Recommendation-as-Retrieval Paradigm

• The recommendation-as-retrieval paradigm is an approach that treats personalized recommendation as an information retrieval task by maximizing recall under strict computational and latency constraints.
• It leverages offline augmentation methods like the RADAR framework to precompute high-quality candidate sets, effectively doubling recall metrics and enhancing engagement.
• The paradigm underpins diverse extensions—from conversational and federated recommendation to generative retrieval—unifying search and recommendation with scalable, adaptive architectures.

The recommendation-as-retrieval paradigm frames recommender systems as information retrieval (IR) problems, treating the task of generating personalized suggestions as the problem of efficiently retrieving a high-recall candidate set from a massive item catalog, followed by precise re-ranking. This paradigm is instantiated through numerous architectures and methodologies, unified by the principle of maximizing recall under strict computational and latency constraints. Recent developments, exemplified by the RADAR framework, introduce recall augmentation via offline, deferred, asynchronous retrieval that bridges the recall–precision gap at billion-scale. The paradigm extends to diverse settings—including shift-robust learning, conversational systems, federated and agentic RAG, multi-round adapters, and generative retrieval—grounding most state-of-the-art recommender system advances.

1. Foundations: Multi-Stage Funnel and the Recall Objective

Large-scale recommender systems universally adopt a multi-stage funnel: retrieval, pre-ranking, and final ranking. The first stage addresses $\max_{R}\;\mathrm{Recall}@N$: $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$ subject to latency and CPU constraints. Here, $U$ and $I$ denote the user and item sets (with $|I| \sim 10^9$ in production), $R(u, I; N)$ is the retrieval function producing $N$ candidates per user, $P(u, C; M)$ the pre-ranking step, and $F(u, C, K)$ the final ranking selecting $K$ items. The true relevant set $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$0 is unobserved and only partially approximated via logs.

Traditional retrieval (e.g., two-tower DNN, KNN, ANN) is fundamentally limited by latency, and at billion-scale, can yield only $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$1 candidates and $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$2 Recall@200 with sub-5ms latency (Jaspal et al., 8 Jun 2025). Downstream pre-ranking and ranking operate on ever-smaller, higher-precision candidate pools but inherit the recall bottleneck from the retrieval stage.

2. Offline Augmentation: RADAR Framework and Recall Lifting

The RADAR framework augments the classical paradigm by leveraging offline, asynchronous computation to precompute high-quality candidate sets per user via the full (expensive) ranking model (Jaspal et al., 8 Jun 2025). This introduces a hybrid candidate pool at inference:

• Online, real-time retrievers yield short-term, session-sensitive candidates.
• Async, offline RADAR stores top-$$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$3 fully-ranked items per user as high-recall long-term candidates.

Pipeline:

• Offline: For each user, aggregate a large ($$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$450$$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$5 larger) set from standard retrievers $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$6; score all candidates with $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$7; store the top $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$8 in a key–value store.
• Online: Parallel retrieval from standard sources and from RADAR; pre-rank online set; bypass pre-ranking for RADAR (they are already ranked); merge, deduplicate, and apply final ranking.

This approach effectively doubles recall at a fixed computational envelope (16.5% Recall@200 vs. 8.1% for DNN baseline), with a statistically significant $$\max_{R} \;\mathrm{Recall}@N = \frac{1}{|U|} \sum_{u \in U} \frac{|\,\mathrm{rel}_u \cap R(u, I; N)\,|}{|\mathrm{rel}_u|}$$9 lift in topline engagement in A/B tests. Empirically, the greatest benefit is observed for moderately active users (+17.3% recall lift) (Jaspal et al., 8 Jun 2025).

3. Extensions and Generalizations Across Retrieval Paradigms

The recommendation-as-retrieval view underpins multiple modern extensions:

• Shift-Robust Retrieval and Distillation: The “Retrieval and Distill” paradigm formalizes a temporal invariance theorem, leveraging a shift-free association probability $U$0 estimated from all historical data, and then distills the resulting slow retrieval module into a small, online-efficient neural module. This structure decouples the retrieval of invariant, transferable representations from the fast adaptation to nonstationary labels $U$1 (Zheng et al., 2024).
• Conversational Recommendation as Retrieval: Conversational tasks are cast as retrieval over expanded item documents (metadata + all training conversations recommending the item), using classical BM25 or sparse lexical matchers. Data augmentation with LLM-generated pseudo-dialogues closes recall gaps for cold-start items (Gupta et al., 2023).
• Agentic Multi-Agent RAG: Multi-stage agentic frameworks (such as ARAG) combine embedding-based retrieval with LLM-driven reranking, natural language inference, and user context summarization. These frameworks structure reasoning into user understanding, NLI filtering, context summarization, and final ranking agents, yielding improvements of up to +42% NDCG@5 over vanilla RAG (Maragheh et al., 27 Jun 2025).
• Federated and Hybrid RAG: In privacy-preserving cross-platforms, hybrid ID-based and text-based retrieval ensures the candidate set covers both locally precise and semantically relevant (cross-client) items. An LLM reranker finalizes the ranking, preventing hallucination and bridging distributional shifts (Zeng et al., 2024).
• Multi-Round Adaptive Retrieval: Multi-round adaptive retrieval (Ada-Retrieval) sequentially refines the candidate set, interleaving feedback from earlier rounds to steer subsequent searches via lightweight adapters, outperforming single-shot retrieval by 3–15% on NDCG/HR metrics (Li et al., 2024).
• Promptable Retrieval and Controllable Retrieval: Recent frameworks enable external signals (e.g., regression targets, prompt control) to modulate the retrieval representation at inference time. Two-tower models inject watch-time (CRM) for target-aware candidate sets (Liu et al., 2024); promptable sequential models (DPR) align collaborative-user and semantic-prompt signals via Mixture-of-Experts and curriculum alignment, supporting steerable retrieval (Lyu et al., 21 Feb 2026).

4. Generative Retrieval: Sequence Models and Unified Pipelines

Generative retrieval replaces nearest-neighbor search with autoregressive sequence generation of item identifiers, learning $U$2 directly (e.g., via semantic tokenization or multi-token SIDs). Key frameworks include:

• RankGR: Introduces listwise direct preference optimization (LDPO) for modeling hierarchical partial orders (purchase > click > exposure), alongside a refined scoring phase allowing candidate–sequence attention. This joint optimization yields +29% HR@20 on Taobao over prior state-of-the-art GR (Fu et al., 9 Feb 2026).
• Hybrid Dense/Generative Models: The LIGER approach combines generative retrieval for fast, large-scale drafting and dense retrieval for cold-start and final ranking, enabling recall to approach dense methods while maintaining generative inference efficiency (Yang et al., 2024).
• Unified Generative Architectures: Sequence-to-sequence Transformers jointly optimize retrieval and ranking heads, bridged by enhancer modules (ranking-influenced mining of hard negatives/positives) and adaptive gradient weighting, as in UniGRF. End-to-end synchronization ensures neither retrieval nor ranking dominates convergence, with consistent NDCG/AUC gains (Zhang et al., 23 Apr 2025).
• Scaling Laws: Unified frameworks (OnePiece) empirically confirm power-law reduction in Next-Token Prediction (InfoNCE, SID cross-entropy) with parameter scaling ($U$3), indicating that generative recommendation enjoys the same data/model scaling regimes as LLMs, and hybrid cascades realized via shared encoder–decoder backbones can unify the strengths of ANN and beam-based retrieval (Cao et al., 8 Dec 2025).

5. Data Representation, Product-Space, and Retrieval Objectives

Effective retrieval depends on the choice of representation and scoring:

• Embeddings: Dual-tower, bi-tower, multi-interest, or user–item graph embeddings remain standard for both dense and generative layers (Jaspal et al., 8 Jun 2025, Yang et al., 2024).
• Item and Query Expansion: Expanding item docs with prior conversations (CRS), graph-entity traversal (G-CRS), or review content (RA-Rec) improves lexical/semantic coverage and recall (Gupta et al., 2023, Qiu et al., 9 Mar 2025, Kemper et al., 2024).
• Objective: All paradigms ultimately optimize for maximum Recall@N under resource constraints; some extend to target-aware or multi-objective retrieval by modulating user queries with explicit business signals or prompt-based semantics (Liu et al., 2024, Lyu et al., 21 Feb 2026).

6. Empirical Performance, Deployment, and Trade-Offs

The paradigm demonstrates robust empirical improvements across benchmarks and commercial deployments:

Framework	Recall@200	NDCG/HR/Lift	Latency	Key Setting(s)
RADAR	16.5%	+0.8% A/B Lift	≈unchanged	Billion-scale, multi-stage
CRM	—	+0.323% avg view time	+O(1)	400M users, 50B logs/d
GPT-FedRec	—	+45% Rec@10	× O(1)	Federated, text+ID, LLM
Ada-Retrieval	—	+2–15% NDCG/HR	O(1)	Iterative, sequential
RankGR	—	+29% HR@20	<10ms QPS	0.9T logs, >10k QPS
UniGRF	—	+2.8% NDCG/AUC	O(1)	Unified, seq2seq, large scale

Offline pipelines consume significant compute and storage but run asynchronously. Online overhead is minimal—often a single additional KV fetch or embedding lookup. Deferred, high-precision candidates (e.g., RADAR) or distilled shift-invariant networks (e.g., RAD) are decoupled from online serving, preserving tail-latency SLAs.

7. Broader Implications and Research Trajectories

The recommendation-as-retrieval paradigm has catalyzed recurring advances:

• Separation of recall maximization (efficient, scalable, resource-constrained) from ultimate ranking (precise, possibly high-latency offline/LLM/graph) (Jaspal et al., 8 Jun 2025, Maragheh et al., 27 Jun 2025).
• Unification of recommendation and search through generative (sequence-to-sequence) or task-instruction (prompt) based modeling, leveraging multi-task mutual information regularization (Zhao et al., 9 Apr 2025, Penha et al., 2024).
• Shrinkage of recall–precision trade-off via offline augmentation, multi-modal retrieval, or agentic orchestration over retrieved candidates.
• Embedding and distillation strategies for data shift invariance and memory/computation efficiency in nonstationary, federated, and privacy-sensitive settings (Zheng et al., 2024, Zeng et al., 2024).

Open questions include data- and model-scaling laws, dynamic retrieval under continuous catalog evolution, explainable/reasoned retrieval (rationales), and optimal fusion of semantic, collaborative, and external signals for universal, efficient recall engines.

References:

• (Jaspal et al., 8 Jun 2025) RADAR: Recall Augmentation through Deferred Asynchronous Retrieval
• (Zheng et al., 2024) Retrieval and Distill: A Temporal Data Shift-Free Paradigm for Online Recommendation System
• (Gupta et al., 2023) Conversational Recommendation as Retrieval: A Simple, Strong Baseline
• (Maragheh et al., 27 Jun 2025) ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation
• (Zeng et al., 2024) Federated Recommendation via Hybrid Retrieval Augmented Generation
• (Fu et al., 9 Feb 2026) RankGR: Rank-Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation
• (Li et al., 2024) Ada-Retrieval: An Adaptive Multi-Round Retrieval Paradigm for Sequential Recommendations
• (Liu et al., 2024) CRM: Retrieval Model with Controllable Condition
• (Lyu et al., 21 Feb 2026) Give Users the Wheel: Towards Promptable Recommendation Paradigm
• (Zhang et al., 23 Apr 2025) Killing Two Birds with One Stone: Unifying Retrieval and Ranking with a Single Generative Recommendation Model
• (Yang et al., 2024) Unifying Generative and Dense Retrieval for Sequential Recommendation
• (Zhao et al., 9 Apr 2025) Unifying Search and Recommendation: A Generative Paradigm Inspired by Information Theory
• (Cao et al., 8 Dec 2025) OnePiece: The Great Route to Generative Recommendation -- A Case Study from Tencent Algorithm Competition
• (Qiu et al., 9 Mar 2025) Graph Retrieval-Augmented LLM for Conversational Recommendation Systems
• (Kemper et al., 2024) Retrieval-Augmented Conversational Recommendation with Prompt-based Semi-Structured Natural Language State Tracking