Update README.md

Arabic Law LLM RAG

README.md

@@ -12,3 +12,37 @@ short_description: Arabic LAW RAG custom
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+
+Arabic Legal Demo: NER & RAG
+
+A rough client demo for Arabic law, extracting legal entities and generating insights using NAMAA-Space/gliner_arabic-v2.1 and Qwen/QwQ-32B in a Retrieval-Augmented Generation (RAG) pipeline. Deployed as a Gradio app in a Hugging Face Space, optimized for NVIDIA H200 GPU.
+Features
+
+    NER: Extracts entities (e.g., person, law) from Arabic legal texts using GLiNER.
+    RAG: Retrieves relevant legal context from a mock corpus using FAISS and generates insights with QwQ-32B.
+    UI: Gradio interface for inputting text, specifying entity types, and viewing entities, context, and insights.
+
+Setup
+
+    Hardware: NVIDIA H200 GPU (141GB VRAM) in a custom/enterprise Hugging Face Space.
+    Files:
+        app.py: Gradio app with RAG pipeline.
+        requirements.txt: Dependencies.
+        legal_corpus.json: Mock legal corpus (replace with real data).
+    Run: Push files to a Hugging Face Space and deploy.
+
+Usage
+
+    Enter Arabic legal text (e.g., "المادة ١٠١ من نظام العمل...").
+    Specify entity types (e.g., "person,law").
+    Click "Analyze" to see extracted entities, retrieved context, and legal insight.
+
+Notes
+
+    Replace legal_corpus.json with a real legal dataset (e.g., MoJ).
+    QwQ-32B uses 4-bit AWQ quantization for H200 efficiency.
+    For non-H200 Spaces (e.g., T4), disable QwQ-32B or use heavier quantization.
+
+
+