Research Publication: 25% Cost Saving for AI Document Review
FOR IMMEDIATE RELEASE
Syntheia Research Shows Structured Retrieval Can Match Full Document AI Review at a Fraction of the Cost
New York City, July 2026
Syntheia today published new research addressing one of the most consequential practical problems in legal AI today: the rising token cost of feeding transactional documents into a large language model's (LLM) context window for question answering.
Most efforts to cut AI costs so far have focused on making the model itself cheaper: smaller prompts, routing to cheaper models, prompt caching. Syntheia's latest research targets a different layer of the problem: the context that goes into the model in the first place. In transactional legal work, the cost of an LLM reading a document is typically far larger than the cost of it reasoning over the answer. In our tests, the length of the final answer barely varied between methods, while the amount of text fed into an LLM varied by up to 30×. That gap is where we can find significant savings. As AI agents increasingly decide for themselves what to retrieve and read, that matters even more.
Syntheia's research team tested two structured retrieval methodologies for transactional legal text, both built on our structure-aware document indexing technology, against full document injection on a 20-question benchmark spanning real credit facility agreements, limited partnership agreements, and share purchase agreements.
Headline results:
Semantic (embedding-based) retrieval, which fetches only the passages most relevant to a question, matched the performance of full document injection on 18 of 20 benchmark questions while cutting tokens processed by 17.3×. A faster, lighter-weight embedding configuration pushed the token reduction further, to nearly 30×, with a modest tradeoff, matching full injection on 15 of 20 questions.
Structured index navigation, a novel format Syntheia developed for LLMs to navigate in order to decide what to retrieve, reasoning over a compact map of a document's clauses rather than the full text, matched full document injection performance on all 20 benchmark questions, while cutting the context the model reads to compose each answer by roughly 56× and total tokens processed by 1.6×.
Both approaches rely on the same underlying capability: Syntheia's document index automatically resolves the cross-references and defined terms in a clause. Critically, the index is used only for navigation, not for the question answering.
"Token economics was always going to become a bottleneck for legal AI, especially as agentic workflows read the same documents over and over," said Horace Wu, CEO of Syntheia. "Most of the cost for legal isn't the model thinking, it's the model reading. We built a retrieval layer specifically to cut the reading cost. The results held up well enough that we have built it into a product. Several law firms and legaltech companies have already begun piloting the new Syntheia Query API, and we are planning to open it up for self-service in the near future."
The full paper, benchmark questions, and evaluation code have been released publicly on GitHub so the results can be independently reviewed and reproduced.
Read the research: https://arxiv.org/abs/2607.05764
Code and test materials: https://github.com/asksyntheia/inject-or-navigate
Media enquiries: hello@syntheia.io
About Syntheia. Syntheia builds precision instruments for transactional lawyers, serving law firms and in-house legal teams who need certainty in their work. Its suite of tools — Compare, Curate, and Query — helps lawyers detect what changed across document versions, manage obligations in fund side letters, and retrieve exact clause and subclause text from large transaction sets, all grounded in a structural, source-traceable understanding of how documents are written.
