one can help. Matching the name is the easy part; Interlinked does it in about 85 microseconds. But the name is the wrong place to look.
The picture is the whole idea. Full-text document search reads the words on the page, so a query lands on the file that contains the phrase, even when its name is a camera string or a scanner counter. Filename search never opens the file, so it can only find what someone already typed into the name.
Why you can't just grep a PDF
A PDF's words are packed. Something has to unpack them first.
The obvious move is to point grep or ripgrep at the folder and search. It comes back empty, and not because the phrase isn't there. A PDF, a Word .docx, an Excel sheet: the readable text is compressed and encoded inside the file. A byte scanner reads those raw bytes, sees what looks like noise, and finds no plain-text match. To search a document you first have to decode its text, then index the words that come out.
This is the real dividing line. Tools that scan bytes on demand (grep, ripgrep, find) can race through plain-text code, but they can't read a PDF or an Office file at all. Tools that decode and index the text can. The only question left is how much of your drive they bother to do it for.
Every option, honestly
Who can read the text inside your documents?
The two famous answers are built into your operating system. Windows Search and Spotlight both decode and index document text (they prove it can be done) but each was built for a person typing into a box, and they differ sharply in how much of the disk they cover.
| The tool | Reads text inside a PDF / Office file? | Coverage | Time · 4.47M files |
|---|---|---|---|
| grep / ripgrep | No, sees encoded bytes | wherever you point it | skips the file |
| find / dir | No, matches names only | names, never the text | walks the tree |
| VS Code search | Text files in the open workspace | one workspace | seconds, in-workspace |
| Windows Search | Yes, in indexed folders | a throttled slice of the drive | 67 s median / empty outside |
| Spotlight (macOS) | Yes, the whole startup volume | macOS only | as fast as you type |
| Interlinked | Yes, every document | whole drive, kept current | 8 ms average |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. The Windows Search figure is a median for content that lives outside its default indexed folders; the Interlinked figure is the average content query on the same corpus (8 ms; 7 to 9 ms typical). The middle rows vary with scope, so they are described rather than timed. Full content methodology in the ripgrep comparison.
Windows · the built-in
Decodes and indexes document text, but only for a subset of folders (your profile, mail), and the indexer is throttled so it never disturbs the rest of the machine. Outside that slice, a content query walks the disk or comes back empty. On 4.47M files the median lands near 67 seconds.
macOS · the good example
Indexes the text of the whole startup volume by default and keeps it current, so a content search almost never falls back to a walk: as fast as you can type. This is the right model. But it is macOS-only, and it was still built for a human at a search box, not an agent making hundreds of lookups a minute.
The lesson of the two anchors is coverage plus freshness. Spotlight feels instant because it covers the whole volume and stays current; Windows Search feels slow because it covers a slice and throttles the rest. On Windows (and for an AI agent that never types into a box) that coverage has to come from somewhere else. (More on the desktop split in Spotlight vs Windows Search.)
Same drive · same documents · measured
Cover every document. Search the text in milliseconds.
Two honest comparisons. For plain-text files, ripgrep is the fast baseline, and Interlinked still returns 58,625× quicker across the test set, because it never walks the disk: on one representative query, ripgrep's 93.8 seconds became 16 milliseconds. For PDFs and Office documents, ripgrep can't read the text at all, so the honest comparison is against the OS indexers, and there the difference is coverage: every document, not a slice. And because the index stays current as files change, edits show up in about a millisecond: the answer is never a day stale.
The fast answer
Whole-drive document search. For you, and for your agents.
Interlinked decodes and indexes the text of every document on the machine (PDF, Word, Excel, PowerPoint and more) alongside your code and everything else, and keeps it current as files change. Content search across the whole drive averages 8 milliseconds; a filename lands in about 85 microseconds. Nothing falls back to a walk, because the whole drive is already covered.
Agents are the worst case: they ask hundreds of times per session, and they can't open a PDF any better than grep can. So Interlinked hands the same decoded-text index to any AI client that speaks MCP: Claude Code, Cursor, Codex, Cline, Continue and more. Same agent, same question: the file hunt that took Claude Code 6 minutes 57 seconds and 71 tool calls comes back in 16 milliseconds and one call, and it can finally search inside the documents on your disk instead of guessing from their names.
Do this today
Search inside every document you own.
Install once. One signed installer indexes the whole drive (names and document text together) and auto-configures 19 AI clients (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest). No indexing settings, no JSON editing.
Search the words, not the name. Type a phrase you remember (a client's name, a line from a contract, a figure from a spreadsheet) and land on the PDF, Word, Excel or PowerPoint file that contains it, whatever it's called.
Trust that it's current. The index tracks changes as they happen, so an edited or newly saved document is searchable about a millisecond later. No nightly rebuild, no stale results.
Let your agent read inside too. Claude Code, Cursor and Codex query the same document index over MCP, so they stop guessing from filenames and start searching the text on your disk.
Everything local is free forever.
The full product (the same document index exposed to every AI client over MCP), free forever, no card. Hosting your projects starts at $5.99/mo. Download for Windows.
Common questions
Straight answers. Before you install.
Can I really search inside PDFs and Word documents, not just their names?
Yes. Interlinked decodes the text of every PDF, Word, Excel and PowerPoint file on the drive and indexes the words, so a phrase that only appears on page 7 lands on that file, whatever the scanner or camera named it. And because the index tracks changes, a document you just saved is searchable about 1 ms later.
Why can't I just run grep or ripgrep on a folder of PDFs?
Because a PDF or .docx stores its text compressed and encoded: a byte scanner reads the raw bytes, sees noise, and returns no match. You have to decode the document's text first, then index it. For plain-text files ripgrep is the fast baseline, and Interlinked still beats it 93.8 s → 16 ms on one query, 58,625× faster.
Doesn't Windows Search already do this?
Only for part of the drive. Windows Search indexes a subset of folders and throttles the rest, so a content query outside that slice walks the disk or comes back empty: a median near 67 s on a 4.47M-file machine, roughly 480,000× slower than covering the whole drive in 7 to 9 ms. Spotlight is the right model, but it's macOS-only.
Can my AI agents search inside documents too?
Yes: the same decoded-text index is exposed to any MCP client (Claude Code, Cursor, Codex, Cline and more). A document hunt that took Claude Code 6m57s and 71 tool calls comes back in 16 ms and a single call (about 7,200,000× faster) and it drops the context it burned from ~58% to nearly zero, all from about 44 MB of memory at rest.
What does it cost?
Everything (the app and the same document index handed to every AI client over MCP) is free forever, no card. Hosting starts at $5.99/mo.