🤖 Hybrid AI Chatbot with Semantic Memory & LLM Learning

A production‑style hybrid AI chatbot that combines classical NLP, machine‑learning intent classification, semantic vector memory, and an LLM fallback used strictly as a teacher.

The system is designed to improve over time by learning from real conversations, discovering new intents, and retraining safely — without blindly trusting LLM outputs.

---

🎯 Project Goals

• Build a realistic AI assistant architecture (not a toy chatbot)
• Minimize LLM usage while still benefiting from it
• Enable offline semantic memory and fast local inference
• Create a safe learning loop from conversations
• Follow professional ML + Git practices

---

✨ Key Features

• Intent‑based chatbot (fast, local, inexpensive)
• Semantic vector memory using FAISS (offline & persistent)
• Knowledge base lookup for deterministic answers
• Hinglish → English auto‑translation
• LLM fallback only when the bot fails
• LLM answers saved and reused for training
• Automatic intent discovery from conversations
• Safe retraining pipeline (no blind auto‑learning)
• Memory importance scoring & controlled forgetting

This is a hybrid AI architecture, similar to how real assistants are built in production systems.

---

🧠 How the Chatbot Thinks (Decision Order)

1. Semantic Vector Memory (FAISS)
2. Knowledge Base Lookup
3. Intent Classification Model
4. LLM Fallback (Teacher Mode)

The LLM is never always‑on. It is only used when the system cannot confidently respond.

---

📁 Project Structure

chatbot-ai/
│
├── app/                # Runtime chatbot logic
│   ├── chatbot_core.py
│   ├── vector_memory.py
│   ├── llm_fallback.py
│   ├── knowledge_graph.py
│   └── __init__.py
│
├── training/           # Offline learning & improvement
│   ├── train_chatbot.py
│   ├── discover_intents.py
│   ├── auto_append_intents.py
│   ├── auto_generate_responses.py
│   └── llm_to_intents.py
│
├── data/               # Knowledge & memory (partially git‑ignored)
│   ├── intents.json
│   ├── knowledge.json
│   └── README.md
│
├── model/              # Trained models (generated locally)
│   └── README.md
│
├── .gitignore
├── requirements.txt
└── README.md

---

🛠 Installation

python -m pip install -r requirements.txt

Some components require an additional spaCy model:

python -m spacy download en_core_web_sm

---

🚀 Running the Chatbot

From the project root:

python -m app.chatbot_text

To exit:

quit

---

🧪 LLM Fallback – Teacher Mode

• Activated only when intent + memory + knowledge fail
• Generates a response using an LLM
• Question + answer are stored in data/llm_memory.json
• These examples are later converted into training data

Environment variable required:

setx OPENAI_API_KEY "your_api_key_here"

The LLM teaches the bot, then steps back.

---

📚 Learning Pipeline (After Conversations)

After chatting with the bot, improve it using:

python training/discover_intents.py
python training/auto_append_intents.py
python training/auto_generate_responses.py
python training/train_chatbot.py

What happens in this pipeline

• Unrecognized queries are clustered
• New intents are created automatically
• Safe, neutral responses are generated
• Intent model is retrained
• Future LLM usage decreases

---

🧠 Semantic Memory

• Uses FAISS + sentence‑transformers
• Fully offline after installation
• Persistent across restarts
• Importance‑weighted storage
• Automatic forgetting of low‑value memories

This allows the chatbot to remember concepts, not just exact phrases.

---

🔐 Files Not Committed to Git

Runtime and personal data are excluded for safety:

• chat history
• unrecognized queries
• LLM memory logs
• FAISS index files
• trained model files

This keeps the repository clean and safe to share.

---

📌 Notes

• Designed for learning and experimentation
• Follows real‑world AI system patterns
• Emphasizes safety, reproducibility, and clarity
• LLM improves the bot over time instead of replacing it

---

📄 License

MIT License