A Survey of Personalization: From RAG to Agent

📝 Paper Summary

RAG-based personalization User-profile based personalization Agentic RAG pipeline

This survey systematically categorizes how personalization is integrated into Retrieval-Augmented Generation (RAG) and agent workflows, framing agents as advanced 'Personalized RAG++' systems with dynamic memory and planning.

Core Problem

Standard RAG and agent systems typically provide generic, one-size-fits-all responses that fail to account for individual user preferences, history, and specific contexts.

Why it matters:

Generic responses in AI systems limit user satisfaction and engagement in applications like healthcare, e-commerce, and education where context is critical
Existing surveys focus on general RAG or generic agents, leaving a gap in understanding how to systematically inject user-specific constraints across the entire pipeline
Personalization is a key step toward Artificial General Intelligence (AGI) by enabling adaptive, context-aware decision-making rather than static information retrieval

Concrete Example: In a medical query, a generic RAG might retrieve standard treatment protocols for 'diabetes'. A personalized system would rewrite the query to include the user's specific age and medical history (e.g., 'diabetes treatment for 70-year-old with heart history'), retrieve documents relevant to those comorbidities, and generate a response adjusting the tone and complexity to the user's medical literacy.

Key Novelty

Unified Taxonomy for Personalized RAG and Agents

Establishes a structural alignment between RAG phases (Rewriting, Retrieval, Generation) and Agent workflows (Understanding, Planning/Execution, Generation), proposing agents as 'Personalized RAG++'
Categorizes personalization into Explicit (user profiles, history) and Implicit (parameters, embeddings) methods across all pipeline stages
Identifies distinct roles for techniques like query rewriting (handling ambiguity) vs. query expansion (handling incompleteness) specifically within personalized contexts

Architecture

Conceptual alignment between the RAG pipeline and Agent workflows

Breakthrough Assessment

7/10

Comprehensive synthesis of a fragmented field. While it is a survey and does not propose a new model, its framing of agents as an evolution of personalized RAG provides a strong conceptual foundation for future work.

⚙️ Technical Details

Problem Definition

Setting: Integrating personalized information 'p' into the standard RAG pipeline stages: Query Processing (Q), Retrieval (R), and Generation (G)

Inputs: User query q, personalized information p (explicit profiles, history, implicit embeddings)

Outputs: Personalized response g* derived from retrieved documents D* and user context

Pipeline Flow

Pre-retrieval: Personalized Query Rewriting / Expansion
Retrieval: Personalized Indexing / Retrieval (Dense/Sparse)
Post-retrieval: Re-ranking / Summarization
Generation: Explicit Prompting / Implicit Fine-tuning

System Modules

Personalized Pre-retrieval

Refine user query using personalized context (p)

Model or implementation: Various (e.g., rewriting models, expansion rules)

Personalized Indexing (Retrieval)

Organize data to facilitate user-specific retrieval

Model or implementation: Embedding models (BERT, DeBERTa) or Knowledge Graphs

Personalized Retriever (Retrieval)

Fetch documents relevant to both query and user profile

Model or implementation: Dense (Bi-encoder), Sparse (BM25), or RL-based

Personalized Generator

Synthesize answer using retrieved docs and user preference

Model or implementation: LLM (GPT, Llama, etc.)

Novel Architectural Elements

Taxonomy mapping Agent workflows (Understanding, Planning, Execution) directly to RAG stages (Rewriting, Retrieval, Generation)
Classification of 'Personalized RAG++' where persistent memory replaces static indexes and tools replace simple retrieval

Comparison to Prior Work

vs. General RAG Surveys (Gao et al., Fan et al.): This work specifically targets personalization across the pipeline rather than general retrieval techniques
vs. Agent Surveys (Li et al., Wang et al.): Focuses on the convergence of Agents and RAG through the lens of personalization, rather than general agent planning/reasoning
vs. Personalized Generation Surveys (Zhang et al., Liu et al.): Extends beyond just generation to include pre-retrieval and retrieval personalization

Limitations

Privacy concerns regarding the storage and usage of sensitive user profiles and history
Long-context limitations in LLMs when ingesting extensive user history
Cold-start problems for new users with no historical interaction data
Lack of standardized benchmarks specifically for 'Personalized RAG' (existing ones are often repurposed recommendation or dialogue datasets)

Reproducibility

Code: https://github.com/Applied-Machine-Learning-Lab/Awesome-Personalized-RAG-Agent

The paper is a survey and does not introduce a single new model with code. However, it provides a curated list of resources at https://github.com/Applied-Machine-Learning-Lab/Awesome-Personalized-RAG-Agent.

📊 Experiments & Results

Evaluation Setup

Review of existing datasets and metrics used in referenced papers (no new experiments performed)

Benchmarks:

LaMP (Language Model Personalization (various tasks))
DuRecDial (Conversational Recommendation)
Persona-Chat (Personalized Dialogue)

Metrics:

Personalization Accuracy
BLEU
ROUGE
Hit Rate
User Satisfaction (often via simulated proxies)
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Personalization in RAG is converging with Agentic workflows, where retrieval maps to planning and generation maps to execution
A key distinction in pre-retrieval is between Query Rewriting (for ambiguity) and Query Expansion (for incompleteness/broadening)
Implicit personalization (via parameters like LoRA) offers fine-grained adaptation but higher training costs compared to Explicit personalization (via prompting)
Future directions include better privacy preservation, handling long-term memory more efficiently, and developing unified benchmarks for personalized RAG agents

📚 Prerequisite Knowledge

Prerequisites

Understanding of the standard RAG pipeline (Indexing, Retrieval, Generation)
Basic knowledge of LLM-based Agents (Planning, Memory, Tool Use)
Familiarity with embedding models and vector databases

Key Terms

RAG: Retrieval-Augmented Generation—AI systems that answer questions by first searching for relevant documents

PEFT: Parameter-Efficient Fine-Tuning—techniques like LoRA to adapt models with minimal parameter updates

LoRA: Low-Rank Adaptation—a PEFT method that injects trainable low-rank matrices into frozen model weights

RLHF: Reinforcement Learning from Human Feedback—training models to align with human preferences using rewards

dense retrieval: Finding relevant documents by comparing learned vector representations (embeddings) rather than keyword matching

sparse retrieval: Keyword-based retrieval methods like BM25 that match exact terms

query rewriting: Reformulating a query to improve clarity or intent matching (e.g., fixing ambiguity)

query expansion: Adding related terms or synonyms to a query to broaden search coverage

AGI: Artificial General Intelligence—hypothetical AI with human-like understanding across diverse tasks

CoT: Chain-of-Thought—prompting technique encouraging models to generate intermediate reasoning steps

knowledge graph: A structured representation of knowledge using entities (nodes) and relationships (edges)