Label-Centric Inverted Indexing

• Label-centric inverted indexing is a novel approach that replaces traditional keyword matching with learned semantic IDs and attribute labels to enhance semantic retrieval.
• It employs neural encoders and quantization techniques to generate discrete labels that serve as keys in posting lists, streamlining query processing and index construction.
• Empirical results demonstrate significant improvements in recall and speed, making this method a scalable and efficient solution for large-scale semantic search applications.

Label-centric inverted file indexing refers to a family of model-based and hybrid approaches that replace traditional term-based inverted indexes with discrete, semantically meaningful labels—such as learned semantic IDs or assigned attribute sets—and construct the index structure around these labels. Rather than relying on lexical token overlap, label-centric methods leverage learned or attributed labels as the primary keys in posting lists, facilitating flexible, efficient, and generalizable retrieval in both semantic search and attribute-constrained nearest neighbor contexts. The two principal lines of work—semantic label indexing as in UniDex (Li et al., 29 Sep 2025) and label-hybrid AKNN as in Elastic Index Select (Yang et al., 6 May 2025)—share the core principle of mapping items (documents or vectors) and queries to sets of discrete labels and inverting over these for retrieval.

1. Foundations and Motivation

Classical inverted files operate on the assumption that surface-level token matches between queries and documents capture relevance. This approach limits the ability to generalize semantically and struggles with queries or items lacking lexical overlap. Label-centric inverted indexing removes the reliance on raw tokens. In UniDex (Li et al., 29 Sep 2025), both queries and documents are mapped—via a neural encoder and quantizer pipeline—to a fixed-size set of learned discrete semantic IDs (SIDs), which then serve as posting keys. In the hybrid AKNN domain (Yang et al., 6 May 2025), high-dimensional embeddings are coupled with attribute label sets, and posting structures are built over these labels.

Label-centric indexing thus generalizes inverted files to settings where "labels"—not direct tokens—encapsulate relevance, enabling: (1) semantic generalization, (2) compositional coverage in hybrid vector+attribute search, and (3) efficient sharing/reuse of index structures across overlapping label sets.

2. Index Architecture and Label Assignment

2.1 UniDex: Semantic ID-Centric Indexing

• Label generation: UniTouch, the encoding module, uses a Transformer-based dual-tower architecture with appended learnable semantic tokens. Each token is projected and quantized via Finite Scalar Quantization (FSQ), yielding $d_q$-dimensional, $K$-ary (typically binary) labels, interpreted as SIDs in $[0, K^{d_q}-1]$.
• Multi-label assignment: Each query receives $M$ SIDs, and each document $N$ SIDs, facilitating polysemy and compositional semantics.
• Inverted file layout: The index consists of posting lists keyed by observed SIDs, mapping each SID $s$ to the list of documents producing $s$.

2.2 Elastic Index Select: Label-Set Hybrid Indexing

• Label definitions: Each data point is represented as $(x_i, L_i)$ where $x_i \in \mathbb{R}^d$ and $L_i \subseteq \mathcal{L}$.
• Index partitioning: Instead of creating an index for every possible query label set, a small family of "covering" label sets (subsets of $K$0) is chosen using a greedy algorithm which optimizes a global coverage/efficiency metric called the elastic factor.
• Partial posting lists: For selected $K$1, posting lists $K$2 are constructed, along with a corresponding vector subindex (e.g., IVF-PQ or HNSW).

3. Query Processing and Retrieval Dynamics

Label-centric query execution departs from token match scoring. In UniDex (Li et al., 29 Sep 2025):

• At runtime, a user query is mapped (online) to its $K$3 SIDs.
• All posting lists for these SIDs are fetched and merged (union), forming the candidate set:

$K$4

where $K$5 is the query's SID set and $K$6 is the posting list for $K$7.

• Ranking is performed using UniRank, a late-interaction transformer-based scorer over semantic tokens.

In Elastic Index Select (Yang et al., 6 May 2025):

• The query $K$8, with $K$9, is dispatched to a pre-selected covering set $[0, K^{d_q}-1]$0 identified via a precomputed inclusion DAG.
• The search operates solely within the vector index constructed over $[0, K^{d_q}-1]$1, returning the top-$[0, K^{d_q}-1]$2 results without further filtering.

4. Index Construction, Maintenance, and Complexity

In model-based label-centric approaches (Li et al., 29 Sep 2025), indexing is a one-time process:

• Each corpus document is encoded, quantized, and all produced SIDs are appended to corresponding posting lists.
• Complexity: Construction is $[0, K^{d_q}-1]$3; per-document updates are $[0, K^{d_q}-1]$4 for posting updates.
• Online query handling involves SID computation for the query (via Transformer + quantization) and posting list lookups.

In label-hybrid schemes (Yang et al., 6 May 2025):

• Index selection via the greedy EIS algorithm runs in $[0, K^{d_q}-1]$5 per round for a query workload of size $[0, K^{d_q}-1]$6.
• Index building per partial index (using IVF or HNSW) is $[0, K^{d_q}-1]$7 in size $[0, K^{d_q}-1]$8 (with $[0, K^{d_q}-1]$9 neighbors).
• Dynamic maintenance is feasible by monitoring label statistics and updating the covering index set as the distribution evolves.

5. Theoretical Guarantees and Elastic Factor

Elastic factor bounds (Yang et al., 6 May 2025) govern the efficiency/space tradeoff. For any query label set $M$0, the elastic factor $M$1 is defined as:

$M$2

where $M$3 is the set of constructed index covers. The slowdown factor $M$4 bounds retrieval cost at most $M$5 times that of an optimal per-query index.

Greedy selection achieves a logarithmic approximation with respect to the theoretical optimum. Empirical analyses demonstrate:

• With $M$6, index space increase is $M$7–$M$8 reified index size and achieves over $M$9–$N$0 speedup versus prior label-filtered graph methods at 95–98% recall.
• The architecture robustly adapts to skewed label distributions and heavy-tailed query mixes by combining statically chosen partial indices with dynamic dispatch.

6. Empirical Results and Practical Impact

UniDex: Model-Based Label Inversion (Li et al., 29 Sep 2025)

Method	Recall@300 (%)	MRR@10 (%)
BM25	49.6	22.2
SPLADE (term-max)	56.6	25.0
Kuaishou (term-touch+rank)	55.3	27.5
UniTouch (6L)+term rank	65.2	32.3
UniTouch (12L)+UniRank	68.6	33.2
UniTouch (24L)+UniRank (UniDex)	70.7	34.1

For comparison, dense retrievers reach 70–73% Recall@300. UniDex approaches this quality at a fraction of the computational and infrastructural cost, yielding −25% end-to-end latency, −20 thousand CPU cores, and −37 TB RAM relative to previous term-index pipelines at industrial scale.

UniDex further demonstrates >14% improved recall for rare (long-tail) queries and reduces the frequency of scarce relevant results, highlighting generalization beyond surface-token dependencies.

Elastic Index Select: Label-Attribute Hybrid Indexing (Yang et al., 6 May 2025)

• ELI-0.2 (elastic factor $N$1): Up to 500× faster AKNN than ACORN/UNG baselines at comparable ($N$298%) recall, with index size only $N$3–$N$4 the size of the unlabeled base index.
• ELI-2.0 (space factor $N$5): Further speed increase.
• Index build time remains under $N$6 of vanilla HNSW.
• Stable performance over large-scale datasets (SIFT 1M, GIST 1M, MS MARCO, OpenAI 1M–100M scale) and distribution types, supporting up to 100 million vectors and 32 attribute labels.

7. Integration and Applicability

Label-centric inverted indexing frameworks are directly compatible with standard search and ANN infrastructures. In practice:

• Model-based label ID systems (e.g., UniDex) supplant term-matching with label matching and plug directly into existing inverted file-based engines.
• Attribute-based partitioning as in EIS integrates via construction of multiple small, query-dispatchable IVF or HNSW indices, with precomputed mapping tables for efficient runtime index selection.
• Both paradigms offer trivial parallelization, since each query resolves to a fixed combination of posting lists or a single selected subindex.

A plausible implication is that as large-scale vector search and semantic retrieval mechanisms become ubiquitous, label-centric inversion provides a scalable, flexible foundation for both performance-critical and semantically rich information retrieval scenarios. Both theoretical analysis and industrial-scale experimentation support their effectiveness across a range of workloads and data regimes (Li et al., 29 Sep 2025, Yang et al., 6 May 2025).