In the fast-evolving world of AI, large language models<\/a> are pushing boundaries in speed, accuracy, and cost-efficiency. The recent release of Deepseek R1, an open-source model rivaling OpenAI\u2019s o1, is a hot topic in the AI space, especially given its 27x lower cost and superior reasoning capabilities. Pair this with Qdrant\u2019s binary quantization for efficient and quick vector searches, we can index over 1,000+ page documents. In this article, we\u2019ll create a Bhagavad Gita AI Assistant, capable of indexing 1,000+ pages, answering complex queries in seconds using Groq, and delivering insights with domain-specific precision.<\/p>\n This article was published as a part of the\u00a0<\/strong><\/em>Data Science Blogathon.<\/strong><\/em><\/a><\/p>\n Deepseek R1 challenges OpenAI\u2019s dominance with\u00a027x lower API costs\u00a0and near-par performance on reasoning benchmarks. Unlike OpenAI\u2019s o1 closed, subscription-based model ($200\/month), Deepseek R1 is\u00a0free, open-source, and ideal for budget-conscious projects and experimentation.\u00a0<\/p>\n Reasoning- ARC-AGI Benchmark: [Source: ARC-AGI Deepseek<\/a>]\u00a0<\/p>\n From my experience so far, Deepseek does a great job with math reasoning, coding-related use cases, and context-aware prompts.\u00a0However, OpenAI retains an edge in\u00a0general knowledge breadth, making it preferable for fact-diverse applications.<\/p>\n Binary quantization (BQ) is Qdrant\u2019s indexing compression technique to optimize high-dimensional vector storage and retrieval. By converting 32-bit floating-point vectors into\u00a01-bit binary values, it slashes memory usage by\u00a040x\u00a0and accelerates search speeds dramatically.<\/p>\n Avoid binary quantization for\u00a0low-dimension vectors (<\/p>\n Below is the flow chart that explains on how we can build Bhagwad Gita Assistant:<\/p>\n Architecture Overview<\/b><\/p>\n Let us now follow the steps on by one:<\/p>\n Let\u2019s set up the foundation of our RAG pipeline using LlamaIndex. We need to install essential packages including the core LlamaIndex library, Qdrant vector store integration, FastEmbed for embeddings, and Groq for LLM access.<\/p>\n Note:<\/b><\/p>\n Once the installation is done, let\u2019s import the required modules.\u00a0<\/p>\n Here, we handle the crucial task of converting raw text into vector representations. The SimpleDirectoryReader loads documents from a specified folder.<\/p>\n Create a folder, i.e., a data directory, and add all your documents inside it. In our case, we downloaded the Bhagavad Gita document and saved it in the data folder.<\/p>\n You can download the ~900-page Bhagavad Gita document here: iskconmangaluru<\/a><\/p>\n Qdrant\u2019s FastEmbed is a lightweight, fast Python library designed for efficient embedding generation. It supports popular text models and utilizes quantized model weights along with the ONNX Runtime for inference, ensuring high performance without heavy dependencies.<\/p>\n To convert the text chunks into embeddings, we will use Qdrant\u2019s FastEmbed. We process these in batches of 50 documents to manage memory efficiently.\u00a0<\/p>\n Time to configure Qdrant client, our vector database, with optimized settings for performance. We create a collection named \u201cbhagavad-gita\u201d with specific vector parameters and enable binary quantization for efficient storage and retrieval.<\/p>\n There are three ways to use the Qdrant client:\u00a0<\/p>\n Note the collection name needs to be unique, after every data change this needs to be changed as well.\u00a0<\/p>\n We first check if a collection with the specified collection_name exists in Qdrant. If it doesn\u2019t, only then we create a new collection configured to store 1,024-dimensional vectors and use cosine similarity for distance measurement.<\/p>\n We enable on-disk storage for the original vectors and apply binary quantization, which compresses the vectors to reduce memory usage and enhance search speed. The always_ram parameter ensures that the quantized vectors are kept in RAM for faster access.<\/p>\n The indexing process uploads our processed documents and their embeddings to Qdrant in batches. Each document is stored alongside its vector representation, creating a searchable knowledge base.<\/p>\n The GPU will be used at this stage, and depending on the data size, this step may take a few minutes.<\/p>\n The search function takes a user query, converts it to an embedding, and retrieves the most relevant documents from Qdrant based on cosine similarity. We demonstrate this with a sample query about the Bhagavad-g\u012bt\u0101, showing how to access and print the retrieved context.<\/p>\n For RAG it\u2019s important to define the system\u2019s interaction template using ChatPromptTemplate. The template creates a specialized assistant knowledgeable in Bhagavad-gita, capable of understanding multiple languages (English, Hindi, Sanskrit).<\/p>\n It includes structured formatting for context injection and query handling, with clear instructions for handling out-of-context questions.<\/p>\n The final pipeline brings everything together in a cohesive RAG system. It follows the Retrieve-Augment-Generate pattern: retrieving relevant documents, augmenting them with our specialized prompt template, and generating responses using the LLM. Here for LLM we will use Deepseek R-1 distill Llama 70 B hosted on Groq, get your keys from here:\u00a0Groq Console<\/a>.<\/p>\n Now what if you need to use this application again? Are we supposed to undergo all the steps again?<\/p>\n The answer is no.\u00a0<\/p>\n There is not much difference in what you have already written. We will reuse the same search and pipeline function along with the collection name that we need to run the query_points.\u00a0<\/p>\n We will use the same above two functions and message_template in the Streamlit app.py.\u00a0<\/p>\n In Streamlit after every user question, the state is refreshed. To avoid refreshing the entire page again, we will define a few initialization steps under Streamlit cache_resource.\u00a0<\/p>\n Remember when the user enters the question, the FastEmbed will download the model weights just once, the same goes for Groq and Qdrant instantiation.\u00a0<\/p>\n If you noticed the response output, the format is On the UI, I want to keep the reasoning under the Streamlit expander, to retrieve the reasoning part, let\u2019s use string indexing to extract the reasoning and the actual response.\u00a0<\/p>\n Initializes a messages history in Streamlit\u2019s session state. A \u201cClear Chat\u201d button in the sidebar allows users to reset this history.\u00a0<\/p>\n Iterates through stored messages and displays them in a chat-like interface. For assistant responses, it separates the thinking process (shown in an expandable section) from the actual answer using the extract_thinking_and_answer function.<\/p>\n The remaining piece of code is a standard format to define the chatbot component in Streamlit i.e., input handling that creates an input field for user questions. When a question is submitted, it\u2019s displayed and added to the message history. Now it processes the user\u2019s question through the RAG pipeline while showing a loading spinner. The response is split into thinking process and answer components.<\/p>\n By combining\u00a0Deepseek R1\u2019s reasoning,\u00a0Qdrant\u2019s binary quantization, and LlamaIndex\u2019s RAG pipeline, we\u2019ve built an AI assistant that delivers\u00a0sub-2-second responses\u00a0on 1,000+ pages. This project underscores how domain-specific LLMs and optimized vector databases can democratize access to ancient texts while maintaining cost efficiency. As open-source models continue to evolve, the possibilities for niche AI applications are limitless.<\/p>\nLearning Objectives<\/h3>\n
\n
Deepseek R1 vs OpenAI o1<\/h2>\n
\n
What is Binary Quantization in Vector Databases?<\/h2>\n
How It Works<\/h3>\n
\n
Why It Matters<\/h3>\n
\n
Building the Bhagavad Gita Assistant<\/h2>\n
![\"Building](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/02\/Building-the-Bhagavad-Gita-Assistant.webp\")
<\/figure>\n\n
Step-by-Step Implementation<\/h2>\n
Step1: Installation and Initial Setup<\/h3>\n
\n
!pip install llama-index\n!pip install llama-index-vector-stores-qdrant llama-index-embeddings-fastembed\n!pip install llama-index-readers-file\n!pip install llama-index-llms-groq <\/code><\/pre>\nimport logging\nimport sys\nimport os\n\nimport qdrant_client\nfrom qdrant_client import models\n\nfrom llama_index.core import SimpleDirectoryReader\nfrom llama_index.embeddings.fastembed import FastEmbedEmbedding\nfrom llama_index.llms.groq import Groq # deep seek r1 implementation<\/code><\/pre>\nStep2: Document Processing and Embedding\u00a0<\/h3>\n
data = SimpleDirectoryReader(\"data\").load_data()\ntexts = [doc.text for doc in data]\n\nembeddings = []\nBATCH_SIZE = 50<\/code><\/pre>\nembed_model = FastEmbedEmbedding(model_name=\"thenlper\/gte-large\")\n\n\nfor page in range(0, len(texts), BATCH_SIZE):\n page_content = texts[page:page + BATCH_SIZE]\n response = embed_model.get_text_embedding_batch(page_content)\n embeddings.extend(response)<\/code><\/pre>\nStep3: Qdrant Setup with Binary Quantization<\/h3>\n
\n
collection_name = \"bhagavad-gita\"\n\nclient = qdrant_client.QdrantClient(\n #location=\":memory:\",\n url = \"QDRANT_URL\", # replace QDRANT_URL with your endpoint\n api_key = \"QDRANT_API_KEY\", # replace QDRANT_API_KEY with your API keys\n prefer_grpc=True\n)\n<\/code><\/pre>\nif not client.collection_exists(collection_name=collection_name):\n client.create_collection(\n collection_name=collection_name,\n vectors_config=models.VectorParams(size=1024,\n distance=models.Distance.COSINE,\n on_disk=True),\n quantization_config=models.BinaryQuantization(\n binary=models.BinaryQuantizationConfig(\n always_ram=True,\n ),\n ),\n )\nelse:\n print(\"Collection already exists\")<\/code><\/pre>\nStep4: Index the document<\/h3>\n
for idx in range(0, len(texts), BATCH_SIZE):\n docs = texts[idx:idx + BATCH_SIZE]\n embeds = embeddings[idx:idx + BATCH_SIZE]\n\n client.upload_collection(collection_name=collection_name,\n vectors=embeds,\n payload=[{\"context\": context} for context in docs])\n\nclient.update_collection(collection_name= collection_name,\n optimizer_config=models.OptimizersConfigDiff(indexing_threshold=20000)) <\/code><\/pre>\nStep5: RAG Pipeline with Deepseek R1<\/h3>\n
Process-1: R- Retrieve relevant document<\/h4>\n
def search(query,k=5):\n # query = user prompt\n query_embedding = embed_model.get_query_embedding(query)\n result = client.query_points(\n collection_name = collection_name,\n query=query_embedding,\n limit = k\n )\n return result\n \nrelevant_docs = search(\"In Bhagavad-g\u012bt\u0101 who is the person devoted to?\")\n\nprint(relevant_docs.points[4].payload['context'])<\/code><\/pre>\nProcess-2: A- Augmenting prompt<\/h4>\n
from llama_index.core import ChatPromptTemplate\nfrom llama_index.core.llms import ChatMessage, MessageRole\n\nmessage_templates = [\n ChatMessage(\n content=\"\"\"\n You are an expert ancient assistant who is well versed in Bhagavad-gita.\n You are Multilingual, you understand English, Hindi and Sanskrit.\n \n Always structure your response in this format:\n Process-3: G- Generating the response<\/h4>\n
os.environ['GROQ_API_KEY'] = \"GROQ_API_KEY\" # replace with your keys\nllm = Groq(model=\"deepseek-r1-distill-llama-70b\")\n\n\ndef pipeline(query):\n # R - Retriver\n relevant_documents = search(query)\n context = [doc.payload['context'] for doc in relevant_documents.points]\n context = \"\\n\".join(context)\n\n # A - Augment\n chat_template = ChatPromptTemplate(message_templates=message_templates)\n\n # G - Generate\n response = llm.complete(\n chat_template.format(\n context_str=context,\n query=query)\n )\n return response\n \n \nprint(pipeline(\"\"\"what is the PURPORT of O my teacher, behold the great\tarmy of\t the sons of P\u0101\u1e47\u1e0du, so\nexpertly arranged by your intelligent disciple, the son of Drupada.\"\"\"))<\/code><\/pre>\nOutput: (Syntax:
![\"output\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/02\/outputtt.webp\")
<\/figure>\nprint(pipeline(\"\"\"\nJayas\ttu\tp\u0101\u1e47\u1e0du-putr\u0101\u1e47\u0101\u1e41\tye\u1e63\u0101\u1e41\tpak\u1e63e\tjan\u0101rdana\u1e25.\nexplain this gita from translation\n\"\"\"))<\/code><\/pre>\n![\"output\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/02\/Screenshot_2025-02-08_at_15.23.10.webp\")
<\/figure>\nStep6: Saved Index Inference\u00a0<\/h3>\n
client = qdrant_client.QdrantClient(\n url= \"QDRANT_URL\",\n api_key = \"QDRANT_API_KEY\",\n prefer_grpc = True\n )\n \n\n# the search and pipeline code remain the same. \n\ndef search(query, client, embed_model, k=5):\n collection_name = \"bhagavad-gita\"\n query_embedding = embed_model.get_query_embedding(query)\n result = client.query_points(\n collection_name=collection_name,\n query=query_embedding,\n limit=k\n )\n return result\n\ndef pipeline(query, embed_model, llm, client):\n # R - Retriever\n relevant_documents = search(query, client, embed_model)\n context = [doc.payload['context'] for doc in relevant_documents.points]\n context = \"\\n\".join(context)\n\n # A - Augment\n chat_template = ChatPromptTemplate(message_templates=message_templates)\n\n # G - Generate\n response = llm.complete(\n chat_template.format(\n context_str=context,\n query=query)\n )\n return response<\/code><\/pre>\nStep7: Streamlit UI<\/h3>\n
import streamlit as st\nfrom time import sleep\nimport qdrant_client\nfrom qdrant_client import models\nfrom llama_index.core import ChatPromptTemplate\nfrom llama_index.core.llms import ChatMessage, MessageRole\nfrom llama_index.embeddings.fastembed import FastEmbedEmbedding\nfrom llama_index.llms.groq import Groq\nfrom dotenv import load_dotenv\nimport os\n\nload_dotenv()\n\n@st.cache_resource\ndef initialize_models():\n embed_model = FastEmbedEmbedding(model_name=\"thenlper\/gte-large\")\n llm = Groq(model=\"deepseek-r1-distill-llama-70b\")\n client = qdrant_client.QdrantClient(\n url=os.getenv(\"QDRANT_URL\"),\n api_key=os.getenv(\"QDRANT_API_KEY\"),\n prefer_grpc=True\n )\n return embed_model, llm, client\n \nst.title(\"\ud83d\udd49\ufe0f Bhagavad Gita Assistant\")\n# this will run only once, and be saved inside the cache\nembed_model, llm, client = initialize_models() <\/code><\/pre>\ndef extract_thinking_and_answer(response_text):\n \"\"\"Extract thinking process and final answer from response\"\"\"\n try:\n thinking = response_text[response_text.find(\"Chatbot Component<\/h4>\n
def main():\n if \"messages\" not in st.session_state:\n st.session_state.messages = []\n\n with st.sidebar:\n if st.button(\"Clear Chat\"):\n st.session_state.messages = []\n st.rerun()\n\n # Display chat messages\n for message in st.session_state.messages:\n with st.chat_message(message[\"role\"]):\n if message[\"role\"] == \"assistant\":\n thinking, answer = extract_thinking_and_answer(message[\"content\"])\n with st.expander(\"Show thinking process\"):\n st.markdown(thinking)\n st.markdown(answer)\n else:\n st.markdown(message[\"content\"])\n\n # Chat input\n if prompt := st.chat_input(\"Ask your question about the Bhagavad Gita...\"):\n # Display user message\n st.chat_message(\"user\").markdown(prompt)\n st.session_state.messages.append({\"role\": \"user\", \"content\": prompt})\n\n # Generate and display response\n with st.chat_message(\"assistant\"):\n message_placeholder = st.empty()\n with st.spinner(\"Thinking...\"):\n full_response = pipeline(prompt, embed_model, llm, client)\n thinking, answer = extract_thinking_and_answer(full_response.text)\n \n with st.expander(\"Show thinking process\"):\n st.markdown(thinking)\n \n response = \"\"\n for chunk in answer.split():\n response += chunk + \" \"\n message_placeholder.markdown(response + \"\u258c\")\n sleep(0.05)\n \n message_placeholder.markdown(answer)\n \n # Add assistant response to history\n st.session_state.messages.append({\"role\": \"assistant\", \"content\": full_response.text})\n\nif __name__ == \"__main__\":\n main()<\/code><\/pre>\nImportant Links<\/h4>\n
\n
Conclusion<\/h2>\n
Key Takeaways<\/h3>\n
\n
Frequently Asked Questions<\/h2>\n
![\"Tarun](\"https:\/\/av-eks-lekhak.s3.amazonaws.com\/media\/lekhak-profile-images\/converted_image_4h3yIko.webp\")
<\/p>\n