MultiFaceted Learnable Index (MFLI)

• MultiFaceted Learnable Index (MFLI) is a retrieval framework that jointly learns multifaceted item embeddings and index structures, eliminating the need for traditional ANN search.
• It employs a multifaceted hierarchical codebook with residual quantization to organize item representations, ensuring efficient and semantically coherent retrieval.
• Innovations such as split-and-merge index balancing and GPU-friendly lookup enhance cold-content delivery, recall performance, and serving efficiency.

MultiFaceted Learnable Index (MFLI) is a scalable, real-time retrieval paradigm for large-scale recommendation systems that jointly learns multifaceted item embeddings and their indices within a unified framework and eliminates approximate nearest neighbor (ANN) search at serving time. In the formulation introduced in "Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System," item representations are organized through a multifaceted hierarchical codebook learned by residual quantization, and serving uses direct index lookup followed by reranking rather than per-request nearest-neighbor search over a continuous embedding corpus (Zhang et al., 18 Feb 2026). The resulting system is positioned as a response to two structural problems in industrial retrieval: the decoupling of embedding learning from offline index construction, and the serving cost of ANN at the scale of hundreds of millions of items and billions of users.

1. Problem formulation and design objectives

In a modern large-scale recommender system, retrieval is the first stage after candidate generation: given a user or a trigger item, the system must quickly retrieve a few hundred or thousand promising candidates from a corpus that may contain up to billions of items. The conventional ANN-based retrieval pipeline consists of learning dense item and user/query embeddings with a DNN, building an ANN index such as IVF, PQ, or HNSW over item embeddings offline, and then, for every request, computing the user/query embedding and running ANN search to retrieve the $K$ nearest items (Zhang et al., 18 Feb 2026).

The MFLI proposal begins from the claim that this pipeline has two structural limitations. The first is the decoupling of embedding learning and index construction. Embeddings may be updated continuously by online training, whereas indices are rebuilt offline on a separate cadence. This creates staleness and inconsistency, because the serving index can be constructed from an earlier checkpoint and no longer reflect the geometry of the current embedding space. It also makes real-time refresh expensive, since frequent offline reconstruction may require procedures such as k-means, PQ, or graph construction. Newly created items are especially affected: they may not be embedded or quantized consistently with exposed items, and may be missing or poorly placed in the ANN structure, which degrades cold-content delivery (Zhang et al., 18 Feb 2026).

The second limitation is serving-time ANN cost. Even when query complexity is sublinear, IVF or HNSW still compute distances to many candidates per query; graph and tree structures can have high memory overhead and poor GPU parallelization characteristics; and in multi-facet or multi-tower setups, serving cost grows roughly linearly in the number of query vectors. At industrial scale, ANN search itself becomes a serving bottleneck rather than merely an implementation detail (Zhang et al., 18 Feb 2026).

MFLI is therefore defined around a specific set of objectives: jointly learn embeddings and indices without offline index construction, eliminate ANN search at serving time, improve recall on engagement metrics, greatly improve cold-content delivery, improve semantic relevance, reduce popularity bias, and maintain real-time freshness through a delta-update pipeline. A common misconception is to treat MFLI as a minor variation on ANN acceleration. The formulation in (Zhang et al., 18 Feb 2026) is stronger: the serving path is designed to avoid ANN search altogether.

2. Multifaceted representations and joint optimization

The central representational unit in MFLI is the multifaceted item embedding

$v_i \in \mathbb{R}^{F \times d},$

where $F$ is the number of facets and $v_i[f] \in \mathbb{R}^d$ is the embedding for facet $f$. The paper describes facets as capturing distinct aspects such as aggregate engagement, semantic relevance, and, in some experiments, time-spent-related signals. Lower layers of the item tower are shared, while facets use separate heads or projections. This design is motivated by the claim that different engagement objectives and semantics often conflict, so a single embedding cannot capture them all well (Zhang et al., 18 Feb 2026).

Training uses co-engagement pairs, with a trigger item $i$, a recently engaged candidate item $j$, and sampled negatives $V_k$. The central loss is a multifaceted Sampled Softmax loss: $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$ where $w_{i,j}[f]$ is the facet-specific label. Each facet is thus trained with its own labels and potentially its own auxiliary tasks, including a semantic relevance loss applied at facet $v_i \in \mathbb{R}^{F \times d},$0 in the default configuration (Zhang et al., 18 Feb 2026).

The indexing mechanism is learned through a three-dimensional hierarchical codebook

$v_i \in \mathbb{R}^{F \times d},$1

with a facet-specific sub-codebook at each layer. For an item embedding $v_i \in \mathbb{R}^{F \times d},$2, each facet is quantized independently by multifaceted residual quantization. Starting from residual $v_i \in \mathbb{R}^{F \times d},$3, the method assigns, for each layer $v_i \in \mathbb{R}^{F \times d},$4 and facet $v_i \in \mathbb{R}^{F \times d},$5, the nearest codeword in $v_i \in \mathbb{R}^{F \times d},$6, subtracts it from the residual, and reconstructs the embedding after $v_i \in \mathbb{R}^{F \times d},$7 layers as

$v_i \in \mathbb{R}^{F \times d},$8

This is standard residual quantization applied per facet in parallel and is described as fully GPU-friendly because facets and layers can be processed in batched matrix operations (Zhang et al., 18 Feb 2026).

A key feature is that quantized embeddings enter the training objective together with raw embeddings: $v_i \in \mathbb{R}^{F \times d},$9 The stated purpose is to force codebooks such that quantized embeddings preserve retrieval performance, not merely reconstruction quality. This jointly learned embedding–index geometry is the core technical distinction from pipelines that first train embeddings and then build ANN indices offline (Zhang et al., 18 Feb 2026).

3. Hierarchical indexing, balancing, and storage layout

Once each item has facet-specific codes, MFLI turns those codes into a serving structure intended to be compact, balanced, and directly addressable. The paper emphasizes that raw quantization alone is insufficient, because some codewords may become overly popular and induce very large clusters while others remain empty. To mitigate this, MFLI adds a codebook regularization term, following Paria et al., that penalizes codewords that are overall too close to many residuals and thereby encourages more uniform cluster assignment. Training is further stabilized by delayed start, in which the embedding model is warmed up before codebook learning begins, and by layer-wise activation of quantization layers (Zhang et al., 18 Feb 2026).

Serving-time guarantees require a further offline rebalance step. MFLI defines size thresholds $F$0 and $F$1. Oversized indices are split by running k-means on the last-layer residual embeddings of items in that index and dividing them into $F$2 subindices. Undersized indices under the same parent are merged by reassigning their items to the smallest or a nearby index. The paper attributes substantial importance to this step: removing split-and-merge causes significant recall drops in ablations and also reduces semantic relevance (Zhang et al., 18 Feb 2026).

For each facet $F$3, the layer tuple is mapped to a scalar index

$F$4

After split-and-merge, each facet has $F$5 distinct indices. These are then offset into a contiguous global index space: $F$6 The global offsetting ensures that every unified index ID is unique across facets while still allowing the facet identity to be recovered from the interval in which the ID falls (Zhang et al., 18 Feb 2026).

The serving layout consists of an item-to-index mapping and an index-to-item mapping.

Structure	Stored content	Role
Item-to-index tensor	$F$7 dense tensor	O(1) item-to-index lookup
Direct ID map	item ID $F$8 row offset	O(1) row access
Count tensor	item counts per unified index	locate slice boundaries
Item ID tensor	all item IDs sorted by unified index	contiguous retrieval slice

The count tensor $F$9 stores the number of items in each unified index, and the item ID tensor $v_i[f] \in \mathbb{R}^d$0 stores all item IDs sorted by unified index. Items under index $v_i[f] \in \mathbb{R}^d$1 occupy a contiguous slice in $v_i[f] \in \mathbb{R}^d$2. The paper states that lookup is $v_i[f] \in \mathbb{R}^d$3 per index and memory is $v_i[f] \in \mathbb{R}^d$4, with a layout designed to be cache-friendly and GPU-friendly because it avoids pointer-chasing and graph traversal (Zhang et al., 18 Feb 2026).

The paper also allows per-facet item pruning through an invalid index. This is used when a facet is specialized, for example when facet $v_i[f] \in \mathbb{R}^d$5 is used only for cold content; masked items are reassigned to the invalid index so that only designated items are retrievable through that facet (Zhang et al., 18 Feb 2026).

4. Serving without ANN and real-time refresh

Serving follows an item-to-item retrieval pattern. A request first yields a set $v_i[f] \in \mathbb{R}^d$6 of trigger items extracted from the user interaction history. For each trigger item $v_i[f] \in \mathbb{R}^d$7, the system retrieves its multifaceted unified index vector $v_i[f] \in \mathbb{R}^d$8. Across triggers and facets, this yields $v_i[f] \in \mathbb{R}^d$9 unified indices with $f$0 total lookup cost (Zhang et al., 18 Feb 2026).

Index selection is then formulated as a multi-interest mining step. For each facet $f$1, the system builds a histogram

$f$2

normalizes it to $f$3, and can sample selected indices from a tempered distribution

$f$4

The paper also describes recent-interest boosting, long-tail exploration through neighboring indices sharing upper-level codes, and interest-strength-based quotas

$f$5

Selected indices from all facets are merged into a set $f$6 with $f$7 (Zhang et al., 18 Feb 2026).

Candidate retrieval then proceeds by direct slice lookup. For each selected unified index, the system obtains items from the full snapshot, maps the index back to associated original indices in the delta snapshot if split-and-merge has changed the topology, retrieves fresh items from those original indices, and unions the two candidate pools. Because split-and-merge ensures $f$8, the total candidate pool has size

$f$9

A reranker, described as a user-to-item reranker such as a MoL/MTMH-style tower, scores candidates within each selected index, retains the top quota, and passes the merged top lists to downstream ranking stages (Zhang et al., 18 Feb 2026).

This serving design is paired with a snapshot-based refresh scheme. A full snapshot, generated every $i$0 such as $i$1 minutes, re-embeds and re-quantizes all items, runs split-and-merge, and rebuilds the serving tensors. A delta snapshot, generated every $i$2 such as $i$3 minute, processes only fresh items, uses current model and codebooks without split-and-merge, and assigns original indices. Because the number of fresh items is described as approximately $i$4 smaller than the full pool, delta updates are comparatively cheap. A mapping from fine-tuned indices back to original indices reconciles full and delta structures at serving time (Zhang et al., 18 Feb 2026).

The resulting complexity profile is the basis for the claim that MFLI removes ANN from retrieval serving. The operations are reduced to index lookup over triggers, index-to-item lookup over selected indices, and reranking over $i$5 items. The paper explicitly contrasts this with centroid distance computations, graph traversal, and tree traversal, and reports a $i$6 throughput increase (QPS) versus MTMH, a strong multi-ANN baseline (Zhang et al., 18 Feb 2026).

5. Experimental results and observed system effects

The reported experiments use an industrial short-video recommendation system with hundreds of millions of items and billions of users, training on real data from period $i$7 and evaluating on a later period $i$8. The default MFLI configuration uses $i$9 facets, a two-layer codebook of size $j$0, facet $j$1 trained with co-engagement only, and facet $j$2 trained with co-engagement plus a semantic relevance auxiliary loss. Retrieval selects $j$3 indices, reranks per index, keeps $j$4 items per index, and evaluates recall@1000 after reranking. Baselines are NeuCF, MoL, HLLM, HSTU, MTMH, and VQIndex (Zhang et al., 18 Feb 2026).

The offline results are reported against the NeuCF baseline as follows.

Metric	MFLI result	Relative change vs NeuCF
VVC recall@1000	24.04%	+11.76%
Like recall@1000	26.72%	+9.33%
LWT recall@1000	27.49%	+6.34%
CCD recall@1000	11.86%	+57.29%
$j$5 topic match	52.31%	+110.93%
$j$6 topic match	38.70%	+126.32%

The abstract separately states that MFLI improves recall on engagement tasks by up to $j$7, cold-content delivery by up to $j$8, and semantic relevance by $j$9 compared with prior state-of-the-art methods. The detailed results further state that MFLI outperforms all baselines on semantic relevance and most on recall, with particularly strong gains on cold-content delivery. Compared with VQIndex, MFLI is described as combining high recall and high semantic coherence through its multifaceted design and split-merge balancing (Zhang et al., 18 Feb 2026).

Ablations identify split-and-merge as the most consequential systems component. Removing it reduces VVC from $V_k$0 to $V_k$1 and LWT from $V_k$2 to $V_k$3, and also lowers semantic relevance. Removing index selection causes semantic relevance to rise slightly but produces notable engagement-recall losses, including a $V_k$4 change on VVC, indicating the trade-off between tightly coherent semantic clusters and engagement-oriented index choice. Removing delayed start causes smaller degradations, which the paper interprets as evidence that it stabilizes training but is less critical than split-and-merge (Zhang et al., 18 Feb 2026).

The index distribution analysis is intended to show that the learned structure remains operationally well behaved at scale. With a two-layer codebook of size $V_k$5 over approximately $V_k$6M items, $V_k$7 of indices contain fewer than $V_k$8k items, approximately $V_k$9 lie between $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$0 and $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$1k items, more than $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$2 of items reside in indices with fewer than $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$3k items, and more than $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$4 of indices are non-empty and remain active over $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$5 days. The paper further reports that item pairs from the same index have much higher topic match rates than item pairs from different indices, which is presented as evidence that indices capture semantically coherent item clusters (Zhang et al., 18 Feb 2026).

Hyperparameter studies emphasize that multi-facet structure is material rather than decorative. Moving from one facet to the default two-facet setting improves VVC, LWT, and topic metrics; moving from two facets to three yields slightly higher LWT recall but somewhat lower topic match and about $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$6 more serving overhead. Similarly, codebook size controls a recall–semantic-purity trade-off: single-layer $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$7k gives worse recall but higher semantic relevance; two-layer $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$8 and $$L^{s}(v_i,v_j,V_k) = -\sum_{f=1}^{F} w_{i,j}[f] \log \frac{ e^{v_{i}^\top[f]\,v_{j}[f]} }{ e^{v_{i}^\top[f]\,v_{j}[f]} + \sum_{v\in V_k} e^{v_{i}^\top[f]\,v[f]} },$$9 give the best recall; and $w_{i,j}[f]$0 improves topic match but causes a slight recall drop (Zhang et al., 18 Feb 2026).

The online A/B test, conducted over $w_{i,j}[f]$1 days, reports reallocation of traffic toward less popular content, stronger exposure of fresh items, improved engagement efficiency, and higher serving capacity. Video views in the low-popularity buckets $w_{i,j}[f]$2 and $w_{i,j}[f]$3 increase by $w_{i,j}[f]$4 and $w_{i,j}[f]$5, while the highest bucket $w_{i,j}[f]$6 decreases by $w_{i,j}[f]$7. Exposures for items aged $w_{i,j}[f]$8 increase by $w_{i,j}[f]$9, and for items aged $v_i \in \mathbb{R}^{F \times d},$00 by $v_i \in \mathbb{R}^{F \times d},$01. Explicit engagement rises by $v_i \in \mathbb{R}^{F \times d},$02, engagement diversity by $v_i \in \mathbb{R}^{F \times d},$03, and throughput by $v_i \in \mathbb{R}^{F \times d},$04 QPS on GPU servers (Zhang et al., 18 Feb 2026).

6. Conceptual position, nomenclature, and limitations

Within the broader learned-index literature, MFLI occupies a specialized position. General surveys of learned multi-dimensional indexes organize the field around axes such as projected-space versus native-space methods, immutable versus mutable layouts, and pure versus hybrid learned structures (Al-Mamun et al., 2024). In that landscape, the recommendation-system MFLI of (Zhang et al., 18 Feb 2026) is neither a general-purpose dictionary index nor a conventional spatial learned index. It is a retrieval-specific learned index over embedding facets whose primary serving operation is candidate generation rather than predecessor, range, or kNN search in the database sense.

The acronym also has a second usage in the supplied literature. "LMSFC: A Novel Multidimensional Index based on Learned Monotonic Space Filling Curves" describes a MultiFaceted Learnable Index as a more general concept in which multiple learnable facets—such as multiple space-filling curves, workload-specific mappings, or different models per region—can coexist within a learned multidimensional indexing framework (Gao et al., 2023). That use is conceptually broader and rooted in learned monotone SFCs, paging, and recursive query splitting. By contrast, the recommendation MFLI uses multifaceted embeddings, hierarchical codebooks, unified index vectors, and direct lookup. The shared term therefore marks a family resemblance around workload-shaped, learnable indexing, but not an identical technical object.

A related contrast appears with the PGM-index literature, which treats indexing as learning the mapping from ordered keys to rank and develops multicriteria, compressed, and distribution-aware learned structures with explicit guarantees (Ferragina et al., 2019). MFLI shares the broader learned-index premise that index structures should adapt to the geometry induced by the target workload, but it applies that premise to item embedding retrieval rather than the dictionary problem. This suggests a point of contact between recommendation retrieval and database learned-index research: both frame indexing as a learnable approximation problem, yet they optimize different objectives and operate over different state spaces.

The limitations and trade-offs stated for MFLI are correspondingly system-specific. The method requires storing item-to-index tensors and index-to-item tensors for both full and delta snapshots; joint optimization of multi-facet embeddings, hierarchical codebooks, and balancing regularization is more complex than a standard two-tower plus ANN pipeline; performance depends on facet definition, codebook size, and size thresholds such as $v_i \in \mathbb{R}^{F \times d},$05 and $v_i \in \mathbb{R}^{F \times d},$06; and the full snapshot cadence assumes that behavior does not shift so rapidly that a $v_i \in \mathbb{R}^{F \times d},$07-minute lag becomes catastrophic (Zhang et al., 18 Feb 2026). The paper suggests extensions including more sophisticated quantization, additional facets for graph-based similarity or fairness objectives, joint modeling with generative retrieval, and tighter integration with GPU inference systems such as SilverTorch (Zhang et al., 18 Feb 2026).

The main interpretive conclusion is therefore narrow but consequential. MFLI does not merely compress ANN search; it reframes retrieval as learning discrete, multifaceted, semantically and behaviorally aligned clusters offline and routing requests through direct index lookup online. In the specific large-scale recommendation setting studied in (Zhang et al., 18 Feb 2026), that reframing is associated with improved engagement-oriented recall, much stronger cold-content delivery, reduced popularity bias, and substantially higher serving efficiency.