Mean Average Precision (MAP)

A ranking evaluation metric that measures the usefulness of retrieved documents based on their graded relevance and position in the result list. Unlike binary relevance metrics, NDCG assigns higher weights to relevant documents appearing at top ranks and penalizes relevant documents ranked lower. The metric normalizes against an ideal ranking, producing scores from 0 to 1. In legal retrieval, NDCG is particularly valuable when documents have multi-level relevance judgments—such as directly on-point precedent versus tangentially related case law—rather than simple relevant/irrelevant binary labels.

An algorithm that combines ranked result lists from multiple retrieval systems by computing a reciprocal rank score for each document across all input rankings. The formula sums 1/(k + rank) for each document, where k is a constant (typically 60) that mitigates the impact of high rankings from outlier systems. RRF effectively merges sparse and dense search results without requiring score calibration between heterogeneous retrievers. In legal search pipelines, RRF is commonly used to fuse BM25 lexical results with dense embedding similarity scores, producing a unified ranking that captures both exact statutory citations and conceptual legal relevance.

A two-stage retrieval refinement model that jointly encodes a query and candidate document through full cross-attention to compute a fine-grained relevance score. Unlike bi-encoders that encode queries and documents independently, cross-encoders process the concatenated pair simultaneously, enabling nuanced semantic comparison. This architecture is typically deployed as a second-pass reranker over top-k candidates from a faster first-pass retriever. In legal applications, cross-encoders excel at distinguishing between superficially similar but legally distinct documents—such as cases that share factual patterns but reach opposing holdings—significantly improving precision at the cost of higher inference latency.

A training data curation strategy that identifies documents which are superficially similar to a query but not relevant, improving the discriminative power of embedding models. Hard negatives are challenging distractors that share vocabulary or topical overlap with relevant documents yet fail to satisfy the information need. Training with hard negatives forces the model to learn fine-grained semantic distinctions rather than relying on shallow lexical cues. In legal retrieval, hard negatives might include cases from the same jurisdiction discussing the same statute but reaching different conclusions, or contracts with similar clause headings but materially different obligations.

A technique that augments a user's original search query with related terms or synonyms to improve recall in retrieval systems. Expansion can leverage legal thesauri, statutory definitions, or generated text from language models to capture broader relevant documents. Common approaches include:

• Lexical expansion: Adding synonyms and morphological variants from domain vocabularies
• Pseudo-relevance feedback: Extracting terms from top-ranked documents in an initial retrieval pass
• Generative expansion: Using LLMs to produce alternative phrasings or hypothetical relevant passages In legal search, query expansion helps bridge the vocabulary gap between layperson queries and formal legal terminology.

A data augmentation method that uses a language model to generate plausible queries for unlabeled legal documents, creating training pairs for fine-tuning dense retrieval models. The process typically involves prompting an LLM to produce questions or search queries that a given document would answer. This technique addresses the scarcity of labeled query-document pairs in specialized legal domains where human annotation is prohibitively expensive. Generated queries can cover diverse information needs—from statutory interpretation questions to fact-pattern analogies—enabling robust training of embedding models without manual relevance judgments.