faster

README.md

@@ -1,16 +1,22 @@
 ---
-title: Arabic Rag Chat
-emoji: 💬
-colorFrom: yellow
-colorTo: purple
+title: Arabic RAG Chat
+emoji: 📚
+colorFrom: green
+colorTo: blue
 sdk: gradio
-sdk_version: 5.42.0
+sdk_version: 6.5.1
 app_file: app.py
 pinned: false
-hf_oauth: true
-hf_oauth_scopes:
-- inference-api
-short_description: Arabic RAG Chat
+short_description: Arabic PDF Q&A with RAG
 ---
 
-An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
+# Arabic RAG Chat
+
+Upload Arabic PDF documents and ask questions. Powered by:
+- **Embeddings**: Multilingual-E5-Small (GPU-accelerated via ZeroGPU)
+- **LLM**: Cohere Command R
+- **Vector Store**: FAISS (CPU - fast enough for RAG workloads)
+
+## Setup
+
+Add your `COHERE_API_KEY` in Space Settings → Secrets.

app.py

@@ -5,13 +5,13 @@ import spaces
 from langchain_community.document_loaders import PyPDFLoader
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_huggingface import HuggingFaceEmbeddings
-from langchain_community.vectorstores import Chroma
+from langchain_community.vectorstores import FAISS
 from langchain_cohere import ChatCohere
 from langchain_classic.chains import RetrievalQA
 
 # 1. SETUP: DEFINING THE MODELS
-# We use a multilingual embedding model specifically tuned for retrieval
-embedding_model_name = "intfloat/multilingual-e5-large"
+# We use a smaller/faster multilingual embedding model for retrieval
+embedding_model_name = "intfloat/multilingual-e5-small"
 
 # Detect device - use CUDA if available, otherwise CPU
 device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -43,8 +43,8 @@ def process_and_chat(file, query):
         if not api_key:
             return "**Error:** COHERE_API_KEY environment variable is not set."
         
-        # Initialize LLM (Cohere Command R+ via API)
-        llm = ChatCohere(model="command-a-03-2025", temperature=0.3, cohere_api_key=api_key)
+        # Initialize LLM (Cohere Command R - faster model)
+        llm = ChatCohere(model="command-r7b-12-2024", temperature=0.3, cohere_api_key=api_key)
         
         # 2. LOAD & PROCESS DOCUMENT
         # Gradio 6.x returns file path as string directly
@@ -60,10 +60,10 @@ def process_and_chat(file, query):
         )
         texts = text_splitter.split_documents(documents)
         
-        # 3. CREATE VECTOR STORE (In-Memory)
+        # 3. CREATE VECTOR STORE (In-Memory FAISS - faster than Chroma)
         # This turns your Arabic text into searchable vectors
-        db = Chroma.from_documents(texts, get_embeddings())
-        retriever = db.as_retriever(search_kwargs={"k": 5}) # Retrieve top 5 chunks
+        db = FAISS.from_documents(texts, get_embeddings())
+        retriever = db.as_retriever(search_kwargs={"k": 3})  # Retrieve top 3 chunks (faster)
         
         # 4. RAG CHAIN
         qa_chain = RetrievalQA.from_chain_type(
@@ -93,8 +93,8 @@ iface = gr.Interface(
         gr.Textbox(label="Ask a question in Arabic", placeholder="ما هي النقاط الرئيسية في هذا المستند؟")
     ],
     outputs=gr.Markdown(),
-    title="Arabic RAG (Command R+)",
-    description="Upload a PDF and ask questions. Powered by Cohere Command R+ and Multilingual-E5 embeddings."
+    title="Arabic RAG (Command R)",
+    description="Upload a PDF and ask questions. Powered by Cohere Command R and Multilingual-E5-Small embeddings."
 )
 
 iface.launch(share=True)

requirements.txt

@@ -6,6 +6,6 @@ langchain-community
 langchain-text-splitters
 langchain-cohere
 langchain-huggingface
-chromadb
+faiss-cpu
 pypdf
 sentence-transformers