Xetrieval: Explainable Embedding Retrieval

• Xetrieval is a framework in explainable retrieval that decomposes dense embeddings into sparse, human-interpretable features to clarify ranking decisions.
• It leverages methods like modified LIME, reasoning internalizers, and sparse autoencoders to generate both local word-level and global mechanistic explanations.
• Empirical evaluations show that targeted feature steering and intervention studies validate that selected latent features significantly influence retrieval performance.

Searching arXiv for the cited Xetrieval and EXS papers to ground the article in the current literature. Xetrieval denotes a line of work in explainable retrieval that makes ranking decisions inspectable rather than treating retrieval scores as irreducibly opaque. In an earlier instantiation, Xetrieval was operationalized by EXS, an explainable-search system that adapts LIME to retrieval in order to answer three questions: what the ranker infers as the query’s intent, why one document is ranked above another, and why a document is relevant to a query (Singh et al., 2018). In its more recent formulation, Xetrieval is an embedding-level mechanistic framework for explaining dense retrieval by enriching embeddings with reasoning-oriented information, decomposing them into sparse human-interpretable features, and explaining retrieval decisions through feature co-activation between query and document representations (Cai et al., 28 May 2026).

1. Early explainable-search formulation

As instantiated by EXS, Xetrieval is a web-based search front-end backed by a REST API. Documents such as AP news, LATIMES, Robust04, and Financial Times are indexed with Lucene. A user selects both a corpus and a retrieval model, including classic BM25 and pointwise neural rankers such as DRMM or DESM implemented in TensorFlow. For a query $q$, the selected ranker $R(q,\cdot)$ scores documents and returns a top-$k$ list $D_q^k$, displayed in a title-plus-snippet layout. EXS adds three explanation modes: per-document relevance, document-pair ordering, and query intent. When an explanation is requested, a modified LIME interpreter produces word-level importance scores rendered as horizontal bar charts (Singh et al., 2018).

The central methodological move in EXS is to reinterpret ranking as a local relevance-classification problem. LIME’s original classifier-oriented pipeline samples perturbed instances by randomly deleting words, weights them by a locality kernel $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$, queries the black-box model, and fits a sparse linear surrogate in bag-of-words space. EXS preserves that structure but constructs $P(X=\mathrm{relevant}\mid q,d')$ from ranking outputs using three schemes: top-$k$ binary, score-based, and rank-based. It then samples 2,000 perturbed documents near $d$, fits a sparse linear regressor on $\phi(d')$, and interprets the sign and magnitude of $w_j$ as each word’s contribution to local relevance.

The surrogate task is stated as

$R(q,\cdot)$0

with $R(q,\cdot)$1 converted into a probability of relevance, $R(q,\cdot)$2 the perturbed samples, and $R(q,\cdot)$3 enforcing sparsity. Positive coefficients are visualized as green bars and negative coefficients as red bars. In the worked example for the query “Rail Strikes” and document AP890710-0178, EXS highlights unrest and walkouts as strong positive terms and jimmy as a negative term. For pairwise explanation, the system uses the score or rank of $R(q,\cdot)$4 as the locality threshold and explains why $R(q,\cdot)$5 appears above $R(q,\cdot)$6; in the same query, railway and union are identified as decisive positive features that push AP890710-0178 above AP890713-0045. For intent explanation, coefficient vectors are aggregated across top-$R(q,\cdot)$7 results as $R(q,\cdot)$8, yielding transport, strikes, and union as the inferred intent terms.

EXS did not report a large-scale user study or automatic metrics; the evidence presented is qualitative and anecdotal. The paper nevertheless argues that the system helps users verify whether learned intent matches expectations, isolate terms driving ordering, and inspect how document editing might improve rank. This suggests a foundational distinction in Xetrieval’s development: the early system explains retrieval through local lexical surrogates, whereas later work seeks explanations at the embedding level.

2. Dense retrieval problem and mechanistic objective

In the 2026 formulation, Xetrieval addresses dense retrieval, where a query $R(q,\cdot)$9 and each document $k$0 are mapped into a shared $k$1-dimensional vector space. The query encoder produces $k$2, the document encoder produces $k$3, and relevance is computed either by dot product,

$k$4

or by cosine similarity,

$k$5

Retrieval then reduces to nearest-neighbor search in $k$6 over pre-indexed document embeddings (Cai et al., 28 May 2026).

The explanatory problem is motivated by the fact that dense retrievers make decisions through opaque high-dimensional embeddings. Existing explanations are characterized as often focusing on surface signals, including lexical matches, token alignments, or post-hoc textual rationales, and therefore providing limited insight into latent factors at the embedding level. Xetrieval’s objective is to decompose each relevance score $k$7 into a small set of sparse, human-interpretable features that are co-activated in $k$8 and $k$9.

This reframes explanation from local word attribution to mechanistic factorization of embedding space. Instead of asking which observed tokens dominate a local surrogate, the framework asks which latent feature directions are jointly active and therefore explanatory of a retrieval decision. A plausible implication is that the explanatory unit shifts from lexical evidence to reusable latent mechanisms that may generalize across queries, documents, and tasks.

3. Reasoning internalizer

A central component of Xetrieval is the reasoning internalizer, introduced to address the entanglement of relevance factors in off-the-shelf embeddings. For each of three reasoning aspects,

$D_q^k$0

the internalizer applies a lightweight MLP:

$D_q^k$1

where $D_q^k$2, $D_q^k$3, $D_q^k$4 is element-wise, and $D_q^k$5 denotes $D_q^k$6 normalization. Each $D_q^k$7 has one hidden layer of size $D_q^k$8, with $D_q^k$9 given as an example (Cai et al., 28 May 2026).

Training proceeds by generating reasoning texts $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$0 for each original document $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$1 through prompting a LLM, then embedding both the original text and the reasoning text with the same encoder:

$\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$2

The internalizer is trained by minimizing

$\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$3

After training, each $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$4 approximates a chain-of-thought embedding in one feed-forward pass, avoiding expensive autoregressive generation at inference time.

The intended effect is to enrich sentence embeddings with reasoning-oriented information while retaining low inference cost. The paper’s description is explicit that the internalizer approximates chain-of-thought reasoning directly in the embedding space with a single forward pass. This does not claim full symbolic reasoning inside the retriever; rather, it inserts an intermediate representational layer designed to make subsequent decomposition more interpretable.

4. Sparse feature decomposition and explanation construction

Once reasoning-enhanced embeddings are available, Xetrieval decomposes embeddings with a sparse autoencoder. For any embedding $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$5, the encoder maps $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$6 to a sparse code $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$7 using a Top-K pass:

$\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$8

where $\pi_d(d')=\exp(-\mathrm{Dist}(d,d')^2/\sigma^2)$9 projects the input into $P(X=\mathrm{relevant}\mid q,d')$0 feature scores and all but the top $P(X=\mathrm{relevant}\mid q,d')$1 entries by magnitude are zeroed out. The decoder reconstructs

$P(X=\mathrm{relevant}\mid q,d')$2

The training loss is

$P(X=\mathrm{relevant}\mid q,d')$3

with sparsity enforced either by $P(X=\mathrm{relevant}\mid q,d')$4 or, in TopK-SAE, by the hard constraint $P(X=\mathrm{relevant}\mid q,d')$5 (Cai et al., 28 May 2026).

Training data for the sparse autoencoder includes both raw embeddings and reasoning-enhanced embeddings from the internalizer. After convergence, each column $P(X=\mathrm{relevant}\mid q,d')$6 is interpreted as a feature direction in $P(X=\mathrm{relevant}\mid q,d')$7, and a nonzero $P(X=\mathrm{relevant}\mid q,d')$8 indicates that feature $P(X=\mathrm{relevant}\mid q,d')$9 is active. To make features human-readable, the method retrieves top-activating documents for each feature, presents them to a LLM, and asks for a concise natural-language summary $k$0.

Explanations for a query–document pair are then built from shared activations. For the query, Xetrieval computes $k$1 and binarizes it as

$k$2

For a document $k$3, it forms multiple views,

$k$4

computes a sparse code for each view, and binarizes those activations in the same way. Shared features are defined by

$k$5

The explanation is

$k$6

The resulting explanatory claim is explicitly mechanistic at the feature level: feature $k$7 explains the match because both the query and at least one document-side view activate that feature. This differs from lexical rationalization in that the explanatory unit is a sparse latent factor with an associated hypothesis rather than a directly observed token.

5. Empirical evaluation and intervention studies

Xetrieval is evaluated on seven benchmarks—BRIGHT, NQ, MuTual, TREC-NEWS, Signal-1M, ArguAna, and Robust04—and eight retriever backbones: e5-small, e5-base, e5-large, gte-base, gte-large, Snowflake, Qwen3-0.6B, and Qwen3-4B. The primary retrieval metric is NDCG@10 (Cai et al., 28 May 2026).

Feature quality is analyzed through three procedures. First, the work examines reconstruction error of the sparse autoencoder as a function of sparsity level. Second, it evaluates mono-semanticity through intruder detection: for each feature, nine top-activating documents and one intruder are shown to a LLM, and detection accuracy is measured. The reported conclusion is that Xetrieval’s features achieve higher coherence than raw-SAE or random controls. Third, retrieval retention is tested by replacing original embeddings with decoded reconstructions; the paper reports that nearly full NDCG is retained at TopK=256.

Faithfulness is examined through pair-level intervention. For a subset $k$8 of selected explanation features, the method forms the decoder submatrix $k$9 and computes its ridge projection of $d$0. Two interventions are then defined:

$d$1

for erase, and

$d$2

for retain. The score change is measured by

$d$3

According to the reported results, Xetrieval’s selected features yield the largest negative $d$4 when erased and the largest positive $d$5 when retained, outperforming direct SAE decomposition and non-overlap features in intervention effect.

These analyses are significant because they evaluate explanations not only by readability but also by whether manipulating the purported explanatory features changes retrieval scores in the expected direction. A plausible implication is that Xetrieval treats interpretability as an intervention-sensitive property of the representation rather than merely a descriptive annotation layer.

6. Task-level feature steering, interpretation, and limits

Beyond explaining individual query–document pairs, Xetrieval introduces task-level feature steering. For a feature $d$6, the co-activation indicator is

$d$7

Given positive pairs $d$8 and negative pairs $d$9, the Retrieval Utility Score is defined as

$\phi(d')$0

Features are ranked by RUS to form a key set $\phi(d')$1. At inference time, steering modifies selected sparse-code entries by a scalar $\phi(d')$2: compute the raw code $\phi(d')$3, set $\phi(d')$4 for $\phi(d')$5, then decode $\phi(d')$6 and normalize (Cai et al., 28 May 2026).

The reported result is that amplifying key features consistently improves NDCG, while suppressing them degrades it; steering non-key features has smaller, inconsistent effects. The paper further states that Xetrieval’s key features yield stronger steering than raw SAE’s. This extends the framework from explanatory analysis to controlled modification of retrieval behavior, with feature sets functioning as globally useful or harmful factors.

Several interpretive boundaries are explicit. The framework is designed for dense retrieval and focuses on embedding-level latent factors rather than surface lexical evidence. The paper identifies limitations and future directions including finer-grained circuit-level explanations inside the encoder, extension to multi-modal and cross-lingual retrieval, and integration of more powerful sparse decomposition methods such as Transcoder. In relation to EXS, a common misconception would be to treat all Xetrieval work as a single technique. The available evidence instead supports two distinct but related explanatory regimes: EXS provides local, model-agnostic word-level explanations for ranking behavior, whereas the 2026 Xetrieval framework provides sparse, mechanistic, embedding-level explanations for dense retrieval.