Build RAG Chatbot with Haystack, Pgvector, NVIDIA Llama 3 70B Instruct, and jina-colbert-v2

Introduction to RAG

Retrieval-Augmented Generation (RAG) is a game-changer for GenAI applications, especially in conversational AI. It combines the power of pre-trained large language models (LLMs) like OpenAI’s GPT with external knowledge sources stored in vector databases such as Milvus and Zilliz Cloud, allowing for more accurate, contextually relevant, and up-to-date response generation. A RAG pipeline usually consists of four basic components: a vector database, an embedding model, an LLM, and a framework.

Key Components We'll Use for This RAG Chatbot

This tutorial shows you how to build a simple RAG chatbot in Python using the following components:

Haystack: An open-source Python framework designed for building production-ready NLP applications, particularly question answering and semantic search systems. Haystack excels at retrieving information from large document collections through its modular architecture that combines retrieval and reader components. Ideal for developers creating search applications, chatbots, and knowledge management systems that require efficient document processing and accurate information extraction from unstructured text.
Pgvector: an open-source extension for PostgreSQL that enables efficient storage and querying of high-dimensional vector data, essential for machine learning and AI applications. Designed to handle embeddings, it supports fast approximate nearest neighbor (ANN) searches using algorithms like HNSW and IVFFlat. Since it is just a vector search add-on to traditional search rather than a purpose-built vector database, it lacks scalability and availability and many other advanced features required by enterprise-level applications. Therefore, if you prefer a much more scalable solution or hate to manage your own infrastructure, we recommend using Zilliz Cloud, which is a fully managed vector database service built on the open-source Milvus and offers a free tier supporting up to 1 million vectors.)
NVIDIA Llama 3 70B Instruct: A high-performance AI model optimized by NVIDIA for complex instruction-following tasks, combining Meta's Llama 3 70B architecture with NVIDIA’s hardware-accelerated efficiency. Strengths include rapid inference, scalability on GPUs, and nuanced context understanding. Ideal for enterprise-grade chatbots, technical support automation, and data-driven decision-making in resource-intensive environments.
Jina-ColBERT-v2: A dense passage retrieval model optimized for semantic search and document ranking. It combines ColBERT's contextualized late interaction with efficient indexing, delivering high accuracy in understanding query intent and matching relevant text. Ideal for large-scale enterprise search, Q&A systems, and content discovery platforms requiring nuanced semantic understanding and rapid retrieval.

By the end of this tutorial, you’ll have a functional chatbot capable of answering questions based on a custom knowledge base.

Note: Since we may use proprietary models in our tutorials, make sure you have the required API key beforehand.

Step 1: Install and Set Up Haystack

import os
import requests

from haystack import Pipeline
from haystack.components.converters import MarkdownToDocument

from haystack.components.preprocessors import DocumentSplitter
from haystack.components.writers import DocumentWriter

Step 2: Install and Set Up NVIDIA Llama 3 70B Instruct

To start using models self-hosted with NVIDIA, we need to install the nvidia-haystack package first.

pip install nvidia-haystack

To use LLMs with NVIDIA, you need to specify the correct api_url and your API key. You can get your API key directly from the catalog website. You also need to get an NVIDIA API key to build this pipeline. Here, we will use the NVIDIA_API_KEY environment variable by default. Otherwise, you can pass an API key at initialization with api_key, as in the following example.

from haystack.utils.auth import Secret
from haystack_integrations.components.generators.nvidia import NvidiaGenerator

generator = NvidiaGenerator(
    model="meta/llama3-70b-instruct",
    api_url="https://integrate.api.nvidia.com/v1",
    api_key=Secret.from_token("<your-api-key>"),
    model_arguments={
        "temperature": 0.2,
        "top_p": 0.7,
        "max_tokens": 1024,
    },
)
generator.warm_up()

Step 3: Install and Set Up jina-colbert-v2

pip install jina-haystack

from haystack_integrations.components.embedders.jina import JinaTextEmbedder
from haystack_integrations.components.embedders.jina import JinaDocumentEmbedder

text_embedder = JinaTextEmbedder(api_key=Secret.from_token("<your-api-key>"), model="jina-colbert-v2")
document_embedder = JinaDocumentEmbedder(api_key=Secret.from_token("<your-api-key>"), model="jina-colbert-v2")

Step 4: Install and Set Up Pgvector

To quickly set up a PostgreSQL database with pgvector, you can use Docker:

docker run -d -p 5432:5432 -e POSTGRES_USER=postgres -e POSTGRES_PASSWORD=postgres -e POSTGRES_DB=postgres ankane/pgvector

To use pgvector with Haystack, install the pgvector-haystack integration:

pip install pgvector-haystack

import os
from haystack_integrations.document_stores.pgvector import PgvectorDocumentStore
from haystack_integrations.components.retrievers.pgvector import PgvectorEmbeddingRetriever

os.environ["PG_CONN_STR"] = "postgresql://postgres:postgres@localhost:5432/postgres"

document_store = PgvectorDocumentStore()
retriever = PgvectorEmbeddingRetriever(document_store=document_store)

Step 5: Build a RAG Chatbot

Now that you’ve set up all components, let’s start to build a simple chatbot. We’ll use the Milvus introduction doc as a private knowledge base. You can replace it your own dataset to customize your RAG chatbot.

url = 'https://raw.githubusercontent.com/milvus-io/milvus-docs/refs/heads/v2.5.x/site/en/about/overview.md'
example_file = 'example_file.md'
response = requests.get(url)
with open(example_file, 'wb') as f:
    f.write(response.content)
file_paths = [example_file]  # You can replace it with your own file paths.

indexing_pipeline = Pipeline()
indexing_pipeline.add_component("converter", MarkdownToDocument())
indexing_pipeline.add_component("splitter", DocumentSplitter(split_by="sentence", split_length=2))
indexing_pipeline.add_component("embedder", document_embedder)
indexing_pipeline.add_component("writer", DocumentWriter(document_store))
indexing_pipeline.connect("converter", "splitter")
indexing_pipeline.connect("splitter", "embedder")
indexing_pipeline.connect("embedder", "writer")
indexing_pipeline.run({"converter": {"sources": file_paths}})

# print("Number of documents:", document_store.count_documents())
  
question = "What is Milvus?"  # You can replace it with your own question.

retrieval_pipeline = Pipeline()
retrieval_pipeline.add_component("embedder", text_embedder)
retrieval_pipeline.add_component("retriever", retriever)
retrieval_pipeline.connect("embedder", "retriever")

retrieval_results = retrieval_pipeline.run({"embedder": {"text": question}})

# for doc in retrieval_results["retriever"]["documents"]:
#     print(doc.content)
#     print("-" * 10)

from haystack.utils import Secret
from haystack.components.builders import PromptBuilder

retriever = PgvectorEmbeddingRetriever(document_store=document_store) 
 text_embedder = JinaTextEmbedder(api_key=Secret.from_token("<your-api-key>"), model="jina-colbert-v2")

prompt_template = """Answer the following query based on the provided context. If the context does
                     not include an answer, reply with 'I don't know'.\n
                     Query: {{query}}
                     Documents:
                     {% for doc in documents %}
                        {{ doc.content }}
                     {% endfor %}
                     Answer: 
                  """

rag_pipeline = Pipeline()
rag_pipeline.add_component("text_embedder", text_embedder)
rag_pipeline.add_component("retriever", retriever)
rag_pipeline.add_component("prompt_builder", PromptBuilder(template=prompt_template))
rag_pipeline.add_component("generator", generator)
rag_pipeline.connect("text_embedder.embedding", "retriever.query_embedding")
rag_pipeline.connect("retriever.documents", "prompt_builder.documents")
rag_pipeline.connect("prompt_builder", "generator")

results = rag_pipeline.run({"text_embedder": {"text": question}, "prompt_builder": {"query": question},})
print('RAG answer:\n', results["generator"]["replies"][0])

Optimization Tips

As you build your RAG system, optimization is key to ensuring peak performance and efficiency. While setting up the components is an essential first step, fine-tuning each one will help you create a solution that works even better and scales seamlessly. In this section, we’ll share some practical tips for optimizing all these components, giving you the edge to build smarter, faster, and more responsive RAG applications.

Haystack optimization tips

To optimize Haystack in a RAG setup, ensure you use an efficient retriever like FAISS or Milvus for scalable and fast similarity searches. Fine-tune your document store settings, such as indexing strategies and storage backends, to balance speed and accuracy. Use batch processing for embedding generation to reduce latency and optimize API calls. Leverage Haystack's pipeline caching to avoid redundant computations, especially for frequently queried documents. Tune your reader model by selecting a lightweight yet accurate transformer-based model like DistilBERT to speed up response times. Implement query rewriting or filtering techniques to enhance retrieval quality, ensuring the most relevant documents are retrieved for generation. Finally, monitor system performance with Haystack’s built-in evaluation tools to iteratively refine your setup based on real-world query performance.

pgvector optimization tips

To optimize pgvector in a Retrieval-Augmented Generation (RAG) setup, consider indexing your vectors using GiST or IVFFlat to significantly speed up search queries and improve retrieval performance. Make sure to leverage parallelization for query execution, allowing multiple queries to be processed simultaneously, especially for large datasets. Optimize memory usage by tuning the vector storage size and using compressed embeddings where possible. To further enhance query speed, implement pre-filtering techniques to narrow down search space before querying. Regularly rebuild indexes to ensure they are up to date with any new data. Fine-tune vectorization models to reduce dimensionality without sacrificing accuracy, thus improving both storage efficiency and retrieval times. Finally, manage resource allocation carefully, utilizing horizontal scaling for larger datasets and offloading intensive operations to dedicated processing units to maintain responsiveness during high-traffic periods.

NVIDIA Llama 3 70B Instruct optimization tips

Optimize inference speed by leveraging model quantization (e.g., 16-bit or 8-bit) to reduce memory usage without significant accuracy loss. Use NVIDIA’s TensorRT-LLM for kernel fusion and efficient GPU utilization, and enable dynamic batching to process multiple queries concurrently. Fine-tune retrieval relevance thresholds to balance precision and recall, minimizing unnecessary context. Cache frequent retrieval results and precompute embeddings. Profile memory usage to avoid bottlenecks, and employ mixed-precision training if fine-tuning. Regularly update drivers and libraries (e.g., CUDA, PyTorch) to leverage hardware acceleration and software optimizations.

jina-colbert-v2 optimization tips

To optimize jina-colbert-v2 in a RAG setup, ensure input text is preprocessed by truncating or chunking documents to fit its 512-token limit, preserving context. Use batch inference for dense embeddings to maximize GPU utilization. Fine-tune on domain-specific data to improve relevance. Leverage ColBERT’s late interaction by precomputing document embeddings and caching them for faster retrieval. Adjust the compression ratio for query-document token tensors to balance speed and accuracy. Filter irrelevant documents early using metadata to reduce computational overhead. Monitor retrieval latency and accuracy to iteratively refine parameters.

By implementing these tips across your components, you'll be able to enhance the performance and functionality of your RAG system, ensuring it’s optimized for both speed and accuracy. Keep testing, iterating, and refining your setup to stay ahead in the ever-evolving world of AI development.

RAG Cost Calculator: A Free Tool to Calculate Your Cost in Seconds

Estimating the cost of a Retrieval-Augmented Generation (RAG) pipeline involves analyzing expenses across vector storage, compute resources, and API usage. Key cost drivers include vector database queries, embedding generation, and LLM inference.

RAG Cost Calculator is a free tool that quickly estimates the cost of building a RAG pipeline, including chunking, embedding, vector storage/search, and LLM generation. It also helps you identify cost-saving opportunities and achieve up to 10x cost reduction on vector databases with the serverless option.

Calculate your RAG cost now.

Calculate your RAG cost

What Have You Learned?

By now, you’ve unlocked the magic of building a powerful RAG system from the ground up! This tutorial showed you how to weave together four game-changing components: Haystack as your orchestration framework, PgVector as your lightning-fast vector database, NVIDIA Llama 3 70B Instruct as your powerhouse LLM, and jina-colbert-v2 as your precision-focused embedding model. You learned how Haystack acts as the glue, connecting your data pipeline to retrieval and generation steps, while PgVector efficiently stores and retrieves context-rich embeddings generated by jina-colbert-v2’s semantic understanding. Then, Llama 3 70B steps in to craft human-like responses that feel grounded and relevant. Along the way, you saw how tweaking chunk sizes, optimizing indexing strategies, or even using hybrid search can supercharge performance—and let’s not forget that free RAG cost calculator to keep your projects budget-friendly!

But this is just the beginning. You’re now equipped with the tools to turn raw data into intelligent, responsive applications that feel almost alive. Imagine the possibilities: chatbots that truly get your users, research assistants that dig through mountains of data in seconds, or creative tools that spark new ideas. The tutorial didn’t just hand you a blueprint—it gave you a launchpad. So go ahead! Experiment with different models, fine-tune your retrieval logic, and let your creativity run wild. The world of RAG is yours to explore, optimize, and innovate. Build something bold, share it with the world, and remember: every line of code you write brings us closer to smarter, more intuitive AI. Let’s make it happen—your next breakthrough is just a prompt away! 🚀

Further Resources

🌟 In addition to this RAG tutorial, unleash your full potential with these incredible resources to level up your RAG skills.

How to Build a Multimodal RAG | Documentation
How to Enhance the Performance of Your RAG Pipeline
Graph RAG with Milvus | Documentation
How to Evaluate RAG Applications - Zilliz Learn
Generative AI Resource Hub | Zilliz

We'd Love to Hear What You Think!

We’d love to hear your thoughts! 🌟 Leave your questions or comments below or join our vibrant Milvus Discord community to share your experiences, ask questions, or connect with thousands of AI enthusiasts. Your journey matters to us!

If you like this tutorial, show your support by giving our Milvus GitHub repo a star ⭐—it means the world to us and inspires us to keep creating! 💖