Enhanced Similar Case Retrieval Methods

• Similar case retrieval enhancement is the set of methods that transform raw cases into structured abstractions by decomposing similarity for robust and interpretable retrieval across domains.
• It improves retrieval by employing domain-specific reformulations such as legal facts, masked SQL templates, or high-attention image patches, ensuring precision in retrieval.
• The approach decomposes similarity into distinct subproblems and employs evidence fusion techniques, enhancing performance and interpretability in legal, healthcare, and pathology applications.

Searching arXiv for the cited and closely related papers to ground the article. arXiv search query: similar case retrieval legal case retrieval case-based reasoning retrieval enhancement Similar case retrieval enhancement denotes the family of methods that improve retrieval of prior cases, precedents, examples, or analogous records by changing what is represented as a “case,” what counts as similarity, and how ranking evidence is combined. Across recent work, the dominant shift is away from raw whole-text or raw example nearest-neighbor matching toward domain-specific reformulations such as legal facts and legal issues, crime-specific sub-facts, masked SQL templates, and pathology-focused image regions, followed by retrieval mechanisms that separately model structure, entities, statutes, dispute focus, or salient visual evidence (Tang et al., 2023, Nguyen et al., 5 Mar 2026, Hashimoto et al., 2021, Deng et al., 2024).

1. Scope and conceptual foundations

Similar-case retrieval arises in several technically distinct settings. In legal retrieval, the query is usually a long judgment or fact description, and relevance is defined by precedent utility rather than generic semantic resemblance. In healthcare text-to-SQL, the “case” is a prior question-SQL pair, and the relevant analogue must preserve SQL logic while also grounding the right entities and schema. In histopathology, the case is a whole slide image together with associated diagnostic metadata, and similarity must reflect clinically meaningful morphology rather than background tissue. Despite these differences, the same failure mode recurs: direct retrieval over raw inputs often conflates salient and non-salient information, or forces one similarity function to solve multiple incompatible subproblems at once (Tang et al., 2023, Nguyen et al., 5 Mar 2026, Hashimoto et al., 2021).

A particularly explicit diagnosis appears in legal query reformulation. On LeCaRD, the average query length was 495.5 characters with stop words, while the average human annotation of salient content was only 72.14 characters, yielding an average compression rate of 19.06% with stop words and 14.52% without. This quantifies the claim that long legal case queries are dominated by noise relative to the legally decisive content (Zhou et al., 2023). A parallel argument appears in healthcare text-to-SQL, where single-step retrieval over raw question-SQL pairs is described as unstable because the retriever must simultaneously recover the correct logical SQL pattern and the correct medical entities and schema grounding (Nguyen et al., 5 Mar 2026).

Domain	Retrieval enhancement mechanism	Representative papers
Legal case retrieval	Query/case reformulation into legal facts, legal issues, sub-facts, or legal elements	(Tang et al., 2023, Deng et al., 2024, Li et al., 2023, Zhou et al., 2023)
Healthcare text-to-SQL	Structure-first case retrieval plus separate entity grounding	(Nguyen et al., 5 Mar 2026)
Histopathology	High-attention region selection plus metric learning aligned with IHC similarity	(Hashimoto et al., 2021)
Legal QA with RAG	Structured case fields and hybrid similarity over question, support, and entities	(Wiratunga et al., 2024)

2. Reformulating the case before retrieval

A central line of work improves retrieval by reformulating the input case into a representation that suppresses incidental variation and preserves legally or clinically decisive structure. In legal retrieval, PromptCase extracts legal facts and legal issues, prepends the prompts “Legal facts:” and “Legal issues:”, encodes them separately and jointly, and concatenates the resulting embeddings as the final case representation. This directly targets the two problems it names legal feature alignment and legal context preservation, replacing whole-case encoding with a compact representation of determinative legal content (Tang et al., 2023). KELLER makes the same move at a finer granularity: it reformulates a case into a set of crime-specific sub-facts, where each sub-fact has a crime as title and a concise fact summary as body, produced from prompts conditioned on the crime and its relevant law articles. The reformulated query becomes $q=\{q_1,\ldots,q_m\}$, and the candidate case becomes $d=\{d_1,\ldots,d_n\}$, so retrieval operates over structured sub-fact sets rather than monolithic long texts (Deng et al., 2024).

In healthcare text-to-SQL, CBR-to-SQL performs an analogous abstraction over question-SQL pairs. Its offline Case Retain stage uses entity masking to replace concrete mentions with broad categories such as CONDITION, MEASUREMENT, PROCEDURE, DRUG, EQUIPMENT, CODE, NAME, and TIME in both the natural-language question and the SQL query. The indexed object is therefore a masked natural-language template rather than a raw domain-specific utterance, and dense embeddings of these masked questions are stored in Qdrant (Nguyen et al., 5 Mar 2026). In legal SCR benchmarking, MUSER operationalizes reformulation as a multi-view legal-element representation. It annotates sentence-level legal elements for legal fact, dispute focus, and law statutory, with 22 first-level labels, 190 second-level labels, and 505 third-level labels overall, and represents a case at a given level by a binary label vector $\mathcal{L}=(l_1,\dots,l_n)$ for view-wise cosine comparison (Li et al., 2023).

Reformulation is not limited to text. In similar image retrieval for malignant lymphoma, attention-based multiple instance learning is used to identify high-attention patches, and only the top $M=10\%$ of patches in each bag are retained as high-attention patches for subsequent metric learning and retrieval. This changes the retrieval unit from all sampled tissue patches to pathologically important regions that are more likely to reflect tumor morphology (Hashimoto et al., 2021). Across these systems, enhancement begins by deciding that the retrievable object should be a legally or clinically meaningful abstraction, not the raw artifact.

3. Decomposing similarity into distinct subproblems

A second major theme is the decomposition of similarity itself. CBR-to-SQL is the clearest statement of this principle. It argues that standard RAG retrieval implicitly tries to satisfy two different requirements at once: finding examples with the right logical SQL pattern and finding examples with the right entities and schema grounding. Its solution is a two-stage CBR-style pipeline: Template Construction retrieves structurally similar masked cases with top-$k=5$ dense nearest neighbors by cosine similarity, then Source Discovery performs entity retrieval directly against a lookup table built from MIMIC-III using a medical embedding model, top-100 semantic retrieval, Levenshtein reranking, and top-5 candidate retention per entity (Nguyen et al., 5 Mar 2026).

Legal retrieval papers decompose similarity in related ways. PromptCase separates facts and issues, encodes them in dual and cross views, and shows that legal relevance is better aligned with these facets than with full-text similarity (Tang et al., 2023). CBR-RAG defines a structured legal case $c=\langle Q,S,E,A\rangle$ and compares queries through three channels: question-question similarity, query-support similarity, and query-entity similarity. Its hybrid retrieval score weights these channels as $[0.25, 0.40, 0.35]$, reflecting the claim that evidence-bearing support text and entities should not be collapsed into a single undifferentiated embedding space (Wiratunga et al., 2024). LCM-LAI pushes this farther by distinguishing semantic interaction from legal-rational interaction. Its legal interaction module computes a sentence-level law distribution vector and defines legal correlation as $c_{i,j}^{(L)}=\cos(\lambda_i^X,\lambda_j^Y)$, thereby comparing sentences in law-distribution space rather than only in embedding space (Xu et al., 25 Feb 2025).

Contrastive legal retrieval models make the same distinction through supervision rather than architecture. MVCL introduces case-view contrastive learning, which pulls globally relevant cases together, and element-view contrastive learning, which preserves legally decisive elements while perturbing non-element sentences. The latter is explicitly motivated by the fact that cases may be semantically similar yet legally different because of one decisive legal factor, such as an interest rate that turns a private lending dispute into usury (Wang, 2022). Caseformer similarly constructs two unsupervised contrastive tasks from legal structure: Legal Judgment Prediction, which discriminates between factually similar cases using crimes and legal provisions, and Factual Description Matching, which discriminates between judgment-similar cases using factual proximity (Su et al., 2023).

This decomposition thesis is cross-domain. It suggests that similar-case retrieval improves when the system is allowed to use different representations, similarity spaces, and evidence sources for different aspects of relevance, rather than forcing one embedding and one nearest-neighbor operation to stand in for all of them.

4. Ranking architectures and evidence fusion

Once cases are reformulated and similarity is decomposed, ranking quality depends on how local evidence is fused. KELLER uses a late-interaction scheme over sub-facts. After encoding each query sub-fact $q_i$ and document sub-fact $d_j$, it builds a similarity matrix $d=\{d_1,\ldots,d_n\}$0 and computes the final case relevance score as

$d=\{d_1,\ldots,d_n\}$1

This MaxSim + Sum design enforces the idea that each query sub-fact should find its best counterpart in the candidate case, and it yields a directly inspectable alignment structure for interpretability (Deng et al., 2024).

Histopathology retrieval uses an analogous partial-matching score. For a query case $d=\{d_1,\ldots,d_n\}$2 and a candidate case $d=\{d_1,\ldots,d_n\}$3, the case distance is

$d=\{d_1,\ldots,d_n\}$4

so each high-attention query patch finds its nearest high-attention counterpart in the candidate case. This yields case-level retrieval together with matched patch pairs that expose the local morphological correspondences behind the ranking (Hashimoto et al., 2021).

Several papers revisit the final ranking layer itself. In legal SCR, replacing the usual fully connected classifier with a linear RankSVM over extracted $d=\{d_1,\ldots,d_n\}$5 features improves top-ranked ordering on LeCaRDv1 and for Lawformer on LeCaRDv2, with the explicit motivation that SCR is fundamentally a ranking problem rather than a pointwise binary classification problem (Liu et al., 16 Feb 2025). Graph-based approaches attempt to inject higher-order structure into case representations before cosine ranking. CaseLink models cases and charges as nodes, but a reproducibility study found a nuanced outcome: graph augmentation can help, yet it is not uniformly superior to strong text baselines, and a more sophisticated heterogeneous graph representation with HGT did not improve over the original homogeneous graph (Donabauer et al., 11 Apr 2025). Law-Match offers a different fusion principle by decomposing a case embedding into a law-related fitted part and a law-unrelated residual, then reconstructing a representation that balances statutory mediation and direct factual effect before pairwise matching (Sun et al., 2022).

These systems differ in modeling assumptions, but they share a common architectural lesson: effective similar-case retrieval rarely emerges from a single monolithic score. It emerges from a controlled aggregation of local matches, structurally typed signals, or pairwise preferences.

5. Evaluation paradigms and empirical evidence

Recent work evaluates enhancement not only by top-line accuracy but also by robustness, sample efficiency, clause-level or sentence-level behavior, and human preference. In healthcare text-to-SQL, CBR-to-SQL achieved $d=\{d_1,\ldots,d_n\}$6 and $d=\{d_1,\ldots,d_n\}$7 in the complete MIMICSQL training environment, compared with RAG-to-SQL’s $d=\{d_1,\ldots,d_n\}$8 and $d=\{d_1,\ldots,d_n\}$9. Under the sparse Incomplete Database setting, RAG-to-SQL fell to $\mathcal{L}=(l_1,\dots,l_n)$0, whereas CBR-to-SQL reached $\mathcal{L}=(l_1,\dots,l_n)$1. Its brittleness also remained lower: in CDB, RAG-to-SQL had $\mathcal{L}=(l_1,\dots,l_n)$2 and $\mathcal{L}=(l_1,\dots,l_n)$3, while CBR-to-SQL dropped only $\mathcal{L}=(l_1,\dots,l_n)$4 and $\mathcal{L}=(l_1,\dots,l_n)$5 (Nguyen et al., 5 Mar 2026). This is direct evidence that abstraction and retrieval decomposition improve robustness when near-duplicate support is unavailable.

Legal retrieval papers report comparable gains from reformulation. PromptCase improved SAILER on LeCaRD from $\mathcal{L}=(l_1,\dots,l_n)$6, $\mathcal{L}=(l_1,\dots,l_n)$7, $\mathcal{L}=(l_1,\dots,l_n)$8 to $\mathcal{L}=(l_1,\dots,l_n)$9, $M=10\%$0, $M=10\%$1, and on COLIEE2023 from $M=10\%$2, $M=10\%$3, $M=10\%$4 to $M=10\%$5, $M=10\%$6, $M=10\%$7 (Tang et al., 2023). LEAD scales the supervision side of the problem rather than the architecture: it constructs 100,060 synthetic criminal query-case pairs across 210 charges, and a LEAD-trained dense retriever reaches $M=10\%$8, $M=10\%$9, $k=5$0 on asymmetric LeCaRD, above SAILER’s $k=5$1, $k=5$2, and $k=5$3, while also improving on CAIL2022-LCR (Gao et al., 2024). KELLER, which combines knowledge-guided reformulation with sub-fact matching, reaches $k=5$4, $k=5$5, $k=5$6 on LeCaRD and $k=5$7, $k=5$8, $k=5$9 on LeCaRDv2, with especially large improvements on controversial queries (Deng et al., 2024).

Outside text, the same pattern holds. In malignant lymphoma retrieval, the proposed attention-plus-IHC-metric method achieved top-5 IHC staining accuracy $c=\langle Q,S,E,A\rangle$0 in the 40x+5x multi-scale setting, above all reported baselines, and pathologists preferred its retrievals significantly more often than those of subtype-based metric learning with high-attention patches, with average binary preference $c=\langle Q,S,E,A\rangle$1 and $c=\langle Q,S,E,A\rangle$2 when only confident responses were counted (Hashimoto et al., 2021). In legal QA, structured case-based retrieval also improves downstream generation: the best CBR-RAG configuration, Hybrid AnglEBERT with $c=\langle Q,S,E,A\rangle$3 full-case context, achieved answer similarity $c=\langle Q,S,E,A\rangle$4 versus $c=\langle Q,S,E,A\rangle$5 for No-RAG, an average 1.94% increase in performance (Wiratunga et al., 2024).

The evaluation methodology itself has expanded. MUSER introduces a multi-view relevance rubric in which legal fact, dispute focus, and law statutory are scored separately and combined, with relevance defined by a total score of at least 7. Its legal-element retriever outperformed standard lexical and neural baselines on $c=\langle Q,S,E,A\rangle$6 and all reported NDCG metrics, indicating that explicit legal structure improves top-rank usefulness (Li et al., 2023). CBR-to-SQL introduces brittleness as a perturbation-based robustness measure, and several legal papers now distinguish zero-shot from fine-tuned performance, reflecting a shift from pure peak accuracy to retrieval behavior under realistic constraints (Nguyen et al., 5 Mar 2026, Su et al., 2023).

6. Limitations, misconceptions, and open directions

A recurring misconception in this literature is that more raw context or more complex architecture automatically yields better similar-case retrieval. Several papers argue against this directly. CBR-to-SQL identifies an anti-pattern in which practitioners enlarge the static demonstration pool when single-step retrieval underperforms; the paper argues that this increases noise, redundancy, and compute cost while leaving the underlying conflation of structure matching and entity grounding unresolved (Nguyen et al., 5 Mar 2026). PromptCase shows a similar pattern: legal features are the main driver of retrieval gains, while prompt-only reformulation without legal-feature extraction does not help (Tang et al., 2023). Query-content-selection experiments likewise show that summary reformulation helps more consistently than naive extraction, even though key sentence selection has higher overlap with human salience annotations, because fluent reformulations better preserve usable legal context for the retriever (Zhou et al., 2023).

Another misconception is that domain-specific encoders or formally richer graph models are always superior. In CBR-RAG, LegalBERT did not outperform AnglEBERT, and the paper attributes this to the importance of contrastive optimization for retrieval rather than legal-domain pretraining alone (Wiratunga et al., 2024). In the CaseLink reproducibility study, the heterogeneous HGT formulation underperformed the simpler homogeneous graph, and graph-based retrieval was found to be competitive but not uniformly superior to strong text baselines such as BM25 (Donabauer et al., 11 Apr 2025). This suggests that representation-task alignment may matter more than nominal structural sophistication.

The literature also contains evaluation caveats. “Judgement Citation Retrieval using Contextual Similarity” reports 90.9% accuracy for ANN + KMeans + embeddings, but this is cluster-classification accuracy rather than a direct retrieval-ranking metric, so it supports semantic candidate generation more clearly than validated similar-case ranking (Dasula et al., 2024). The Calcutta High Court RAG system indexes two-step summaries and retrieves top-3 similar documents, but its quantitative evidence is summarization-oriented rather than retrieval-oriented, so the claim of improved similar-case retrieval remains architectural and indirect (Banerjee et al., 28 Jun 2025). Law-aware matching frameworks raise a different limitation: Law-Match requires cited or predicted law articles, and LCM-LAI is naturally suited to civil-law settings where codified law articles are explicit; transfer to common-law settings remains future work in both cases (Sun et al., 2022, Xu et al., 25 Feb 2025).

Several datasets and methods also make clear that the problem is not solved. MUSER explicitly states that incorporating legal elements can benefit SCR models, but further efforts are still required to address the remaining challenges posed by the dataset, including complex multi-view similarity and difficult legal-element prediction, especially for dispute focus (Li et al., 2023). Similar limitations appear elsewhere as entity-tagging errors in CBR-to-SQL, long-document truncation in dense legal retrievers, and the computational cost of LLM-based reformulation in KELLER (Nguyen et al., 5 Mar 2026, Deng et al., 2024). This suggests that successful enhancement is less about a single universally superior model than about matching representation, supervision, and ranking design to the domain’s operative notion of relevance.

Across domains, the most stable principle is therefore not “use a larger model” or “use a more complex retriever,” but “retrieve over the right abstraction.” When cases are reformulated into legally or clinically decisive units, when similarity is decomposed into the components that actually govern reuse, and when ranking exposes the local evidence behind a match, similar-case retrieval becomes more robust, more sample-efficient, and more interpretable (Tang et al., 2023, Hashimoto et al., 2021, Deng et al., 2024).