What is Retrieval-Augmented Generation?
Retrieval-Augmented Generation (RAG) is an AI architecture that combines the power of large language models (LLMs) with external knowledge retrieval. Instead of relying solely on the model’s training data, RAG systems:
- Retrieve relevant information from a knowledge base
- Augment the user’s query with retrieved context
- Generate accurate responses based on both the query and retrieved information
RAG is particularly valuable for domain-specific applications where accuracy and source attribution are critical, such as Islamic knowledge systems.
How DeenPAL Implements RAG
DeenPAL uses RAG to provide accurate Hadith-based responses by ensuring every answer is grounded in authentic Islamic sources (Sahih Al-Bukhari and Sahih Al-Muslim). This approach prevents the LLM from generating unsupported information.
The system is designed to:
- Deliver personalized responses tailored to user queries
- Utilize reliable sources from trusted Hadith collections
- Provide citations with book numbers, Hadith numbers, and chapters
- Offer explanations that connect retrieved Hadiths to user questions
The RAG Pipeline
DeenPAL’s RAG architecture follows a four-stage pipeline:
1. Data Loading
Hadith PDFs are loaded from the data/ directory and processed:
# From loader.py
folder_path = "data/"
loader = PyPDFDirectoryLoader(folder_path)
documents = loader.load()
for doc in documents:
split_source = (doc.metadata['source'].split("/")[-1])
exact_source_with_ext = split_source.split('_', maxsplit=1)[1]
exact_source = exact_source_with_ext.split('.')[0]
doc.metadata = {'source': exact_source}
2. Text Splitting & Embedding
Documents are split into semantic chunks based on Hadith structure:
# From loader.py
pattern = r"(?:Chapter\s\d+:)|(?:Book\s\d+,\sNumber\s\d+:)"
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000,
chunk_overlap=0,
separators=[pattern],
is_separator_regex=True
)
chunks = text_splitter.split_documents(documents)
DeenPAL uses regex-based splitting to preserve each Hadith as a complete semantic unit, rather than arbitrary character-based chunks. This maintains the integrity of each Hadith.
embeddings = HuggingFaceEmbeddings(
model_name='sentence-transformers/all-MiniLM-L6-v2'
)
3. Vector Storage & Retrieval
Embedded chunks are stored in ChromaDB for efficient similarity search:
# From loader.py
persist_directory = 'database/chroma_db'
db = Chroma.from_documents(
documents=chunks,
embedding=embeddings,
persist_directory=persist_directory
)
# From chains.py
retriever = db.as_retriever(
search_type="mmr",
search_kwargs={"k": 4, "fetch_k": 10}
)
4. Response Generation
The LLM receives the user’s question along with retrieved Hadith context:
# From chains.py
llm = ChatOpenAI(
model="deepseek/deepseek-chat-v3-0324:free",
base_url="https://openrouter.ai/api/v1"
)
question_answer_chain = create_stuff_documents_chain(llm, qa_prompt)
rag_chain = create_retrieval_chain(retriever, question_answer_chain)
Benefits of RAG for Hadith Retrieval
1. Accuracy & Attribution
Every response is backed by authentic Hadith sources with proper citations, preventing hallucinations.
2. Up-to-date Knowledge
The system can be updated with new sources without retraining the LLM.
3. Transparency
Users can verify responses by checking the cited Hadith sources.
4. Semantic Understanding
The embedding model understands meaning, not just keywords, allowing for natural language queries.
5. Scalability
New Hadith collections can be added to the vector store without modifying the core system.
DeenPAL uses caching with @st.cache_resource to ensure data loading happens only once per session, preventing redundant processing and improving response times.
Architecture Diagram
┌─────────────────┐
│ User Query │
│ "What does │
│ Islam say..." │
└────────┬────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ RETRIEVAL PHASE │
│ │
│ 1. Query → Embedding Model │
│ (sentence-transformers/all-MiniLM-L6-v2) │
│ │
│ 2. Vector Search in ChromaDB │
│ (MMR: k=4 from fetch_k=10) │
│ │
│ 3. Retrieve Top 4 Diverse Hadiths │
└────────┬────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ GENERATION PHASE │
│ │
│ 1. Inject Retrieved Context into Prompt │
│ │
│ 2. LLM Processes Query + Context │
│ (DeepSeek Chat via OpenRouter) │
│ │
│ 3. Generate Response with Citations │
└────────┬────────────────────────────────────────┘
│
▼
┌─────────────────┐
│ Response: │
│ • Hadiths │
│ • Explanation │
│ • Answer │
└─────────────────┘
