Why Context Modeling Could Replace RAG – Insights from DeepVista CEO Jing Conan Wang

Introduction

Yesterday afternoon I spent two hours talking with DeepVista CEO Jing Conan Wang, which made me rethink the direction of AI development.

Jing previously worked at Google Brain on conversational recommendation and reinforcement learning. In July he published an article titled Context Modeling: The Future of Personalized AI , introducing a new concept—context modeling—to replace the widely used RAG and prompt‑engineering approaches.

During the conversation he highlighted the main problems of current AI agents: traditional methods lack technical barriers, and RAG systems suffer from rigid rules, poor personalization, and limited adaptability.

Core Technology Analysis: What Is Context Modeling?

From Engineering to Modeling – The Fundamental Difference

Context Engineering relies on manually crafted rules such as keyword search or cosine similarity to retrieve relevant content, which brings several obvious issues:

Rules are hard‑coded and inflexible.

All users share the same rule set.

Constant prompt and rule tuning is exhausting.

Customization for different users is impossible.

Context Modeling takes a different approach: a learnable system dynamically generates context, offering benefits such as:

Learning from data and self‑optimizing.

Per‑user customization.

No need for manual parameter tweaking.

Intelligent translation between user and LLM.

Technical Architecture Principles

The idea draws from the two‑stage design of recommendation systems:

Step 1: Fast Filtering

A lightweight model quickly narrows down the content space.

Step 2: Precise Ranking

A more complex algorithm then ranks the candidates for maximum relevance.

Architecture Comparison

Problems with Current RAG Systems

User Query → Rule Retrieval → Document Ranking → LLM Processing → Response Generation

The above pipeline suffers from:

Fixed Thinking : LLM behaves like a "smart but stubborn colleague".

Lack of Flexibility : Larger models become more rigid.

Limited Developer Control : Hard to influence internal reasoning.

Unstable Context Quality : Simple rules may retrieve irrelevant content.

Advantages of Context Modeling

The new architecture inserts a context model that:

User Query → Adaptive Context Model → Context Planning → Core LLM → Intelligent Reply

Key improvements include:

Intelligent Mediator : Understands user intent and LLM behavior.

Dynamic Generation : Creates the most suitable context on the fly.

Continuous Learning : Improves from user interactions.

Personalization : Tailors context per user and scenario.

Dual‑Model Architecture: The Optimal Future Solution

Two specialized models work together:

1. Fast Context Model

Specialized : Dedicated to context retrieval and generation.

Optimization Goal : Extreme speed.

Technical Traits : Lightweight and highly optimized.

Function : Instantly identifies and generates the most relevant context.

2. Powerful Core Model

Specialized : Focused on reasoning, synthesis, and generation.

Optimization Goal : Intelligence and accuracy.

Technical Traits : Large‑scale with complex reasoning capabilities.

Function : Performs deep thinking based on planned context.

This separation solves the speed‑vs‑intelligence trade‑off of single‑model "thinking" architectures.

Practical Case: DeepVista

DeepVista, marketed as an "AI Chief Assistant", applies context modeling in real products. Core features include:

Automatic Context Collection

Multi‑channel integration (email, Slack, meeting notes).

Real‑time synchronization of a context database.

Intelligent filtering of business‑relevant information.

Smart Content Generation

Investor reports generated from accurate data.

Customer communication handling escalations and relationship maintenance.

Strategic emails that advance business goals.

Priority Management

Automatic identification of urgent messages.

Opportunity capture and intelligent attention allocation.

Technical Implementation Highlights

Proactive collection – the system maintains context without user prompts.

Deep business integration – embedded in daily workflows.

Action‑oriented output – generates executable content, not just information.

Efficiency boost – compresses hour‑long tasks into seconds.

Technical Value and Market Opportunities

Core Value

User Experience Improvement

Reduced context‑switching cost.

Higher work efficiency.

Lower cognitive load.

System Performance Optimization

More precise information retrieval.

More relevant content generation.

Reduced compute waste.

Business Value Creation

30% time savings on email handling.

Better decision quality.

New commercial opportunities.

Infrastructure Opportunities

Context modeling represents a universal AI infrastructure need:

Specialized model development for different domains.

System optimization balancing speed and accuracy.

Seamless integration tools.

Platform‑as‑a‑service offerings and API economy.

Developer tools and community ecosystems.

Implementation Strategy & Best Practices

Technical Implementation Path

Assess Current State

Analyze existing context management.

Identify performance bottlenecks and user pain points.

Prioritize improvements.

Progressive Migration

Pilot specific scenarios.

Gradually expand to more use cases while maintaining stability.

Model Training

Collect high‑quality training data.

Establish evaluation metrics.

Continuously optimize model performance.

Success Factors

Speed First : Ensure the context model responds instantly.

Quality Assurance : Generated context must be highly relevant and accurate.

User Experience : Seamlessly integrate into existing workflows.

Continuous Learning : Build feedback loops for ongoing optimization.

Future Outlook

Key evolution directions include model specialization (SQL, Cypher, GraphQL for different data sources), deeper system integration, privacy mechanisms, open‑source tooling, standard APIs, and community‑driven innovation.

Conclusion

Moving from static context engineering to dynamic context modeling is a fundamental shift in AI thinking. Early adopters of this technology will gain a decisive advantage in building truly personalized AI assistants that act proactively rather than reactively.