Inject or Navigate? Token-Efficient Retrieval for LLM Analysis of Transactional Legal Documents

Jul 16, 2026

Answering questions over transactional legal documents via repeated full-corpus injection into an LLM context maximizes recall but scales token usage with the corpus and suffers long-context degradation. The study replaces injection within a legal-document analysis system by evaluating structured retrieval against two modes over structure-aware chunking: embedding retrieval (navembed) and LLM navigation over a compact structured index (navindex). On a 20-question benchmark, semantic retrieval with reranking matches injection on 16/18 cases while using far fewer tokens and costs; navindex also reduces total token footprint, answering context size, and dollar cost. A caching crossover rule quantifies when cached injection remains cheaper and where uncertainty arises.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

Introduction

Baseline: full-corpus injection

Costs and retrieval challenges in legal domain

Structured retrieval setup and controlled chunking

Comparison 1: NavigateEmbeddings (navembed)

Comparison 2: Navigate-Index (navindex)

arXiv :2607 .05764v 1 [ cs .CL] 7 Jul 2026 Inject or Navigate? Token-Efficient Retrieval for LLM Analysis of Transactional Legal Documents Mahmoud Hany* Mourad ElSheraey* Mahmoud Said* Peter Naoum [hany@syntheia.app](hany@syntheia.app) [mourad@syntheia.io](mourad@syntheia.io) [mahmoud@syntheia.app](mahmoud@syntheia.app) [peter@syntheia.app](peter@syntheia.app) Syntheia Pty Ltd Abstract. Answering questions over a set of transactional legal documents is most simply done by injecting the whole corpus into the LLM’s context window on every query. That baseline maximises retrieval recall, but its token footprint scales with the corpus rather than the question, and long-context degradation scales with it. We report what it took to replace full-corpus injection in a legal-document analysis system, comparing it against two structured retrieval modes over our proprietary structure-aware chunking: embedding retrieval (navembed) and LLM navigation over a compact structured index (navindex) . On a 20-question benchmark with verified ground-truth answers, a position-bias-controlled, reference-anchored pairwise judge scored semantic retrieval with reranking tied with injection on 16 of 18 document-bound questions (injection preferred on 2) while attending to 17.3 × fewer input tokens (a general-text-embedding (GTE) configuration reaches 29.9 × at a lower tie rate); both modes were judged tied on the 2 out-of-scope controls. navindex was judged tied on all 18 at a 1.61 × smaller total token footprint, a ∼56 × smaller answering context, and 25% lower dollar cost. We derive a closed-form caching-crossover rule: cached injection is cheaper in dollars only while the corpus stays below roughly ten times the retrieval payload. Scope and uncertainty are quantified in Section 8 . 1 Introduction Large language models have made it practical to interrogate documents directly: a user poses a question in natural language and receives an answer grounded in the source text, with no hand-built query interface and no manual review. Few professions stand to benefit more from this than law, whose day-to-day work consists largely of locating, reconciling, and interpreting provisions spread across long, densely cross-referenced instruments. Syntheia provides the data layer for this workload: a structured document index over coordinated transaction sets (an original agreement plus its amendments, side letters, schedules, and disclosure documents) served through an API, and directly to LLMs such as Claude via the Model Context Protocol [1], for document tools to build on. This paper reports the evaluation that decided how retrieval over that index should work: what it took to replace full-corpus injection with structured retrieval without giving up answer quality, and what each alternative costs in tokens and dollars. The simplest method is to place the entire corpus in the model’s context window on every query. We call this approach inject, and we adopt it as our baseline deliberately, because it represents the upper bound on retrieval recall. Every clause, defined term, and crossreference is in context on every query, so there is nothing ∗ Equal contribution. Mahmoud Hany built and ran the evaluation program across all three phases and wrote the paper’s initial draft. Mourad ElSheraey revised the analysis, added the uncertainty quantification and threats-to-validity treatment, and prepared the final manuscript. Mahmoud Said implemented the crossencoder reranking pipeline and metadata fields, and contributed the cross-reference extraction. Peter Naoum implemented the initial codebase, and supervised the experimental work. for a retriever to miss and nothing to misconfigure. An alternative that a careful judge cannot distinguish from inject is therefore a stronger result than one that merely outperforms a lossy retriever; the corpus was deliberately sized to keep inject runnable (Section 3) . Two costs limit inject as the corpus grows. The t

FAQs

为什么“注入整套语料（inject）”在法律文档问答中会产生高成本？

inject会在每次查询都把整个语料放入上下文，token占用随语料规模增长，并且会受到长上下文信息衰减影响，同时法律文本的相关条款往往分散在多处。

文档研究评估了哪些结构化检索替代inject的方法？

研究在相同的结构感知分块表示下比较两种结构化检索：navembed（基于嵌入相似度检索节点）以及navindex（在紧凑结构索引上进行LLM引导式导航）。

在基准评测中，结构化检索相对inject的质量与资源消耗如何？

在20个问题的评测中，受控的语义检索与重排在16/18个“文档相关”问题上与inject打平，且输入token显著更少；navindex在所有18个问题上保持打平，同时进一步降低总token占用、回答上下文规模与成本。

研究如何判断何时缓存的inject在金钱上更划算？

论文给出一个封闭形式的“缓存-交叉”规则：当语料规模低于大约为检索负载的十倍时，缓存注入在美元成本上更便宜，并且在相应部分量化了范围与不确定性。

Jul 16, 2026

Upload by Theodore