filehugger.comAI methodology
This page exists so the word “AI” on this site can be checked rather than taken on faith. Feature by feature: is it deterministic code, a model on your device, or something on our server — and what leaves your machine in each case.
Deterministic code, in your browser
No model is involved at all. Same input, same output, every time — these paths are unit-tested for exactly that.
- File Doctor diagnosisByte and structure parsing
- File inspection card (all tools)Header parsing
- Settings AdvisorPure rules, unit-tested for identical output on identical input
- Quality Scorecard, compare, SSIM/PSNRMeasurement in the page
- File Intelligence baselineParsing, extraction and scoring
A model, but running in your browser
The model is downloaded to your machine and runs there. The file stays in the tab.
- OCR (image and PDF)Tesseract, fetched on first use and then cached
Server software — not AI
These need a full engine that cannot run in a browser. No model, no training, no retention beyond the job.
- Office / HTML to PDF, PDF/ALibreOffice and headless Chromium; held in memory only
- Video / audio conversionNative FFmpeg with ffprobe validation and fixed profiles, in an isolated job directory that is removed automatically
A server model, and only after you click
Appears as a separate button after the local report already exists. The deterministic result is produced first and stands on its own.
- Optional Intelligence second opinionOpenAI Responses API via our server, capped and rate-limited, store: false
The rule we build against
Our architecture decision record is blunt about it: the deterministic parser is the source of truth, and a model may only explain evidence the parser already produced — never introduce a fact. Three things follow, and all three are enforced in code rather than promised:
- No evidence, no claim. Every finding carries a reference to the byte range, inspection field or page it came from. A finding without one is rejected by the contract validator before it can render — you see "report withheld", not a plausible sentence.
- Unprovable absolutes are blocked. The validator refuses phrasing like "virus-free", "100% safe", "fully repaired" or "certified", because a parser cannot establish them.
- Low confidence is not a verdict. A finding below 50% confidence must carry an explicit limitation instead of a conclusion, and "cannot determine" is a first-class result.
Your file as data, never as instructions
Text inside a document is untrusted input. It is never placed where instructions go, and it is never rendered as HTML — the report is built as text nodes only. A document that contains "ignore previous instructions" is just a document containing that sentence, and it is part of our fixture set.
How the optional server model is constrained
Selected File Intelligence tools expose a second-opinion button only after the local report exists. The request sends extracted text—not the original file—is capped at 256 KB, is rate-limited, keeps credentials on the server and sets store: false. The model is instructed to preserve uncertainty, abstain when the supplied text does not support an answer and never replace the deterministic findings. Provider retention can still depend on account configuration and current terms, so this feature is optional and unsuitable when third-party processing is prohibited.
How to check us
Open your browser's network tab and run the File Doctor or a File Intelligence local report: you will see no document upload. Pressing the separately labelled AI button creates the documented extracted-text request. JSON exports contain the same evidence references shown on screen. Our testing methodology covers the unit and fixture suites, and security & file privacy covers per-tool processing.