AppFlowy AI: Intelligent Workspace Orchestration 🚀

Integrating local-first productivity with powerful LLM orchestration. This project embeds an AI Assistant directly into AppFlowy to automate database creation, data structuring, and workspace management.

---

🎯 Project Goal

The goal is to create a seamless AI-to-Workspace bridge where users can describe their data needs in natural language and have AppFlowy automatically configure the appropriate views, properties, and data structures locally.

Key Objectives:

• Local-First AI: All intelligence runs locally via Ollama, preserving privacy.
• Automated Structuring: Intelligent detection of property types (Text, Number, Checkbox, etc.).
• Clean Workspace: Automated cleanup of default boilerplate properties.
• Natural Interaction: A persistent AI overlay that understands workspace context.

---

🏗️ Architecture Stack

The system is built on a decoupled architecture for maximum flexibility and local performance.

graph TD
    User((User)) -->|Prompt| Flutter[AppFlowy Flutter Frontend]
    Flutter -->|HTTP/JSON| FastAPI[Python AI Orchestrator]
    FastAPI -->|LangChain| LLM[Ollama: gpt-oss:20b]
    LLM -->|Parsed Action| FastAPI
    FastAPI -->|Success/Error| Flutter
    Flutter -->|Backend SDK| Rust[AppFlowy Rust Core]
    Rust -->|Storage| DB[(Local Collab DB)]

Loading

📦 Components:

• Frontend (Dart/Flutter): The user interface hosting the AiChatOverlay and AiAgentClient.
• Orchestrator (Python/FastAPI): A robust agent logic layer using LangChain to parse intent and resolve context.
• Engine (Rust): AppFlowy's high-performance core that handles CRDTs and data persistence.

---

📈 Progress & Implementation

Phase 1: AI Chat Integration ✅

• Implemented a non-blocking AiChatOverlay widget.
• Created AiAgentClient for secure local communication with the orchestrator.

Phase 2: Database Automation ✅

• Primary Field Mapping: Automatically renames the default "Name" field to "ID" for cleaner data modeling.
• Smart Cleanup: Automatically deletes default "Type" and "Done" columns upon AI creation.
• Prop Type Detection: Real-time detection of data types (e.g., "age" as number, "lead" as text).
• Duplicate Prevention: Resolved issues where "name" properties were incorrectly skipped.

Phase 3: Reliability & Performance ✅

• Improved database name extraction from various prompt formats.
• Migrated to gpt-oss:20b for superior reasoning capabilities.
• Created parallel startup orchestration via start.sh.

---

🚀 Getting Started

1. Prerequisites

• Ollama: Download & Install
• Model: Pull the research model used:

  ollama pull gpt-oss:20b

2. Installation

Clone the repository and run the setup:

# Install Python dependencies
python3 -m venv venv
source venv/bin/activate
pip install -r agent/requirements.txt

# Fetch Flutter dependencies
cd appflowy/frontend/appflowy_flutter
flutter pub get

3. Execution

Launch both the AI service and the AppFlowy application with a single command:

./start.sh

---

🛠️ Project Structure

Directory	Description
agent/	Python FastAPI Orchestrator & LLM Logic
appflowy/	Core AppFlowy source code (Rust + Flutter)
appflowy_flutter/	Main frontend application
lib/ai/	Dart implementation of AI Agent client and handlers

---

Tip

Try this prompt: "Create a CRM database with name, company, email, and deal_value properties." The AI will automatically set deal_value to a Number type and name as a Text property while setting the primary field to ID.

---

Built with ❤️ for the Open Source Community.