filehugger.comSemantic Document Compare
Compare two documents locally, align similar paragraphs, isolate additions/removals and list changed numeric tokens.
Analyse your own file
Choose the original document
Drop it here or browse this device
Choose the revised document
Drop it here or browse this device
Your original file is analysed in this browser and is not uploaded. An optional AI second-opinion button appears only after the local report; if you choose it, extracted text—not the original file—is sent to our server.
What happens to the file
- Default processing
- Browser-only JavaScript, PDF.js, pdf-lib and Tesseract where the task needs them
- Accepted input
- PDF and text-based documents
- Report contract
- File Intelligence JSON schema 1.0 plus a flattened CSV view
- Optional AI
- Explicit second opinion; extracted text only; original file is not uploaded
Synthetic service terms change notice period from 30 to 45 days, price from EUR 120 to EUR 135 and add a backup clause.
How the result is calculated
Text is extracted, split into paragraphs, represented as term-frequency vectors and aligned greedily above a cosine-similarity threshold. Exact numeric tokens are compared separately so a small price/date change is not hidden by high overall similarity.
Overall similarity is cosine(term-frequency vectors) × 100. Paragraphs align at ≥42%; >98.5% is unchanged. These thresholds are published and versioned, not learned from user files.
Supported and unsupported cases
| Input or situation | What version 1.0.0 checks | What it does not claim |
|---|---|---|
| Text/selectable PDF | Paragraph additions, removals, similarity and numeric-token changes | Pixel/layout differences |
| Reordered similar clauses | Best unmatched paragraph alignment | Reliable legal equivalence |
| Tables represented as text | Some row/number changes | Merged cells or spatial relationships |
| AI summary | Optional narrative of deterministic evidence | Substitution for legal review |
What the confidence score means
Confidence describes how much of the file the published parser could inspect and how specific the evidence was. It is deliberately separate from the result score. A 95/100 confidence value does not mean “95% likely to be true,” and a high health/readiness score does not certify safety, compliance or factual correctness.
Comparison with other approaches
| Approach | Useful strength | Important limitation |
|---|---|---|
| Local semantic compare | Local paragraph alignment plus numeric changes | Not a legal redline |
| Line diff | Exact and deterministic | Noisy when paragraphs wrap or move |
| Visual PDF compare | Finds layout/pixel changes | Can miss meaning or treat reflow as change |
| Word processor track changes | Best when edit history exists | Requires compatible editable sources |
5/5 fixtures passed (100%)
Dataset: FH-INT-SYNTH-2026-07. Scope: Aligned change, addition, removal, numeric-token change and identical-control fixtures. Fixtures contain generated data only; no visitor file or personal record is collected.
When this result can be wrong
Boilerplate can inflate similarity, word order is weakly represented, a negation may look deceptively similar, and extracted PDF reading order can be wrong. Review every changed paragraph and number.
Keep the original file, test the output in the receiving application and escalate high-stakes legal, financial, medical, accessibility or security decisions to a qualified reviewer.
Version and accuracy change log
| Release | Date | Calculation/coverage change | Re-test requirement |
|---|---|---|---|
| 1.0.0 | 2026-07-29 | Initial public engine, report schema, thresholds, synthetic fixtures and sample files. | All published fixtures and one browser sample per tool. |
Downloadable evidence, not a black-box verdict
After analysis, use Download JSON report for structured evidence or Download CSV report for review in a spreadsheet. Both include engine version, timestamp, score, confidence, metrics, findings, recommendations and limitations. Sensitive-data reports mask candidate values and no report includes the complete extracted document text.
Related tools and independent checks
Method or accuracy correction? Read the testing methodology and contact info@brothers.net.tr. Material scoring changes are published in the change log.