AI-powered Legal Summaries

LexiBrief: Legal Summarizer

Quickly transform complex legal documents* into insightful summaries using advanced NLP models. Choose from a variety of models to suit your needs, and get started in seconds.

* Currently optimized for English legislative and statutory text — bills, statutes, regulations and contracts. The models are fine-tuned on US Congressional bills (BillSum), so judicial opinions and case law fall outside their training domain and summaries of those are unreliable. Working to support multilingual documents soon!

If you want to learn more about how this website came to be, read the paper.

Why LexiBrief

Instant Summaries

Get concise summaries of complex legal documents in seconds.

Advanced AI Models

Choose from state-of-the-art NLP models tailored for legal text.

Privacy First

Your documents are processed securely and never stored.

How It Works

1

Upload

Upload or paste your legal document.

2

AI Analysis

Our model analyzes key clauses and terms.

3

Get Summary

Receive a concise summary instantly.

4

Export & Share

Download or share your summary easily.

Backed by Research

LexiBrief is the applied companion to a master's thesis on transformer-based legal text summarization — every model choice here was benchmarked, not guessed.

4

Specialized models compared, from 1k to 16k token context windows

2

Independent AI judges (Claude & Gemini) scoring summaries blind, A/B/C/D

~90%

Run-to-run scoring consistency per judge, across repeated evaluation passes

3

Evaluation lenses: lexical overlap (ROUGE), semantic yield (BLANC), human + AI judgment

Read the full thesis for methodology and results →

A smoother workflow, end to end

Scroll or drag through the highlights, then jump straight into the summarizer.

Drag to explore

Secure processing
Secure
Privacy-aware processing

Summaries are generated on demand. No document retention in the UI-only preview.

Gavel representing legal documents
Legal
Built for long legal documents

Pick a model based on length. Longer-context models unlock larger agreements and case files.

Fast insights
Fast
From text to clarity

Paste text or upload a file—then generate a concise summary with one click.

Continue to summarizer

Implementation Notes

The four models here are the ones the thesis evaluates, summarising to the same 256 tokens. What surrounds them differs in three ways, because this site takes documents uploaded by visitors and runs on a free CPU host.

It reads real documents

The thesis works from clean dataset text. Here, PDFs are parsed on the server and repaired first — layout padding removed, and words the PDF split apart put back together.

It fits a small server

Beam search narrows on very long inputs and generation is time-capped. Documents past a model’s window are summarised in sections and condensed, rather than cut off.

It tidies the text

Cleanup runs on the input before summarising, where the thesis runs it on the summaries when scoring. Stray markup is stripped and summaries end on a complete sentence.

See every difference, step by step
Step Thesis pipeline This site Why it differs
Input BillSum dataset text, already plain and clean PDF, DOCX or TXT uploaded by the visitor and parsed server-side Real documents arrive as files, not as dataset rows
Extraction repair Not applicable Collapses padding whitespace and dot leaders; converts single line breaks to spaces; rejoins words the PDF text layer split apart (“ackn owledged” → “acknowledged”) when the joined form is a dictionary word PDF text layers position glyphs, not words, so extraction splits and pads text
Text normalisation
Roman numerals, SymSpell, acronym expansion		 Applied to the generated summaries, when computing evaluation metrics Applied to the input before summarising. SymSpell is also restricted to lowercase alphabetic words of four or more characters, and keeps punctuation Uploaded text is noisy where benchmark text is not. The restriction stops party names, citations and statute numbers being rewritten
Beam search 4 beams for every document 4 beams up to 2,048 input tokens, 2 up to 8,192, greedy beyond that Four beams over a 16k-token input exceeds the memory of the free-tier CPU host
Documents longer than the context window Truncated to the window Summarised in up to four windows, then condensed into one summary Truncating a long contract silently discards most of it
Generation limits Maximum 256 tokens, early stopping The same, plus a 64-token minimum, no repeated 3-grams, and a 150-second cap Guards against degenerate repetition and requests that never return
Output cleanup None Strips HTML entities and trims the summary back to its last complete sentence BillSum’s bill text is scraped HTML, so the models learned to emit entities such as  . The length cap can also stop generation mid-clause

None of the steps above were involved in producing the results reported in the thesis. Those numbers come from the unmodified training and evaluation scripts in the repository (bart_base.py, led_base.py, longt5_base.py), published exactly as they were run.

Your Legal Document

Select AI Model

Pick a model based on document length and desired quality.

Tip: Longer models handle longer documents.

Models grey out when they can’t handle the token count.

Document Tokens: 0
Status: Ready
Character count: 0/64,000
Drop your file here or click to browse

Supports PDF, DOCX, and TXT files (Max 16MB)

Generating summary... This may take a moment.

Processing Progress
0%

Initializing...

Summary Results
Actions

Disclaimer

Not a substitute for professional legal advice. Summaries may omit critical details. Always consult a qualified attorney for legal matters.