📚 A semantic legal assistant built with Retrieval-Augmented Generation (RAG) to simplify access to Indian tax law using open-source LLMs and modern NLP tools.
- 🔍 Semantic Retrieval using FAISS and MiniLM-SBERT embeddings
- 🔁 Reranking with Cohere for improved context precision
- 🧠 Local Answer Generation using Meta’s LLaMA 2–7B
- 📊 Model Evaluation using BLEU, ROUGE, METEOR, BERTScore, and QuESTEval
- 💡 User-Controlled Temperature for tuning creativity
- 🖥️ Built with Streamlit for an interactive chat UI
| Component | Tool/Library |
|---|---|
| Embeddings | all-MiniLM-L6-v2 (SBERT) |
| Vector Store | FAISS |
| Reranker | Cohere Rerank API |
| LLM | LLaMA 2–7B (4-bit quantized, local) |
| Frontend | Streamlit |
| Evaluation | BLEU, ROUGE, METEOR, BERTScore, QuESTEval |
git clone https://github.com/akasha456/TaxMate.git
cd TaxMate
pip install -r requirements.txtflowchart TD
A[User Query] --> B[Embed Query]
B --> C[Retrieve Top-k using FAISS]
C --> D[Rerank with Cohere]
D --> E[Construct Prompt with Context]
E --> F[Generate Answer with LLaMA 2]
F --> G[Display Answer in Streamlit UI]
| Model | BLEU | METEOR | ROUGE-1 | BERTScore F1 |
|---|---|---|---|---|
| Mistral AI | 13.97 | 39.99 | 49.19 | 89.16 |
| LLaMA 2–7B | 7.98 | 41.83 | 36.83 | 87.06 |
| Model | Cosine MRR | Recall@5 |
|---|---|---|
| MiniLM-SBERT | 1.0000 | 1.0000 |
| BERT | 0.8333 | 1.0000 |
| RoBERTa | 0.7500 | 1.0000 |
- 🧑⚖️ Expand to other domains like GST or Labor Law
- 📲 Web + Mobile deployment
- 🔔 Integrate real-time updates from government portals
- 🗣️ Voice input/output for accessibility
- 🤖 Fine-tuned LLM on Indian legal corpus
This project is licensed under the MIT License.
- Meta AI for LLaMA 2
- Cohere for reranking API
- SentenceTransformers for embeddings
- Streamlit for the frontend
