AI System
PageSpace treats AI not as an isolated chatbot, but as contextual intelligence embedded within the workspace hierarchy. AI conversations are pages, they inherit context from their location, and they participate in the same collaborative, permission-based ecosystem as documents, folders, and channels.
Architecture
Pages as AI Containers
AI conversations are AI_CHAT page types — first-class citizens in the PageSpace ecosystem:
📁 Project Alpha/
├── 📄 Requirements.md
├── 📁 Research/
│ ├── 🤖 Research Assistant ← AI_CHAT page
│ └── 📄 Market Analysis.md
└── 🤖 Project Planning AI ← AI_CHAT page
This means:
- AI conversations inherit permissions from their drive
- AI can reference and understand sibling documents
- AI conversations appear in search, mentions, and navigation
- Multiple AI contexts can exist at different hierarchy levels
Database-First Message Persistence
Every message is immediately persisted to PostgreSQL as an individual row:
chat_messages:
id | pageId | userId | role | content | toolCalls | agentRole
msg-1 | page-123 | user-1 | user | "Analyze this" | NULL | PARTNER
msg-2 | page-123 | NULL | assistant | "Here's my..." | [{...}] | PARTNER
msg-3 | page-123 | user-2 | user | "Also check..." | NULL | PARTNER
This enables:
- Multi-user collaboration — multiple people chat with the same AI
- Real-time sync — all participants see messages as they arrive
- Message attribution — clear record of who said what
- Cross-conversation search — find information across all AI interactions
- Tool call persistence — tool calls and results stored for context and debugging
- Message versioning — support for editing and regeneration via
isActiveflag
Contextual Intelligence
AI conversations understand their position in the workspace hierarchy:
📁 Marketing Campaign/
├── 📄 Brand Guidelines
├── 📄 Target Audience
└── 🤖 Campaign AI ← Sees Brand Guidelines and Target Audience
Context flows upward: an AI can reference parent and sibling pages (with permission). Context is limited by the user's access permissions — AI can only see what you can see.
Multi-Provider Support
PageSpace supports 7 AI providers with a unified interface via the Vercel AI SDK:
| Provider | Key Models | Key Required |
|---|---|---|
| PageSpace | Free models via OpenRouter | No |
| OpenRouter | 200+ models including Claude, GPT, Gemini | User's key |
| Google AI | Gemini 2.5 Pro, Gemini Flash | User's key |
| OpenAI | GPT-5, GPT-4.1, o3 | User's key |
| Anthropic | Claude 4.1 Opus, Claude Sonnet | User's key |
| xAI | Grok 4 | User's key |
| Ollama | Any local model | Self-hosted |
API keys are stored encrypted per-provider. You can set different providers per drive or per individual AI page.
Tool Calling
AI agents have access to 13+ workspace automation tools organized into 6 categories:
| Category | Tools | Capability |
|---|---|---|
| Core Page Ops | list_drives, list_pages, read_page, create_page, rename_page, move_page | Navigate and manage workspace |
| Content Editing | replace_lines | Precise line-based document editing |
| Trash Ops | trash, restore | Soft delete and recovery |
| Search | regex_search, glob_search, multi_drive_search | Pattern and cross-workspace search |
| Task Management | update_task | Create and update tasks on task lists |
| Agent Management | list_agents, multi_drive_list_agents, ask_agent, update_agent_config | Agent discovery and collaboration |
Tools are filtered by agent role and can be customized per page. See Tool Calling for details.
Agent Roles
Three built-in agent roles with different capabilities:
| Role | Read | Write | Delete | Use Case |
|---|---|---|---|---|
| PARTNER | Yes | Yes | Yes | Collaborative AI partner with full capabilities |
| PLANNER | Yes | No | No | Strategic planning — read-only analysis |
| WRITER | Yes | Yes | Yes | Execution-focused — minimal conversation, maximum output |
Each role gets a different system prompt and filtered set of tools. You can further customize by setting enabledTools on individual AI pages.
Real-Time Collaboration
AI messages are broadcast to all conversation participants via Socket.IO:
- User A sends a message → saved to database → broadcast to all users
- AI responds → response streamed to all users in real-time
- User B can continue the conversation immediately
- Tool calls and results are visible to all participants
Model Capability Detection
PageSpace automatically detects model capabilities:
- Vision: Whether a model can process images (checked via static maps and pattern matching)
- Tool calling: Whether a model supports function calling (checked via OpenRouter API for OpenRouter models, pattern-based for others)
If a model doesn't support tools, PageSpace falls back to text-only mode and suggests tool-capable alternatives.
Learn More
- Providers & Models — Detailed provider configuration and model list
- Tool Calling — Complete tool reference with parameters and examples
- Agents — Custom agents, system prompts, and agent-to-agent communication