Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System

Jiang Zhang; Yubo Wang; Wei Chang; Lu Han; Xingying Cheng; Feng Zhang; Min Li; Songhao Jiang

cs.IRFeb 18, 2026

Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System

Authors: Jiang Zhang, Yubo Wang, Wei Chang, Lu Han, Xingying Cheng, Feng Zhang, Min Li, Songhao Jiang,

Organizations: Meta Platforms, Inc.

Abstract

Approximate nearest neighbor (ANN) search is widely used in the retrieval stage of large-scale recommendation systems. In this stage, candidate items are indexed using their learned embedding vectors, and ANN search is executed for each user (or item) query to retrieve a set of relevant items. However, ANN-based retrieval has two key limitations. First, item embeddings and their indices are typically learned in separate stages: indexing is often performed offline after embeddings are trained, which can yield suboptimal retrieval quality-especially for newly created items. Second, although ANN offers sublinear query time, it must still be run for every request, incurring substantial computation cost at industry scale. In this paper, we propose MultiFaceted Learnable Index (MFLI), a scalable, real-time retrieval paradigm that learns multifaceted item embeddings and indices within a unified framework and eliminates ANN search at serving time. Specifically, we construct a multifaceted hierarchical codebook via residual quantization of item embeddings and co-train the codebook with the embeddings. We further introduce an efficient multifaceted indexing structure and mechanisms that support real-time updates. At serving time, the learned hierarchical indices are used directly to identify relevant items, avoiding ANN search altogether. Extensive experiments on real-world data with billions of users show that MFLI improves recall on engagement tasks by up to 11.8%, cold-content delivery by up to 57.29%, and semantic relevance by 13.5% compared with prior state-of-the-art methods. We also deploy MFLI in the system and report online experimental results demonstrating improved engagement, less popularity bias, and higher serving efficiency.

Figures & tables

Figure 1. Traditional ANN-based retrieval contains three stages: item embedding learning, offline item indexing, and online ANN search at serving time. In contrast, LI-based retrieval jointly learns item embeddings and indices during training, and performs fast index lookup during serving.

Figure 2. Overview of MFLI framework. During training phase, it learns a multifaceted embedding for each item, and conducts residual quantization on candidate item embeddings by facet in parallel. The codebook and item embeddings are jointly trained via a multifacted loss and a quantized multifaceted loss. The serving part consists of four modules: 1) index lookup, which looks up the multifaceted indices of query item ids; 2) index selection module that selects top

K

indices; 3) item lookup, which fetches all items in each selected index; 4) per-index rerank, which selects top

N

items for each selected index.

Figure 3. Illustration of index rebalancing, where undersized indices are merged, oversized indices are split, and mask items are assigned with an invalid index.

Figure 4. Delta update of MFLI’s indexing structure. Every

\Delta t

time (e.g., 1 min), a delta snapshot storing the indexing structure of fresh items is produced. Every

\Delta T

time (e.g., 30 minutes), a full snapshot storing the refreshed indexing structure of the full item pool is produced. MFLI fetches items from both the full and delta snapshots during serving.

Baseline	Recall				Semantic Relevance
	VVC	Like	LWT	CCD	T1 topic	T2 topic
NeuCF ( He et al., 2017 )	21.51%	24.44%	25.85%	7.54%	24.80%	17.10%
MoL ( Zhai et al., 2023 )	22.67% (+5.39%)	25.78% (+5.48%)	26.84% (+3.83%)	7.60% (+0.80%)	39.65% (+59.88%)	26.84% (+56.96%)
HLLM ( Chen et al., 2024 )	22.39% (+4.09%)	25.28% (+3.44%)	26.61% (+2.94%)	7.97% (+5.70%)	38.94% (+57.02%)	26.48% (+54.85%)
HSTU ( Zhai et al., 2024 )	23.64% (+9.90%)	28.07% (+14.85%)	26.46% (+2.36%)	12.08% (+60.21%)	28.00% (+12.90%)	19.07% (+11.52%)
MTMH ( Zhang et al., 2025 )	22.85% (+6.23%)	26.11% (+6.83%)	27.23% (+5.34%)	8.18% (+8.49%)	46.10% (+85.89%)	33.55% (+96.20%)

Table 1. Performance comparison of MFLI and baseline methods. For recall, we report recall@1000 on VVC (Video View Complete), Like, LWT (Long Watch Time), and CCD (Cold Content Delivery) tasks. For semantic relevance, we report the average topic match rate between the trigger and retrieved items at the

T_{1}

(coarse-grained) and

T_{2}

(fine-grained) levels.

Baseline	Recall				Semantic Relevance
	VVC	Like	LWT	CCD	T1 topic	T2 topic
MFLI	24.04%	26.72%	27.49%	11.86%	52.31%	38.70%
- Split-and-merge	19.60% (-18.47%)	23.05% (-13.74%)	22.65% (-17.61%)	11.32% (-4.55%)	49.35% (-5.66%)	34.49% (-10.88%)
- Index selection	21.61% (-10.11%)	25.15% (-5.88%)	25.30% (-7.97%)	12.50% (+5.40%)	53.39% (+2.06%)	40.17% (+3.80%)
- Delayed start	23.60% (-1.83%)	26.38% (-1.27%)	27.64% (+0.55%)	12.24% (+3.20%)	51.70% (-1.17%)	38.28% (-1.09%)

Table 2. Ablation study results of MFLI. We independently remove split-and-merge index rebalancing, top-k index selection, and delayed start to quantify the impact of each component on MFLI’s recall and semantic relevance. VVC, Like, LWT, and CCD denote the video view complete, like, long watch time, and cold content delivery tasks, respectively.

Panel of Figure 5 — Figure 5. Analysis of learnt indices in MFLI. (a) Index-size distribution (items per index). (b) Item distribution across index-size buckets (items per index-size buckets). (c) Index usage (percentage of non-empty indices) over training snapshots. (d) Intra-index vs. inter-index relevance.

Table 3. Effects of number of facets and codebook size.

Category	Metric	Relative change
Video view	[0, 5k)/[5k, 50k)	+279%/+54%
	[50k, 100k)/[100k, $\infty$ )	+11%/-2%
Freshness	[0, 6h)	+221%
	[0, 24h)	+10%
Engagement	Explicit signals	+0.08%
	Diversity	+0.30%

Table 4. Aggregated 7-day online A/B testing results. MFLI reduces popularity bias by shifting exposure toward low-VV items and improves freshness, engagement efficiency (more explicit and diverse engagement ), and serving throughput.

Num. of facets	Recall				Semantic relevance		Throughput
	VVC	Like	LWT	CCD	$I_{1}$ topic	$I_{2}$ topic
One	20.72% (-13.81%)	24.19% (-9.47%)	24.49% (-10.91%)	10.53% (-11.21%)	45.52% (-12.98%)	31.51% (-18.58%)	+0.0%
Two (default)	24.04%	26.72%	27.49%	11.86%	52.31%	38.70%	-18.8%
Three	24.13% (+0.37%)	27.53% (+3.03%)	28.19% (+2.55%)	12.23% (+3.12%)	50.60% (-3.27%)	37.11% (-4.11%)	-23.9%

Table 5. Effects of number of facets on MFLI’s performance. Note that VVC, Like, LWT, CCD represent video view complete, like, long watch time, cold content delivery tasks respectively.

Codebook size	Recall				Semantic relevance
	VVC	Like	LWT	CCD	$I_{1}$ topic	$I_{2}$ topic
10k	22.54% (-6.24%)	25.51% (-4.53%)	26.71% (-2.84%)	10.82% (-8.77%)	55.10% (+5.33%)	43.10% (+11.37%)
256 $\times$ 128	24.07% (+0.12%)	26.82% (+0.37%)	27.53% (+0.15%)	11.43% (-3.63%)	52.21% (-0.19%)	38.76% (+0.16%)
512 $\times$ 128 (default)	24.04%	26.72%	27.49%	11.86%	52.31%	38.70%
1024 $\times$ 256	23.18% (-3.58%)	25.97% (-2.81%)	27.15% (-1.24%)	11.98% (+1.01%)	52.73% (+0.80%)	39.49% (+2.04%)

Table 6. Effects of codebook size on MFLI’s performance. Note that VVC, Like, LWT, CCD represent video view complete, like, long watch time, cold content delivery tasks respectively.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Name	Value
Number of GPUs	48 A100s
Batch size	2048
Learning rate	0.01
Optimizer	Adagrad
Training epoch	1
Item embedding dim	128

Appendix

Table 7. Hyperparameters of MFLI.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 508 papers

Jul 29, 2026

OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval

In modern recommendation systems, retrieval serves as a primary stage responsible for filtering billions of candidate items down to thousands prior to refined ranking. To make this massive search effective and efficient, the system relies on ranking accuracy and indexing efficiency. However, these two objectives are traditionally misaligned: while the former optimizes for the alignment between ranking predictions and user behavior, the latter optimizes for a structural grouping of item representations which enables fast search among billions of candidates. Thus, despite extensive efforts to scale up interaction modeling for retrieval, they remain fundamentally limited by the structural misalignment between the ranking objectives and the proximity-learned index. In this work, we address this long-standing dichotomy by proposing a new holistic retrieval framework, OneShot. It is an end-to-end, in-model index learning framework that natively aligns index learning with ranking objectives. Using this joint learning as a structural foundation, OneShot pushes the boundaries of retrieval expressiveness by scaling interaction modeling with neural scoring beyond the persistent dot-product bottleneck. OneShot is fully deployed in Instagram's industrial short-video recommendation system, driving significant wins in user daily sessions, engagement, and time-spent. Additionally, OneShot achieves a

20\%

recall gain at the operational ranking volume and a 10x efficiency improvement at an equivalent recall level.

Ziwei Li, Shuyao Li, Xufeng Cai +12

Meta Platforms, Inc. ∗USA, †Singapore

Learning to Rank Recommender Systems

Ask PDF

Aug 25, 2026

Keeping the Index Open: The Recommendation-Side Cost of Shared Search and Recommendation

A shared search-and-recommendation index must score new items from features alone because search has no exploration slot. In a public log covering both surfaces over one catalog,

38.6\%

of held-out query-search impressions show an item never previously shown or visited. For user-cold engagements, the feature-based tower serves this demand without measurable loss against

99

sampled negatives (

0.9595

Recall@20 versus

0.9510

warm). A lexical baseline reaches similar parity, while a full-catalog check remains statistically undecided. Dual-encoder retrieval therefore keeps the index open to new items, unlike an ID-softmax recommender that requires retraining. We price this openness on recommendation against six sequential baselines, each retrained and tuned through five rounds on corrected targets. A float32 timestamp bug had reordered leave-one-out targets for

19.7\%

of users. On MovieLens-1M, warm accuracy trails the strongest retrained baseline by

5.2\%

Recall@20 and

11.4\%

NDCG@20. On MIND, the gap narrows to

0.8

3.6\%

relative to the five strongest baselines, though the model ranks sixth of seven. Under strict zero-leakage cold-start evaluation, the content tower achieves

0.172 \pm 0.006

Recall@20,

1.4\times

the strongest retrained dedicated method (

0.124 \pm 0.007

) and

3\times

a training-free floor, without cold-specific training. Exact full-softmax training raises Recall@20 by

54\%

on MIND-small and

6.9\%

on MovieLens-1M over sampled InfoNCE, but recomputes the full catalog each step and exhausts accelerator memory at

240

K items. Approximate nearest-neighbor search explains none of the remaining gap, serving cost does not regress against ID-softmax retrieval, and a history-window sweep explains half the post-recipe remainder. Exact-quality training at catalog scale remains the open problem.

Theodore Rogers, Joe Standerfer, Dmitrii Timoshenko +3

Amazon Web Services United States

Recommender Systems Cold-Start Recommendation Dense Retrieval Sequential Recommendation

Ask PDF

Jun 3, 2026

Distributional Approximate Nearest Neighbour Search for Uncertainty-Aware Retrieval

Approximate Nearest Neighbour search indices form the backbone of real-world recommender systems, enabling real-time candidate retrieval over million-item catalogues. Typically, a single point estimate embedding is learnt for every user and every item. At serving time, the user embedding queries the index for relevant items. Since these representations are learnt from sparse interaction data, they are noisy and might fail to capture all the nuances that contribute to ``relevance'' -- ignoring the fundamental uncertainty that is inherent to them. The result is a retrieval pipeline that is systematically biased toward the small minority of popular head items with well-estimated embeddings, at the expense of the long-tail majority of niche, diverse, and serendipitous content. We propose DINOSAUR (Distributional Approximate Nearest Neighbour Search for Uncertainty-Aware Retrieval): a simple and infrastructure-compatible framework to incorporate embedding uncertainty into candidate generation. Rather than indexing point estimates, DINOSAUR samples

S_i

embeddings per item and constructs an index on this augmented set. Analogously, at query time, a user embedding is sampled. This two-sided stochastic retrieval process implicitly marginalises over embedding uncertainty, without requiring changes to model architecture or ANN index infrastructure. On the analytical side, we show that DINOSAUR recovers standard point-estimate retrieval as uncertainty vanishes, and we characterise how increased embedding variance expands the regions of latent space in which uncertain items are retrievable. Reproducible empirical observations align with these expectations, showing large coverage gains with small losses in offline recall.

Olivier Jeunen

aampe · Antwerp, Belgium

Long-Tail Recommendation Recommender Systems Approximate Nearest Neighbor Search

Ask PDF

Abstract

Recommender Systems

Explore similar work

OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval

Keeping the Index Open: The Recommendation-Side Cost of Shared Search and Recommendation

Distributional Approximate Nearest Neighbour Search for Uncertainty-Aware Retrieval