The context window is a hard budget, and legal work spends it fast: a single contract runs tens of thousands of words, a matter is many contracts, and the teams using AI for legal and financial documents process them at staggering volume — a Harvey Vault upload or connector sync can drop millions of documents into the system, and Hebbia, serving finance and legal teams, pushes 250 billion tokens a month. Compression multiplies that budget: shrink every document 4× without losing the relevant info, and whatever window your model has now covers 4× more source material — more documents per call, more agent tool calls before the context overflow forces compaction.
The multiplier is worth more than it looks, because the actual window you command is smaller than what the model release claims. Chroma ran 18 models through deliberately simple long-input tasks and found that performance "grows increasingly unreliable as input length grows" (Context Rot); when the question doesn't share words with the evidence — the normal case in contract QA — 11 of 13 models drop below half of their short-context accuracy when reaching just 32K tokens (NoLiMa, ICML 2025). Anthropic recommended treating context as a finite resource "attention budget". Legal use-cases are among ones affected the most: on multi-document case files exceeding 100K tokens, frontier models omit key facts more often than they hallucinate, and get worse as cases get longer (Gavel). Teams building legal-document analysis report replacing whole-corpus prompting precisely because "long-context degradation scales with it" (Inject or Navigate?).
The catch sits in "without losing the relevant info." Legal text does not forgive careless cuts, and the harder you compress, the more it matters which parts survive
In this report, we present benchmarking results of our new legal data compression model, latte-v2-lex, and compare it against the state-of-the-art general-purpose latte-v2.
Setup
Two readers — gpt-5.4-mini and gemini-3.6-flash — answer the same questions over the same documents, with the context compressed at four rates (~1.3×, ~2×, ~3×, ~4×) by two checkpoints: the general-purpose latte-v2 and the legal specialist latte-v2-lex. The readers' temperature is fixed at 0. We consider three benches, each shown as a side-by-side pair of panels: ContractNLI (n=300, NDA entailment), CUAD (n=300, contract clause retrieval), FinanceBench (n=142, 10-K / 10-Q filings QA).
ContractNLI and CUAD measure the model's legal performance. We include FinanceBench to make sure the specialized chekpoint does not degrade outside of the domain of interest.
Each compressed run is compared against the same reader's uncompressed run with exact McNemar (per-sample comparison). Dots on the charts are accuracy; whiskers are the 95% CI on the paired difference. * marks p < 0.05.
Results
latte-v2
n=300 · dot = accuracy, whiskers = 95% CI paired vs dashed anchor · * = p<0.05
latte-v2-lex
same construction
latte-v2
n=300 · dot = accuracy, whiskers = 95% CI paired vs dashed anchor · * = p<0.05
latte-v2-lex
same construction
latte-v2
n=142 · dot = accuracy, whiskers = 95% CI paired vs dashed anchor · * = p<0.05
latte-v2-lex
same construction
On FinanceBench and ContractNLI, none of the compression runs show a significant difference with the uncompressed baselines, for either checkpoint or reader — up to ~4× compression, performance holds within noise. CUAD is different: v2 degrades with rate, dropping significantly below uncompressed from 3× under gemini (−5.7pp, p=.027) and at ~4× under both readers (mini −6.7pp, p<.001; gemini −6.3pp, p=.009). latte-v2-lex flattens that slope: its only significant drop is under mini at ~4× (−4.3pp, p=.047), and under gemini it holds within noise at every rate.
These results make sense. ContractNLI asks for a document-level entailment verdict, and FinanceBench answers depend on a handful of figures — in both, inputs are highly redundant, so a compressor can drop most of them without sacrificing the actual signal. CUAD is different: the reader has to pinpoint one specific clause, and every word of that clause the compressor deletes is fatal. That is exactly the regime where knowing which legal language matters starts to matter a lot. This reflects in the results: at up to 2×, compression is essentially free for both checkpoints — past that, clause-retrieval work is where generic compression is not enough — and where latte-v2-lex stretches the safe range further, enabling 3× or 4× compression depending on the reader. Remarkably, specialized latte-v2-lex maintains up to 4× lossless compression on the "out-of-distribution" FinanceBench.
Below is one contract end to end, exactly as latte-v2-lex compressed it — what a 4× cut keeps, and what it throws away.
latte-v2-lex is available upon request!
Everything above was measured on public benchmarks, but what matter most is the actual performance on your data. If you want to give your legal agents more reliable context at any given time - please reach out, and we will show how much additional context Compresr unlocks on your data!