Aller au contenu principal

Artificial Intelligence

Philosophy: AI in Synap is not just a chatbot; it is a collaborator that lives alongside your data, understanding context and capable of taking action.

1. The "Intelligence Service"​

Synap uses a dedicated microservice called the Intelligence Service to handle all AI operations. This separation ensures:

  • Scalability: Heavy inference tasks don't block the main application.
  • Security: The "Brain" interacts with the "Body" (Backend) through strict protocols.
  • Modularity: Different AI models or providers can be swapped without changing the core app.

2. Architecture: The 3-Protocol System​

We define three distinct ways for AI to interact with Synap:

Protocol 1: Hub Protocol (Internal)​

  • Purpose: The "Spine" of the system. Secure, high-bandwidth communication between Synap Backend and the Intelligence Service.
  • Mechanism: A specialized tRPC client with identifying scopes (hub-protocol.read, hub-protocol.write).
  • Capabilities:
    • Read full Thread Context (history, linked entities).
    • Search the entire Knowledge Graph.
    • Create/Update Entities and Documents.
    • Manage Branches.

Protocol 2: MCP Server (Tools)​

  • Purpose: To allow external AI tools (like Claude Desktop or generic MCP clients) to access Synap data.
  • Standard: Implements the Model Context Protocol (MCP).
  • Capabilities: "Read-only" or safe actions exposed as secure tools.

Protocol 3: A2A (Agent-to-Agent) [Future]​

  • Purpose: To allow autonomous agents from other systems to negotiate and collaborate with Synap agents.

3. Core Capabilities (MVP)​

1. Context-Aware Chat​

The AI doesn't just see your last message. It sees:

  • Conversation History: The full thread.
  • Linked Context: Entities or Documents you explicitly attached to the thread.
  • User Memory (Mem0): It remembers your preferences and past facts (via Mem0 integration).

2. Reactive Actions​

The AI can perform actions, but (for the MVP) it is reactive—it acts only when you ask.

  • Search: "Find that note about the Q3 roadmap."
  • Create: "Make a task for the design review."
  • Update: "Add a 'High Priority' tag to this project."

3. Proposals & Diffs​

When the AI wants to make significant changes (like rewriting a document/code), it doesn't just overwrite.

  1. Proposal: The AI submits a "Proposal" (e.g., a Document Diff).
  2. Review: You see a visual diff (Before vs. After).
  3. Decision: You click "Approve" to commit the change, or "Reject" to discard.

4. Technical Flow​

  1. User Message: "Refactor utils.ts to be cleaner."
  2. Backend: Stores message, forwards to Intelligence Service via Redis/Queue.
  3. Intelligence Service:
    • Fetches Thread Context via Hub Protocol.
    • Injects "Linked Documents" (utils.ts) into the System Prompt.
    • Calls LLM (e.g., Claude 3.5 Sonnet).
  4. Action: LLM decides to call update_document.
  5. Hub Protocol: Sends createDocumentProposal to Backend.
  6. Frontend: Real-time update shows a "Review Proposal" card in the chat.
Learn more on the website