Feature Request: Layered RAG Query Strategy for Wiki Page Generation

[jorschac]

Is your feature request related to a problem?

In our team's practical scenarios, the current single-pass RAG query approach has the following limitations:

1. Single generic query construction: Uses a one-size-fits-all query pattern like "Generate comprehensive wiki page content for {title} focusing on files: {files}", which lacks specificity and focus.

2. Low information density: The generic query often results in overly broad retrieval results that dilute relevant information with less pertinent content.

3. Underutilization of wiki structure metadata: The current approach fails to leverage rich structural information available in the wiki schema, including:

   • Page importance levels (high/medium/low)
   • Section hierarchies and relationships
   • Related pages and dependencies
   • Parent-child page structures

These limitations result in suboptimal context quality for LLM-based page generation, leading to less accurate and less comprehensive wiki content.

Describe the solution you'd like

Based on repetitive tests and validations, we Implemented a Layered RAG Query Strategy that performs multiple focused queries across three strategic dimensions, then merges and deduplicates the results to provide high-quality context for page generation.

Solution Architecture

graph TD
Start[Wiki Page Generation Request] --> Check{Check Page<br/>Importance}
Check --> Layer1[Layer 1: Core Content Query]
Layer1 --> L1Query["Query: Implementation details<br/>and core functionality<br/>in specified files"]
L1Query --> Layer2[Layer 2: Architecture Query]
Layer2 --> Importance{Page<br/>Importance?}
Importance -->|High| L2High["Query: Architecture integration,<br/>dependencies, and<br/>system interactions"]
Importance -->|Medium| L2Med["Query: Integration patterns<br/>and dependencies"]
Importance -->|Low| Skip[Skip Layer 2]
L2High --> Layer3[Layer 3: Context Query]
L2Med --> Layer3
Skip --> Layer3
Layer3 --> L3Query["Query: Context from<br/>- Section relationships<br/>- Related pages<br/>- Parent/child structure"]
L3Query --> Merge[Merge All Retrieved Documents]
Merge --> Dedup[Deduplicate by Content Hash]
Dedup --> Rank[Rank by Document Length]
Rank --> TopN[Select Top 30 Documents]
TopN --> Generate[Generate Wiki Page with<br/>High-Quality Context]
Generate --> Output[Output: Comprehensive<br/>Wiki Page Content]
style Layer1 fill:#e1f5e1
style Layer2 fill:#fff4e1
style Layer3 fill:#e1f0ff
style Merge fill:#ffe1f0
style Output fill:#f0e1ff

Loading

1. Layer 1 - Core Content Query: Focuses on implementation details and core functionality within the page's specified file paths

2. Layer 2 - Architecture Query: Adaptive query based on page importance:

   • High importance: Deep dive into architecture integration, dependencies, and system interactions
   • Medium importance: Focus on integration patterns and dependencies
   • Low importance: Skip to reduce overhead

3. Layer 3 - Context Query: Leverages wiki structure metadata:

   • Section relationships and hierarchies
   • Related pages connections
   • Parent-child page structures

4. Post-Processing Pipeline:

   • Merge documents from all query layers
   • Deduplicate using content-based hashing (first 200 characters)
   • Rank by document length (longer documents typically contain more information)
   • Select top 30 most relevant documents

Expected Benefits

Page Content Accuracy has been significantly improved in our AI Coding Workflows

• Higher information density: Focused queries retrieve more relevant content
• Adaptive resource usage: Query depth scales with page importance
• Better context quality: Multiple perspectives provide comprehensive coverage
• Improved wiki quality: LLM generates more accurate content with better context

Describe alternatives you've considered

no

Additional context

Performance impact: Increases query count from 1 to 2-3 per page (depending on importance), but significantly improves output quality, making the trade-off worthwhile for high-quality documentation generation.