Architecture
Understanding RA-H's AI system
Overview
RA-H uses a streamlined AI architecture with two chat modes that share the same tool set but differ in model and capabilities. The system is built around three core concepts:
- Nodes - Your knowledge items (papers, ideas, people, projects)
- Edges - How things connect with explanations
- Dimensions - Flexible tags that emerge from your content
Core Elements
Nodes
Knowledge items stored in the database. Each node can be a paper, idea, person, project, video, tweet, or any other piece of information. Nodes have:
- Title and notes (your thoughts)
- Description (AI-generated grounding)
- Dimensions (multi-tag categorization)
- Metadata (structured information)
- Embeddings (for semantic search)
- Links (to external sources)
- Chunk (full source content)
Edges
Directed relationships between nodes. Edges capture how your knowledge connects with explanations - not just simple links, but contextual relationships like "this paper inspired this idea."
Dimensions
Multi-select categorization tags. Unlike rigid folders, nodes can have multiple dimensions. Some dimensions can be "locked" for auto-assignment during ingestion.
Chat Modes
RA-H uses two chat modes, each optimized for different tasks.
Normal Mode (⚡)
- Model: GPT-5 Mini
- Purpose: Fast responses for everyday tasks
- Best for: Quick searches, simple node creation, fast lookups
Agent Mode (🔥)
- Model: Claude Opus 4.5
- Purpose: Deep reasoning with persistent memory
- Max steps: 25 autonomous tool calls
- Memory: Persists across sessions
- Approval: Write operations require confirmation
Unique capabilities:
- Persistent memory file for cross-session context
- Tool approval gates for safety
- Extended reasoning for complex analysis
Mode Switching
Toggle between modes in the Chat panel. Your choice persists across sessions. Context is maintained when switching mid-conversation.
Tool Approval (Agent Mode)
When Agent mode attempts a write operation:
- Approval modal appears showing the action
- Options: Approve, Reject, or Trust Session
- Auto-reject after 120 seconds
This prevents accidental modifications while enabling autonomous operation.
Context Management
RA-H maintains clean context boundaries:
- Auto-context — Top 10 most-connected nodes included automatically
- Focused node — Current working node with content preview
- Agent memory — (Agent mode only) Persistent file for cross-session context
This architecture ensures fast responses while maintaining the flexibility to handle complex, multi-step tasks.