Direct-Access Ranked Retrieval

• Direct-Access Ranked Retrieval (DAR) is a retrieval paradigm that returns the answer at any specified rank without enumerating preceding results, leveraging quasilinear preprocessing and logarithmic access times.
• It integrates database theory, geometric duality, and automata techniques to address rank retrieval across conjunctive queries, numerical datasets, and string evaluations.
• Key implementations range from layered join trees for structured queries to ε-sampling and hierarchical strategies for managing high-dimensional and compressed data.

Searching arXiv for the cited DAR-related papers to ground the article. Direct-Access Ranked Retrieval (DAR) denotes retrieval by ranked position: after preprocessing a dataset for a query model and an order, the system returns the answer at rank $k$ without enumerating the preceding $k-1$ answers. In database theory, DAR is framed as simulating a sorted array of query answers with quasilinear preprocessing and logarithmic access; in interactive data systems, it includes arbitrary query-time linear scoring functions over points; and in logical settings over strings, it asks for the $t$-th lexicographically ordered satisfying assignment of an MSO query (Carmeli et al., 2020, Dehghankar et al., 1 Aug 2025, Hu et al., 9 Jan 2026, Muñoz, 13 Mar 2026).

1. Definition, scope, and terminological boundaries

Within the DAR literature, the central object is not merely a top-$k$ prefix but an arbitrary rank position. The access primitive is therefore stronger than ranked enumeration and different from ordinary top-$k$ retrieval, because the target rank can lie deep in the order and should be reachable without materializing prior answers (Dehghankar et al., 1 Aug 2025, Hu et al., 9 Jan 2026).

The acronym itself is not stable across arXiv. In "Augmenting Document Representations for Dense Retrieval with Interpolation and Perturbation," DAR means Document Augmentation for dense Retrieval, a training-time embedding-space augmentation method for dense retrievers rather than a direct-access retrieval paradigm (Jeong et al., 2022). In "Beyond Text-to-SQL: Autonomous Research-Driven Database Exploration with DAR," DAR means Data Agnostic Researcher, a multi-agent BigQuery system for autonomous database exploration rather than ranked retrieval (Vykhopen et al., 16 Dec 2025).

Some adjacent work has strong conceptual overlap without being canonical DAR. "DynamicRetriever" replaces sparse and dense external indexes with direct document-identifier prediction from model parameters, so ranking is induced by sorting probabilities over docids; however, it does not define a new efficient direct-access algorithm with sublinear access guarantees, and its main formulation is closer to large-vocabulary document classification than to formal DAR (Zhou et al., 2022). This suggests that DAR is best treated as a family of rank-access problems rather than as a single architecture.

2. Formal problem statements and ranking models

For conjunctive queries, the formal DAR task begins with a database instance $I$, a conjunctive query $Q$, and a total order $\preceq$ over the answer set $Q(I)$. Direct access asks for a data structure such that, given an index $k$, it returns the answer occupying position $k-1$0 in the sorted order, with “out-of-bound” returned when $k-1$1 exceeds the number of answers (Carmeli et al., 2020). The tractability yardstick is preprocessing $k-1$2 and access $k-1$3, often written as $k-1$4 (Carmeli et al., 2020).

Two ranking families dominate the CQ theory. The first is lexicographic ordering over free variables, written $k-1$5. The second is an additive order over per-variable weights,

$k-1$6

with DAR defined over any total order refining nondecreasing score (Carmeli et al., 2020).

In the geometric formulation for interactive data systems, the dataset is

$k-1$7

and a query supplies a normalized linear scoring vector $k-1$8. Scores are

$k-1$9

and the DAR query is

$t$0

that is, the tuple at position $t$1 under the query-time ranking function $t$2 (Dehghankar et al., 1 Aug 2025). This formulation differs sharply from classical top-$t$3 retrieval because $t$4 may be anywhere in $t$5, not only near the head of the ranking (Dehghankar et al., 1 Aug 2025).

In MSO query evaluation over strings, the answer set is the set of mappings from open variables to positions satisfying the formula. Equivalently, the paper works with an unambiguous functional vset automaton $t$6, and the rank query asks for the $t$7-th mapping in $t$8 under lexicographic order induced by a variable order $t$9 (Muñoz, 13 Mar 2026). The ordering is therefore canonical lexicographic rank over assignments, not an arbitrary user-defined scoring function (Muñoz, 13 Mar 2026).

3. Tractability theory for conjunctive queries

The main theoretical dichotomies for DAR over conjunctive queries are developed for self-join-free CQs and are organized around query structure and order compatibility. For lexicographic orders, direct access is tractable exactly when the query is free-connex and the order avoids a structural obstruction called a disruptive trio; for partial lexicographic orders, $k$0-connexity is additionally required (Carmeli et al., 2020). A disruptive trio consists of free variables $k$1 such that $k$2 and $k$3 are not neighbors, $k$4 is a neighbor of both, and $k$5 appears after both in the lexicographic order (Carmeli et al., 2020).

The corresponding hardness results are conditional. If the structural conditions fail, then lexicographic DAR is not possible in $k$6 under sparseBMM and Hyperclique assumptions (Carmeli et al., 2020). For sum-of-weights orders, the tractable class is much narrower: direct access is possible only when the query is acyclic and some atom contains all free variables; otherwise, SUM-DAR is not possible in $k$7 assuming 3sum and Hyperclique (Carmeli et al., 2020).

The same paper separates DAR from the weaker selection problem. For lexicographic selection, free-connexity alone suffices for tractability, and the paper gives linear-time selection $k$8; for SUM-selection, the condition is free-connexity together with $k$9, where $k$0 is the number of maximal hyperedges in the free-restricted hypergraph (Carmeli et al., 2020). Consequently, selection can be strictly easier than reusable direct access, even for the same query and order family (Carmeli et al., 2020).

Algorithmically, the lexicographic direct-access construction uses a layered join tree aligned with the order. Preprocessing performs semijoin reduction, sorting, bucketization by already-fixed variables, and dynamic-programming counts. For each tuple $k$1, the structure stores the number of compatible solutions in its subtree, cumulative bucket weights, and range endpoints, so the access procedure can locate the block containing the desired rank by binary search and descend recursively (Carmeli et al., 2020). The same framework extends under unary functional dependencies by replacing the original query $k$2 with its FD-extension $k$3, and for lexicographic orders by similarly reordering to $k$4 (Carmeli et al., 2020).

4. String, compressed-text, and dynamic formulations

A precursor to DAR in ranked document retrieval appears in "Ranked Document Retrieval in (Almost) No Space," which augments Wavelet Trees on Bytecodes so they support top-$k$5 conjunctive and bag-of-words retrieval directly over a compressed text representation. The method computes exact tf-idf scores without storing per-term postings lists, using just $k$6 extra over compressed text space, reported experimentally as $k$7 total size for WTBC-DR and $k$8 for WTBC-DRB relative to original text size (Brisaboa et al., 2012). Query processing is direct over compressed segments via rank/select navigation and best-first segment splitting, but the speed tradeoff is substantial: milliseconds to hundreds of milliseconds rather than the tens to hundreds of microseconds associated with in-memory inverted indexes (Brisaboa et al., 2012).

The logical-string DAR line is more explicit. "Dynamic direct (ranked) access of MSO query evaluation over SLP-compressed strings" studies direct ranked access to answers of MSO queries over strings, represented via unambiguous functional vset automata (Muñoz, 13 Mar 2026). For fixed variable order $k$9, the paper gives

$I$0

time per access after

$I$1

preprocessing on an uncompressed string $I$2. If the variable order is supplied only at access time, preprocessing becomes

$I$3

while the same direct ranked access bounds hold (Muñoz, 13 Mar 2026).

The same asymptotic access time is preserved for SLP-compressed strings. If $I$4 is an SLP representing $I$5, fixed-order preprocessing becomes

$I$6

with logarithmic-time access in the length of the original string (Muñoz, 13 Mar 2026). The core mechanism is selection by counting on lexicographic slices, using matrices $I$7 that count partial runs of the automaton under interval restrictions and compose by matrix multiplication over balanced decompositions (Muñoz, 13 Mar 2026). The dynamic extension incorporates complex edits using strongly balanced SLPs and document-editing machinery adapted from Schmid and Schweikardt; the paper states that this preserves logarithmic direct access after edits, although it also notes a likely typo in one theorem statement where linear rather than logarithmic dependence appears (Muñoz, 13 Mar 2026).

5. Geometric DAR for interactive weighted ranking

The most explicit modern formulation of DAR as a standalone retrieval primitive appears in "Efficient Direct-Access Ranked Retrieval" (Dehghankar et al., 1 Aug 2025). The paper treats DAR as a query-time ranking problem over high-dimensional numerical tuples and proposes two algorithmic families: an exact arrangement-based method with logarithmic query time, and linear-space sampling-based methods that relax or approximate exact rank location (Dehghankar et al., 1 Aug 2025).

The exact method, KthLevel, is built from geometric duality. Each point $I$8 is mapped to the hyperplane

$I$9

and a query scoring function $Q$0 becomes a ray from the origin. The order of intersections of that ray with the dual hyperplanes matches the rank order of points under $Q$1 (Dehghankar et al., 1 Aug 2025). Preprocessing constructs the arrangement of hyperplanes and the $Q$2-th level for every rank $Q$3, so a query $Q$4 reduces to point location on the projected arrangement, yielding $Q$5 query time with $Q$6 space (Dehghankar et al., 1 Aug 2025). In 2D, the paper reports $Q$7 total preprocessing; in 3D and higher, the arrangement complexity becomes prohibitive, reaching $Q$8 in general (Dehghankar et al., 1 Aug 2025).

The paper then shows why exact linear-space DAR is difficult. A DAR structure can be used to solve half-space range counting by binary search on rank, implying the lower bound

$Q$9

for a DAR algorithm $\preceq$0 using $\preceq$1 space (Dehghankar et al., 1 Aug 2025). In the linear-space regime $\preceq$2, this yields roughly $\preceq$3 query time up to logarithmic factors, which explains why logarithmic-time exact DAR requires very large space (Dehghankar et al., 1 Aug 2025).

To address that barrier, the paper introduces Conformal Set Ranked Retrieval (CSR), which returns a small subset $\preceq$4 guaranteed to contain the target tuple, with $\preceq$5 (Dehghankar et al., 1 Aug 2025). The key intermediate problem is Stripe Range Retrieval (SRR) over ranges

$\preceq$6

that is, narrow bands between two hyperplanes orthogonal to the scoring vector (Dehghankar et al., 1 Aug 2025). Using $\preceq$7-samples, the paper derives stripe boundaries from ranks on the sample and proves that the target point lies inside the resulting stripe. In 2D, the exact method Eps2D achieves $\preceq$8 query time with linear space. In higher dimensions, EpsRange gives CSR in

$\preceq$9

time and exact DAR in

$Q(I)$0

time, again with linear space (Dehghankar et al., 1 Aug 2025).

For practical high-dimensional SRR, the paper replaces classical range indexes with a hierarchical sampling structure, EpsHier, built by recursive subsampling and enclosing balls. Its worst-case query time can degrade to $Q(I)$1, but it remains linear in space and is empirically effective on narrow stripes (Dehghankar et al., 1 Aug 2025). On synthetic data, hierarchical sampling yields up to $Q(I)$2 speedup over exhaustive SRR, and the paper reports scalability to millions of tuples and hundreds of dimensions on real and synthetic datasets (Dehghankar et al., 1 Aug 2025).

6. Implementations, adjacent directions, and limitations

"Database Theory in Action: Direct Access to Query Answers" moves DAR from theory to implementation for conjunctive queries and tractable lexicographic and sum orders (Hu et al., 9 Jan 2026). The system automatically analyzes whether a query-order pair falls into a known tractable class and applies the corresponding algorithm. It is compared with PostgreSQL on synthetic joins, including

$Q(I)$3

under full lexicographic order $Q(I)$4 and partial orders such as ordering only by $Q(I)$5 (Hu et al., 9 Jan 2026). The paper reports that direct access and single access outperform PostgreSQL on large join results, while PostgreSQL can be competitive or faster on small results or small $Q(I)$6 because its plans behave more like top-$Q(I)$7 or ranked enumeration strategies (Hu et al., 9 Jan 2026). A central empirical finding is that direct access and single access are largely insensitive to the target position $Q(I)$8, whereas PostgreSQL performance varies significantly with $Q(I)$9; the observed break-even ratio between direct access and single access lies in $k$0, leading the paper to conclude that even a small number of accesses, such as the three quartiles, can justify preprocessing (Hu et al., 9 Jan 2026).

Related work has extended ranked retrieval quality control rather than direct-access indexing itself. "Two-stage Risk Control with Application to Ranked Retrieval" formulates a retrieval-then-ranking pipeline with a retrieval loss

$k$1

and a ranking loss

$k$2

then calibrates thresholds so expected stagewise risks are bounded by user-specified levels (Xu et al., 2024). The method is not a DAR algorithm, but it is directly relevant when a direct-access candidate generator feeds a reranker (Xu et al., 2024).

Across the literature, several limitations recur. First, tractability is highly sensitive to the joint choice of query and order; the same query can admit DAR under one order and be hard under another (Carmeli et al., 2020, Hu et al., 9 Jan 2026). Second, much of the exact theory is tied to canonical order families—lexicographic variable orders, additive weights, or linear scoring functions—and does not support arbitrary ranking semantics (Carmeli et al., 2020, Dehghankar et al., 1 Aug 2025, Muñoz, 13 Mar 2026). Third, exact low-latency access in linear space is obstructed by range-counting lower bounds in the geometric setting (Dehghankar et al., 1 Aug 2025). Fourth, compressed and dynamic string results depend on unambiguous functional vset automata and, under edits, strongly balanced SLP representations (Muñoz, 13 Mar 2026). Finally, adjacent model-based retrieval systems such as DynamicRetriever can directly predict ranked document identifiers without sparse or dense indexes at inference time, but they do not by themselves resolve the core DAR requirement of efficient reusable access by rank position (Zhou et al., 2022).

Taken together, these lines of work define DAR as a broad but coherent research area centered on one primitive: direct retrieval of the answer at a specified rank. Its realizations range from layered join trees and matrix-counting decompositions to geometric arrangements, $k$3-sampling, and compressed self-indexes. The unifying criterion is not the surface domain—database tuples, string assignments, or documents—but the avoidance of prefix enumeration under an explicitly modeled ranking order (Carmeli et al., 2020, Dehghankar et al., 1 Aug 2025, Muñoz, 13 Mar 2026).