NoVa - A Personalized AI Agent for Contextual and Autonomous Interaction

---

Introduction

NoVa is not just a chatbot. It is an autonomous AI agent designed to dynamically adapt to conversation contexts and respond with intelligence and personality.

Built for advanced interaction on Discord, NoVa integrates:

• Multi-model AI
• Contextual memory
• Real-time web search
• External tool execution
• A unique personality reflecting its creator, Michele Grimaldi

Thanks to integration with Hugging Face Inference API, NoVa can generate adaptive responses while leveraging models like LLaMA 3.2-3B , Google Gemma 2-2B and Mixtral-8x7B-Instruct-v0.1.

Goal:

NoVa is evolving into a fully functional AI agent, capable of executing external tools and interacting beyond simple conversation.

---

Problem and Solution

Challenge:

Most chatbots lack:

• Contextual memory → Forget past messages, leading to disconnected responses.
• A defined personality → Responses are often generic and lack character.
• Real-time adaptability → Limited ability to execute tools or retrieve updated information.

Solution: NoVa Overcomes These Limitations

✔ Conversational Memory → Stores and manages up to 15 messages per user for better contextual responses.
✔ Advanced System Prompt → Ensures a unique, consistent personality across interactions.
✔ Multi-model AI → Supports both LLaMA 3.2-3B , Google Gemma 2-2B and Mixtral-8x7B-Instruct-v0.1, allowing real-time switching.
✔ Real-Time Web Search → Uses SmolAgents to retrieve up-to-date information.
✔ Unique Conversational Style → NoVa is direct, ironic, and adds intelligent sarcasm with a touch of dark humor.

---

Technology and Architecture

Core Technologies

• Language: Python
• AI Libraries: Hugging Face Transformers, LangChain (for future integrations), PyTorch
• AI Models:
  • LLaMA 3.2-3B (Meta)
  • Google Gemma 2-2B
  • Mixtral-8x7B-Instruct-v0.1
• APIs and Frameworks:
  • Hugging Face Inference API → Text generation
  • Discord API → Real-time interaction
  • SmolAgents → External tool execution

Contextual Memory

• NoVa retains up to 15 messages per user.
• This enhances continuity and ensures responses align with previous interactions.
• LAST UPDATES
  A file called memory_tool.py has been integrated, this file has the goal of Automatically storing important information.

Use a simple database, in this case sqlite.

Making the memory more dynamic, using a CodeAgent to generate intelligent queries based on the context of the user's request.
Speeding up the retrieval of information, preventing NoVa from always returning generic answers such as "No relevant memories found."
Optimizing the search system, using a SQL index and selecting only the most relevant information based on the user's request.

How the memory system works

• SQLite database → Saves information in nova_memory.db with user_id, timestamp and message.
• store_memory() function → Saves a new message for NoVa to remember.
• forward() function → Retrieves the most recent information related to the user's query.
• Using CodeAgent → Instead of searching for information with fixed queries, CodeAgent generates dynamic queries to find the most relevant information.

THE FILE HAS CURRENTLY BEEN CREATED, BUT NOT INTEGRATED INTO THE MAIN FILE, bot.py.

Current and Future Features

Implemented Features:

✔ Dynamic AI Model Switching → Change between LLaMA and Gemma using Discord commands (use gemma, use llama, use mistral).
✔ Contextual Memory → Stores recent messages for more coherent conversations. (Still in progress with memory_tool.py)
✔ Advanced Response Management → Text cleaning, filtering, and formatting.
✔ Real-Time Web Search → Uses SmolAgents for up-to-date information retrieval.
✔ Tool Execution → Supports:

• SummarizeTextTool → Summarizes a given text.
• SearchWebTool → Performs DuckDuckGo searches.
• VisitWebpageTool → Fetches and analyzes webpage content.

---

WebVoyager Integration into NoVa

Web Interaction without OpenAI → The main WebVoyager script (run.py) has been rewritten to enable direct interaction with the web, using inference from Hugging Face instead of OpenAI.
New Prompt Handling → The agent now utilizes a structured task system, controlled via task_test.jsonl in the data directory (repo link).
Capabilities:

• Navigate webpages, extract content, and follow links.
• Perform Wikipedia searches and retrieve information.
• Identify interactive elements on webpages.
• Retrieve and summarize content.

Current Limitations & Next Steps

Incomplete Hugging Face Inference Integration → Some tasks fail due to API quota or authentication issues.
Agent Memory Handling Needs Refinement → The way tasks are stored and processed can be improved.
Dynamic Tool Expansion → More AI tools will be integrated for better automation.

Final Goal: Full integration of WebVoyager into NoVa, making it a fully autonomous agent capable of reading, analyzing, and executing tasks on the web.

---

Future Developments:

External API Integration → Expand toolset with APIs for automation and research.
Enhanced Conversational Memory → Implement Retrieval-Augmented Generation (RAG) for better long-term recall.
Voice Interaction → Enable spoken commands and responses for a natural user experience.
Performance Optimization → Reduce response latency and improve model efficiency.
WebVoyager Integration into NoVa

---

Progress shown on LinkedIn post

🔗 LinkedIn Post

---

How NoVa Works on Discord

Context-Aware Conversations

Example: Memory Retention

User: "I had too many chocolate cookies today."
NoVa: "Ah, the eternal struggle between self-control and sugar... Who won?"
User: "Definitely the cookies..."
NoVa: "Another fallen soldier in the battle against cravings. RIP self-discipline."