AI代理生产力系统

graph TB
    subgraph "Input Layer"
        TG[Telegram Bot]
        USER[User]
    end
    
    subgraph "Sensory Processing"
        SWITCH{Content Type}
        AUDIO[Audio Transcription]
        IMAGE[Image Analysis]
        DOC[Document Analysis]
        TEXT[Text Input]
        GR[Guardrails<br/>NSFW + Jailbreak]
    end
    
    subgraph "Sensory Memory"
        BUFFER[(Message Buffer<br/>PostgreSQL)]
        WAIT[Wait 3s]
        AGG[Message Aggregator]
    end
    
    subgraph "Cognitive Layer"
        AGENT[Complex Agent<br/>GPT-4.1-mini]
        STM[(Short-term Memory<br/>PostgreSQL)]
        LTM[(Long-term Memory<br/>Vector Store)]
    end
    
    subgraph "Tool Registry"
        THINK[Think Tool]
        CALC[Calculator]
        MCP[MCP Sub-agents]
        SEARCH[Web Search]
    end
    
    subgraph "Output Layer"
        SEND[Telegram Send]
        CLEAN[Buffer Cleanup]
    end
    
    USER -->|Message| TG
    TG --> SWITCH
    SWITCH -->|Text| TEXT
    SWITCH -->|Audio| AUDIO
    SWITCH -->|Image| IMAGE
    SWITCH -->|Document| DOC
    
    TEXT --> GR
    AUDIO --> GR
    IMAGE --> GR
    DOC --> GR
    
    GR -->|Safe| BUFFER
    GR -->|Unsafe| SEND
    
    BUFFER --> WAIT
    WAIT --> AGG
    AGG --> AGENT
    
    AGENT <--> STM
    AGENT <--> LTM
    AGENT <--> THINK
    AGENT <--> CALC
    AGENT <--> MCP
    AGENT <--> SEARCH
    
    AGENT --> SEND
    SEND --> CLEAN
    CLEAN --> BUFFER

graph LR
    INPUT[Input] --> SWITCH{Type?}
    SWITCH -->|text| TEXT[Direct to Guardrails]
    SWITCH -->|voice| VOICE[Get Audio File]
    SWITCH -->|photo| PHOTO[Get Image File]
    SWITCH -->|document| DOC[Get Document File]
    
    VOICE --> TRANS[Transcribe<br/>Gemini 2.0]
    PHOTO --> ANALYZE[Analyze Image<br/>Gemini 2.0]
    DOC --> EXTRACT[Extract Text<br/>Gemini 2.0]
    
    TRANS --> GR[Guardrails]
    ANALYZE --> GR
    EXTRACT --> GR
    TEXT --> GR
    
    GR -->|Pass| BUFFER[(Buffer)]
    GR -->|Fail| REJECT[Send Rejection]

-- 1. Insert into buffer
INSERT INTO message_buffer (chat_id, content, batch_id)
VALUES ($chat_id, $content, NULL);

-- 2. Wait 3 seconds (allows for multiple incoming messages)

-- 3. Atomic marking with batch_id
UPDATE message_buffer
SET batch_id = $execution_id
WHERE chat_id = $chat_id 
  AND batch_id IS NULL
RETURNING content;

-- 4. Aggregation
SELECT STRING_AGG(content, '\n' ORDER BY id) as full_context
FROM message_buffer
WHERE batch_id = $execution_id;

-- 5. Post-processing Cleanup
DELETE FROM message_buffer WHERE batch_id = $execution_id;

graph TB
    subgraph "Agent Core"
        INPUT[User Input] --> THINK[Think Tool<br/>Intent Analysis]
        THINK --> DECISION{Decision Type}
    end
    
    subgraph "Memory Systems"
        STM[(Short-term<br/>PostgreSQL<br/>10 msgs)]
        LTM[(Long-term<br/>Supabase Vector<br/>OpenAI Embeddings)]
    end
    
    subgraph "Tool Registry"
        CALC[Calculator]
        WEB[Web Search<br/>Native GPT-4.1]
        MCP[MCP Sub-agents]
    end
    
    DECISION -->|Retrieval| LTM
    DECISION -->|Action| MCP
    DECISION -->|Compute| CALC
    DECISION -->|Research| WEB
    
    STM -.->|Context| DECISION
    LTM -.->|Memories| DECISION
    
    MCP --> OUTPUT[Response]
    CALC --> OUTPUT
    WEB --> OUTPUT
    LTM --> OUTPUT

🟣 SYSTEM_IDENTITY
🟣 CONTEXT_VARIABLES (date, time, user)
🟣 GLOBAL_CONSTRAINTS (formatting, data integrity)
🟣 DECISION_PROTOCOL (priority order)
🟣 TOOL_REGISTRY (tech specs)
🟣 ORCHESTRATION_PROTOCOL (workflow)
🟣 FEW_SHOT_EXAMPLES

graph LR
    A[New Interaction] --> B[(PostgreSQL<br/>n8n_chat_histories)]
    B --> C{Window Size}
    C -->|Keep| D[Last 10 messages]
    C -->|Archive| E[Long-term Consolidation]
    D --> F[Agent Context]

CREATE TABLE n8n_chat_histories (
    id SERIAL PRIMARY KEY,
    session_id VARCHAR(255),
    message JSONB,
    created_at TIMESTAMP DEFAULT NOW()
);

graph TB
    subgraph "Daily Consolidation (3AM)"
        CRON[Schedule Trigger] --> AGG[Aggregate 24h Messages]
        AGG --> EXTRACT[Information Extractor<br/>Llama 3.3 70B]
    end
    
    subgraph "Extraction Schema"
        EXTRACT --> SCHEMA{Extracted Fields}
        SCHEMA --> T[main_topic]
        SCHEMA --> E[entities]
        SCHEMA --> A[action_taken]
        SCHEMA --> I[relevant_info]
    end
    
    subgraph "Vector Storage"
        SCHEMA --> EMBED[OpenAI Embeddings<br/>text-embedding-3-small<br/>1536 dims]
        EMBED --> VDB[(Supabase pgvector<br/>agent_memory)]
    end
    
    subgraph "Retrieval"
        QUERY[User Query] --> QEMBED[Embed Query]
        QEMBED --> SEARCH[Cosine Similarity]
        VDB --> SEARCH
        SEARCH --> CONTEXT[Top-K Results]
    end

-- 24h Aggregation
SELECT STRING_AGG(message->>'content', E'\n' ORDER BY id) as batch
FROM n8n_chat_histories
WHERE created_at > NOW() - INTERVAL '1 day';

CREATE TABLE agent_memory (
    id BIGSERIAL PRIMARY KEY,
    content TEXT,
    metadata JSONB,
    embedding VECTOR(1536)
);

CREATE INDEX ON agent_memory 
USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 100);

graph TB
    AGENT[Complex Agent] -->|MCP Request| SERVER[MCP Server]
    
    SERVER --> CAL[calendar_agent]
    SERVER --> MAIL[gmail_agent]
    SERVER --> FIN[financial_agent]
    SERVER --> REPORT[financial_report]
    SERVER --> TASK[tasks_agent]
    
    CAL -->|CRUD| GCAL[Google Calendar API]
    MAIL -->|Send/Reply| GMAIL[Gmail API]
    FIN -->|Read/Write| SHEETS[Google Sheets API]
    REPORT -->|Generate Chart| VIZ[Data Visualization]
    TASK -->|CRUD| GTASKS[Google Tasks API]
    
    GCAL --> RESPONSE[MCP Response]
    GMAIL --> RESPONSE
    SHEETS --> RESPONSE
    VIZ --> RESPONSE
    GTASKS --> RESPONSE
    
    RESPONSE --> AGENT

{
  "tool": "sub_agents",
  "params": {
    "agent": "calendar_agent",
    "prompt": "Schedule meeting with Ana on Jan 15th, 2026 at 2 PM",
    "scope": null
  }
}

UPDATE message_buffer
SET batch_id = NULL
WHERE batch_id = '{{ $execution.id }}';

DELETE FROM n8n_chat_histories 
WHERE created_at < NOW() - INTERVAL '30 days';

gantt
    title Execution Timeline (without tools)
    dateFormat SSS
    section Input
    Sensory Processing    :000, 800ms
    Guardrails Check      :800, 400ms
    section Buffer
    Wait Period           :1200, 3000ms
    Message Aggregation   :4200, 200ms
    section Cognitive
    Agent Processing      :4400, 2000ms
    section Output
    Telegram Send         :6400, 300ms

👤 User: "Lunch with Ana tomorrow at 1 PM"
🟣 Mira: [Calls calendar_agent]
       "Done! Scheduled Lunch with Ana 
        for tomorrow (Jan 15) at 1 PM. ✅"

👤 User: "Spent $50 on lunch"
🟣 Mira: "Was this Personal or Business expense?"
👤 User: "Business"
🟣 Mira: [Calls financial_agent]
       "Logged! 💰 $50.00 (Business - Food)"

👤 User: "How much did I spend this month?"
🟣 Mira: "Do you want the Personal or Business report?"
👤 User: "Personal"
🟣 Mira: [Calls financial_report]
       [Sends PNG chart via Telegram]

👤 User: "What did I agree with Carlos in the last meeting?"
🟣 Mira: [Searches long-term memory]
       "In the meeting on Jan 10th, you agreed with Carlos to:
        • Deliver the proposal by Jan 20th
        • Review the cost spreadsheet
        • Next meeting: Jan 25th at 3 PM"