Smart Online Assistant (Deep Historical Awareness)

An intelligent local AI agent featuring autonomous web search, hybrid memory management, and context-aware query rewriting.

📖 Overview

Smart Online Assistant is a privacy-first, local AI agent powered by Ollama (LLMs) and Playwright. Unlike standard chatbots, it bridges the gap between offline privacy and real-time information.

It implements a RAG (Retrieval-Augmented Generation) workflow that autonomously browses the web when necessary, rewrites search queries based on conversation history, and manages context using a unique Hybrid Memory (Raw/Zip) system.

✨ Key Features

• 🌐 Autonomous Web Search: Uses Playwright to simulate human browsing (DuckDuckGo), handling pagination and anti-scraping to fetch real-time data without API costs.
• 🧠 Hybrid Memory Architecture:
  • Raw Mode: Retains full-text conversation history.
  • Zip Mode: Automatically summarizes past interactions to save context window tokens while retaining key facts.
• 🔄 Context-Aware Intent Analysis: The agent analyzes your intent (analyze_intent) and rewrites search keywords using recent history to ensure search results are relevant.
• ⚡ Local & Private: Powered entirely by local models (e.g., qwen3:30b-instruct). Your data stays on your machine.
• 🛠️ Full Control: Force search (/s), force offline (/n), or switch memory modes instantly via commands.

🚀 Quick Start

Prerequisites

1. Python 3.9+ installed.
2. Ollama installed and running.

Installation

1. Clone the repository (if applicable) or download the source.

2. Install Python dependencies:

   cd hybrid_agent_project
   pip install -r requirements.txt

3. Install Playwright Browser:

   playwright install chromium

4. Prepare the Model: Pull the required model (ensure it matches MODEL_NAME in config.py):

   ollama run qwen3:30b-instruct
   # Wait for the model to download

Running the Assistant

Run the main script:

python main.py

Or run as a module:

python -m hybrid_agent_project.main

💡 Usage

Once started, interact with the agent naturally or use the following commands:

Command	Description
/s <text>	Force Search: Rewrites query based on history + <text> and searches the web.
/n <text>	No Search: Forces the LLM to answer from internal knowledge/memory only.
/raw	Switch memory to Full-text mode.
/zip	Switch memory to Summarized mode (saves tokens).
/clear	Clear all conversation history.
exit / q	Quit the application.

📄 Real-World Examples

Check real_chat.txt for a full log of actual conversations, search processes, and memory compression in action.

Example Output

You [📉 compressed]: /s Bill Gates net worth
🔧 [manual force-search] Assistant rewrites query -> Bill Gates net worth 2026
>> 🌐 Performing web search and simulated pagination...
AI is thinking...
AI: ... (Responds based on live search results)

⚙️ Configuration

Edit config.py to customize the agent's behavior:

• MODEL_NAME: The Ollama model to use (default: qwen3:30b-instruct).
• SUMMARY_MODEL: The model used for compressing memory.
• HIDE_WINDOW: Set to False to watch the browser scrape in real-time.
• HISTORY_LIMIT: Number of turns used for intent analysis.
• MEMORY_FILE: Path to the JSON memory file.

📂 Project Structure

hybrid_agent_project/
├── agent.py              # Core Logic: Combines LLM, Memory, and Search
├── search.py             # Web Scraping: Playwright & DuckDuckGo integration
├── memory.py             # Memory System: JSON handling & Summarization
├── config.py             # Settings: Models, timeouts, search limits
├── main.py               # Entry Point: CLI Loop
├── utils.py              # Helpers: Text processing
├── real_chat.txt         # Log of real usage examples
└── requirements.txt      # Dependencies

⚠️ License / Copyright

Copyright (c) 2026 Feng Simo. All rights reserved.

This code is for demonstration purposes only. You may not use, modify, distribute, or sublicense this code without explicit permission.