A RAG-enhanced chatbot with manual fine-tuning capabilities for domain-specific knowledge integration.
SelfSmart addresses a limitation of static LLMs: knowledge cutoff dates. By integrating web crawling with Retrieval-Augmented Generation (RAG), the system can access up-to-date information without retraining. For domain-specific use cases, manual LoRA/DPO fine-tuning scripts allow knowledge to be baked directly into model weights.
Core Thesis: Combine RAG for fresh context access with optional fine-tuning for domain specialization. RAG provides immediate knowledge updates; fine-tuning encodes domain expertise into model weights for efficiency.
The system has 5 phases for knowledge integration. Phases 1-2 (data collection and RAG) are automated. Phases 3-5 (fine-tuning) are manual scripts that require developer execution.
graph TD
subgraph Phase 1: Data Acquisition
Crawler[Web/RSS Crawler] --> JSON[Processed JSON Data]
end
subgraph Phase 2: RAG Pipeline
JSON --> Embed[Sentence Transformers]
Embed --> VectorDB[(ChromaDB)]
VectorDB -.->|Context Injection| Inference
end
subgraph Phase 3: Cognitive Integration
JSON --> Instruct[Instruction Formatting]
Instruct --> LoRA[Supervised Fine-Tuning LoRA]
LoRA --> Merge[Model Merging]
end
subgraph Phase 4: Local Execution
Merge --> MLX[Apple MLX Unified Memory]
MLX --> Inference[Fast Streaming API]
end
subgraph Phase 5: The Data Flywheel
Inference --> User[User UI]
User -.->|Thumbs Up/Down| Feedback[feedback.jsonl]
Feedback --> Gemini[Gemini Synthesis]
Gemini --> Preference[DPO Preference Dataset]
Preference --> DPOTrain[Direct Preference Optimization]
DPOTrain --> MLX
end
classDef primary fill:#2b3137,stroke:#24292e,stroke-width:2px,color:#fff;
classDef secondary fill:#0366d6,stroke:#005cc5,stroke-width:2px,color:#fff;
classDef db fill:#28a745,stroke:#22863a,stroke-width:2px,color:#fff;
class Crawler,JSON,Instruct,LoRA,Merge,MLX,Inference,User,Feedback,Gemini,Preference,DPOTrain primary;
class VectorDB db;
| Phase | Feature | Engineering Details | Automation Status |
|---|---|---|---|
| 1 | Web Crawling | Async data collector pulling HTML/RSS into JSON pipelines via aiohttp and BeautifulSoup4. | Manual script execution |
| 2 | RAG Search Engine | ChromaDB integration with sentence-transformers for semantic retrieval and context-grounding. | Automated in chat flow |
| 3 | SFT LoRA Fine-Tuning | PyTorch peft and trl.SFTTrainer pipelines for domain-specific knowledge integration. | Manual script execution |
| 4 | Apple MLX Inference | mlx-lm for native Apple Silicon Unified Memory execution with streaming SSE responses. | Automated in chat flow |
| 5 | DPO Training | RLHF pipeline using trl.DPOTrainer with Gemini-synthesized preference pairs from user feedback. | Manual script execution |
- Python 3.9 - 3.11
- Node.js 20+
- Apple Silicon Mac (M1/M2/M3) recommended for local MLX inference
git clone https://github.com/genius-0963/SelfSmart.git
cd SelfSmart
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtcp .env.example .envAdd your GEMINI_API_KEY (used for data synthesis and evaluation) inside .env.
# Terminal 1: Backend API
uvicorn src.api.main:app --host 0.0.0.0 --port 8000 --reload
# Terminal 2: Frontend UIcd frontend && npm install && npm run devThe SelfSmart architecture could be adapted for various domain-specific use cases:
Technical Documentation Assistant: Crawl and ingest technical docs, API references, and code repositories. Use RAG for immediate Q&A and fine-tune for domain-specific terminology.
Corporate Knowledge Base: Index internal wikis, Slack conversations, and documentation. RAG provides fresh context; fine-tuning encodes company-specific jargon.
Research Assistant: Digest academic papers and research publications. RAG retrieves relevant papers; fine-tuning adapts to specific research domains.
| Layer | Technology | Purpose |
|---|---|---|
| Inference Engine | Apple MLX (mlx-lm) | High-speed, native Apple Silicon LLM execution |
| Alignment (RLHF) | TRL (DPOTrainer) | Direct Preference Optimization using human feedback |
| Fine-Tuning (SFT) | PEFT (LoRA) | Parameter-efficient weight updates on Kaggle/Local GPUs |
| Vector Store | ChromaDB | Local persistent semantic memory |
| Layer | Technology | Purpose |
|---|---|---|
| Framework | FastAPI | Async REST, Server-Sent Events (SSE) streaming |
| Automation | asyncio | Concurrent web crawling and dataset synthesis |
| Synthesis | Gemini API | Generates missing preference pairs for DPO datasets |
SelfSmart/
├── src/
│ ├── api/main.py # FastAPI application entry point
│ ├── llm_training/
│ │ ├── inference.py # MLX-powered LocalLLMClient
│ │ ├── lora_trainer.py # SFTTrainer Pipeline
│ │ └── dpo_trainer.py # RLHF DPOTrainerManager
│ └── llm/ # RAG logic and Gemini API integrations
│
├── scripts/
│ ├── build_dataset.py # Phase 1: Web Scraper -> Instruction JSON
│ ├── ingest_data.py # Phase 2: Instruction JSON -> ChromaDB RAG
│ ├── build_dpo_dataset.py # Phase 5: Feedback JSONL -> DPO Triplets
│ ├── train_local.py # Local LoRA SFT Execution
│ └── train_dpo.py # Local DPO Execution
│
├── data/ # Persistent SQLite and Feedback logs
├── vector_store/ # ChromaDB storage
├── training_data/ # Raw and Processed JSON datasets
├── models/ # LoRA Checkpoints and Merged Safetensors
│
├── kaggle_train.ipynb # Cloud GPU SFT Notebook
└── kaggle_dpo.ipynb # Cloud GPU DPO Notebook
SelfSmart is designed with a Senior AI/ML Engineering mindset:
| Principle | How It's Applied |
|---|---|
| Hardware Symbiosis | We ditched PyTorch inference on Mac in favor of mlx-lm to tap directly into Apple's Unified Memory, bypassing expensive CUDA constraints. |
| VRAM Optimization | Kaggle notebooks utilize bitsandbytes 4-bit NF4 quantization to squeeze 3.8B parameter models into 16GB T4 instances without OOM crashes. |
| Data Integrity | Instead of manually writing synthetic data, the DPO builder proxies through Gemini to guarantee structurally perfect chosen/rejected pairings. |
| Architectural Agility | The system is highly modular. The LLM engine, RAG database, and UI are fully decoupled, allowing plug-and-play swaps as the ecosystem evolves. |
Contributions are welcome. Please keep PRs focused and ensure they align with the ARCHITECTURE.md.
This project is licensed under the MIT License.
⭐ If SelfSmart helped you envision the future of autonomous AI, give it a star! ⭐
Built with 🧠 and ☕ by Subh | Report a Bug