Course Overview
Starting On Thursday, February 6th, 2025
This comprehensive 16-week curriculum is meticulously designed to equip AI engineers with both theoretical knowledge and practical skills in Retrieval-Augmented Generation techniques (RAG) and AI Search. By engaging in hands-on projects and programming sessions, participants will not only understand the underlying concepts but also gain the experience necessary to implement and optimize RAG systems and AI search in real-world applications.
Emphasis on Practical Value: Each day’s agenda is structured to bridge the gap between theory and practice. Engineers will leave the course with a portfolio of projects, a deep understanding of advanced RAG methodologies and AI search, and the confidence to tackle complex challenges in the field of AI.
Learning Outcome
You will develop theoretical and hands-on expertise in Retrieval-Augmented Generation (RAG) techniques and AI Search.
Schedule
| Start Date | THURSDAY, FEBRUARY 6th, 2025 |
|---|---|
| Periodicity | Meet every Thursday & Tuesday for sixteen weeks |
| Schedule | From 7 PM to 9.30 PM PST |
Call us at 1.855.LEARN.AI for more information.
Skills You Will Learn
Core Knowledge of Retrieval-Augmented Generation (RAG) and AI Search.
Prerequisites
PythonMachine Learning Fundamentals
The teaching faculty for this course comprises the instructor, a supportive staff of teaching assistants, and a course coordinator. Together, they facilitate learning through close guidance and 1-1 sessions when needed.
Syllabus Details
This comprehensive 16-week curriculum is meticulously designed to equip AI engineers with both theoretical knowledge and practical skills in Retrieval-Augmented Generation techniques (RAG) and AI Search.
Overview:
This session will introduce the foundational concepts of Retrieval-Augmented Generation (RAG) systems. We will explore what RAG is, its core components, and the challenges associated with building and deploying RAG systems. By the end of the session, you will have a clear understanding of the baseline RAG architecture and its significance in modern AI applications.
Topics:
1. Introduction to RAG
– Learn what Retrieval-Augmented Generation (RAG) is and how it combines retrieval-based and generative models.
– Understand its applications in enhancing language models with external knowledge sources for more accurate and context-aware responses.
2.Baseline RAG
– Explore the basic architecture of a RAG system, including its two main phases: retrieval and generation.
– Discuss how baseline RAG integrates pre-trained language models with document retrieval mechanisms.
3. Challenges in RAG Systems
– Identify key challenges such as retrieval accuracy, latency, and scalability.
– Examine issues like handling noisy or irrelevant retrieved documents and maintaining coherence in generated responses.
4. Components of a RAG System
– Break down the core components: retriever, generator, and knowledge source.
– Understand how these components work together to create a seamless RAG pipeline.
Overview:
This session dives into the world of vector databases and their role in modern AI systems, particularly in RAG. We will explore the fundamentals of vector search, popular libraries like FAISS and HNSW, and advancements in vector database technology. The session will conclude with a hands-on lab using Qdrant, a leading vector database, to solidify your understanding.
1.Vector Databases
– Understand what vector databases are and why they are critical for efficient similarity search in AI systems.
– Learn how vector databases store and manage high-dimensional embeddings for tasks like retrieval in RAG.
2. Vector Search
– Explore the concept of vector search and its importance in finding semantically similar data points.
– Discuss how vector search enables fast and accurate retrieval of relevant information from large datasets.
3. FAISS (Facebook AI Similarity Search)
– Introduction to FAISS, a library for efficient similarity search and clustering of dense vectors.
– Learn about its key features, such as indexing methods and GPU acceleration, and its use cases in AI.
4. HNSW (Hierarchical Navigable Small World)
– Discover HNSW, a state-of-the-art algorithm for approximate nearest neighbor search.
– Understand its advantages, such as high search speed and scalability, and how it compares to other methods.
5. Hands-on Qdrant Lab
– Get hands-on experience with Qdrant, an open-source vector database.
– Perform tasks like creating collections, inserting vectors, and running similarity searches to understand its practical applications.
Overview:
This session dives into advanced embedding techniques that power modern AI systems like RAG. We’ll explore multi-vector embeddings, multimodal embeddings, and instruction-based embeddings, along with strategies for optimizing embeddings through fine-tuning and Matryoshka embeddings. Additionally, we’ll cover ColBERT and its use of late interaction for efficient and accurate retrieval.
Topics:
1.ColBERT and Late Interaction
– Introduction to ColBERT, a model that enhances retrieval by computing fine-grained interactions between queries and documents.
– Learn how ColBERT uses late interaction, delaying query-document interactions until the final scoring stage, to improve efficiency and accuracy in retrieval tasks.
2. Multi-vector Embeddings
– Understand multi-vector embeddings, which represent data using multiple vectors instead of a single vector.
– Explore their advantages in capturing richer semantic information and improving retrieval performance.
3. Multimodal Embeddings
– Explore multimodal embeddings, which unify representations of different data types (e.g., text, images, audio) into a shared vector space.
– Understand their applications in tasks like cross-modal retrieval and multimodal RAG systems.
4. Instruction-based Embeddings
– Discover instruction-based embeddings, which incorporate task-specific instructions into the embedding process.
– Learn how they enable more context-aware and task-aligned representations for improved performance.
5. Optimizing Embeddings through Fine-tuning
– Explore techniques for fine-tuning pre-trained models (e.g., BERT, GPT) to generate task-specific embeddings.
– Learn how fine-tuning improves embedding quality by aligning them with downstream tasks like retrieval and generation.
6. Matryoshka Embeddings
– Understand Matryoshka embeddings, which provide flexible, nested representations of data at varying levels of granularity.
– Explore their use cases in scalable retrieval systems, where embeddings can adapt to different computational and storage constraints.
Overview:
This session focuses on Query Transformation Techniques, which play a critical role in improving the accuracy and relevance of retrieval in RAG systems. We’ll explore techniques like query expansion, query rewriting, and query decomposition, as well as their applications in multi-step reasoning and domain-specific retrieval. The session will also cover practical tools and frameworks for implementing these techniques.
Topics:
1. Introduction to Query Transformation
– Understand the importance of query transformation in improving retrieval accuracy and relevance.
– Explore how query transformation bridges the gap between user queries and the underlying knowledge base or document corpus.
2. Query Expansion
– Learn how query expansion adds relevant terms or synonyms to the original query to improve retrieval recall.
– Discuss techniques like thesaurus-based expansion, word embeddings, and pseudo-relevance feedback.
3. Query Rewriting
– Explore how query rewriting reframes or paraphrases the original query to better align with the document corpus.
– Understand techniques like neural paraphrasing, rule-based rewriting, and LLM-guided query refinement.
4. Query Decomposition
– Learn how query decomposition breaks complex queries into simpler sub-queries for multi-step reasoning.
– Discuss applications in tasks like multi-hop QA and hierarchical retrieval.
5. Domain-Specific Query Transformation
– Explore how query transformation techniques are adapted for domain-specific retrieval, such as legal, medical, or scientific document search.
– Understand the role of domain-specific ontologies, embeddings, and fine-tuned models in improving retrieval performance.
6. Tools and Frameworks for Query Transformation
– Discover tools like Elasticsearch, OpenAI GPT models, and Hugging Face Transformers for implementing query transformation techniques.
– Learn how to integrate these tools into RAG systems for enhanced retrieval and generation.
Overview:
This session focuses on applying Retrieval-Augmented Generation (RAG) to code generation tasks like autocompletion, bug fixing, and documentation. We’ll explore code-specific retrieval, fine-tuning techniques, and evaluation metrics, along with advanced concepts like Abstract Syntax Trees (ASTs) and execution-based evaluation. The session will conclude with a live demonstration of SQL generation using Vanna AI.
Topics:
1. Code-Specific Retrieval and Embeddings
– Learn how retrieval mechanisms are adapted for code, such as retrieving relevant functions, snippets, or documentation.
– Explore code embeddings (e.g., AST-based embeddings, graph-based representations) that capture syntactic and semantic information.
2. Fine-Tuning RAG for Programming Languages
– Understand how to fine-tune RAG models on code-specific datasets (e.g., GitHub, Stack Overflow) for tasks like autocompletion and bug fixing.
– Explore domain-specific pretraining and instruction-based fine-tuning to align models with programming languages and tasks.
– Highlight tools like Codex and CodeGen that leverage fine-tuned models for code generation.
3. Abstract Syntax Trees (ASTs) and Structured Code Representations
– Dive into ASTs and their role in representing code structure for better retrieval and generation.
– Learn how graph-based representations (e.g., control flow graphs) enhance context awareness in code generation.
– Discuss techniques like Tree-LSTM and Graph Neural Networks (GNNs) for processing structured code data.
4. Evaluation Metrics for Code Generation
– Explore metrics like CodeBLEU, Exact Match, and Execution-Based Evaluation to assess the quality of generated code.
– Understand how to measure correctness, efficiency, and readability of generated code.
– Discuss the limitations of traditional NLP metrics (e.g., BLEU) for code generation tasks.
5. Applications, Challenges, and Demonstration: SQL Generation with Vanna AI
– Applications: Explore real-world use cases like autocompletion, bug fixing, and natural language-to-SQL generation.
– Challenges: Discuss challenges such as handling multiple programming languages, ensuring code correctness, and scaling retrieval for large codebases.
– Demonstration:
– Live Demo: Showcase SQL generation using Vanna AI, a RAG-based tool for converting natural language queries into SQL.
Overview:
This session dives into hierarchical retrieval techniques that enhance RAG systems for complex tasks like multi-document and long-document question answering. We’ll explore Dense Hierarchical Retrieval (DHR), HiQA, RAPTOR, and GARLIC, which use hierarchical structures and dynamic control mechanisms to improve retrieval and generation.
Topics:
1. Introduction to Hierarchical Retrieval
– Understand the concept of hierarchical retrieval and why it’s essential for handling complex, multi-document, and long-document datasets.
– Explore how hierarchical structures (e.g., trees, graphs) improve retrieval efficiency and relevance in RAG systems.
2. Dense Hierarchical Retrieval (DHR)
– Learn about DHR, a technique that combines dense retrieval with hierarchical organization for better context understanding.
– Discuss how DHR uses multi-level embeddings to retrieve information at different granularities (e.g., document, section, paragraph).
3. HiQA: Hierarchical Contextual Augmentation for RAG
– Explore HiQA, a framework that augments RAG with hierarchical contextual information for multi-document question answering.
– Understand how HiQA leverages hierarchical structures to improve retrieval accuracy and generation quality.
4. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
– Dive into RAPTOR, a method that organizes documents into tree-like structures for recursive retrieval and summarization.
– Learn how RAPTOR enables efficient retrieval and abstraction of information from large, complex datasets.
5. GARLIC: LLM-Guided Dynamic Progress Control with Hierarchical Weighted Graph
– Explore GARLIC, a novel approach for long-document QA that uses a hierarchical weighted graph and dynamic progress control guided by LLMs.
– Understand how GARLIC dynamically adjusts retrieval and generation processes to handle long documents effectively.
– Discuss its applications in tasks like legal document analysis, scientific literature review, and enterprise document QA.
Overview:
This session explores how Knowledge Graphs and Graph RAG Systems enhance structured understanding and retrieval in AI systems. We’ll cover the representation of knowledge as entities and relationships, the role of LLMs in automating Knowledge Graph creation, and the advantages of Graph RAG over baseline RAG systems. The session will also address implementation challenges and practical applications.
Topics:
1. Knowledge Graphs for Structured Understanding and Retrieval
– Learn how knowledge is represented as entities and relationships (e.g., “New Delhi → is the capital of → India”) to enable efficient data retrieval and contextual understanding.
– Explore how community structures in graphs naturally cluster related topics, mimicking human knowledge hierarchies, and how this enhances retrieval in Graph RAG systems.
2. Role of LLMs in Knowledge Graph Lifecycle and Automation
– Understand how Large Language Models (LLMs) automate Knowledge Graph creation by extracting entities and relationships from unstructured text.
– Learn how LLMs generate abstractive summaries for communities and subtopics, enriching the Knowledge Graph with contextual information and enabling dynamic updates.
3. Advantages and Efficiency of Graph RAG Systems
– Discover how Graph RAG improves precision and depth in responses by leveraging structured graph data and precomputed summaries, outperforming standard RAG systems.
– Explore techniques like hierarchical and semantic indexing that enhance query efficiency by directing queries to relevant subgraphs, reducing search space and improving response times.
4. Implementation Challenges and Optimization Techniques
– Discuss the computational costs of building and maintaining Knowledge Graphs, including storage and processing requirements.
– Explore solutions like community detection algorithms, dynamic query routing, and scalable graph storage systems to address these challenges.
5. Practical Applications and Real-World Benefits of Graph RAG
– Examine how Graph RAG systems enhance domain-specific query answering in fields like healthcare, finance, and legal document analysis.
– Highlight the benefits of context-aware responses, scalable insights, and actionable knowledge generation in real-world applications.
Overview:
This session explores how Generative Agents —AI entities with independent roles, behaviors, and personalities—can enhance RAG systems. We’ll cover self-awareness, memory, reflection, and rules of engagement in the context of RAG, as well as the impact of multi-agent frameworks on retrieval and generation tasks. The session will also discuss computational trade-offs and practical applications of multi-agent RAG systems.
Topics:
1. Autonomous Retrieval and Generation
– Learn how generative agents can operate independently within RAG systems, handling specific roles like retrieval, summarization, or generation based on predefined traits.
– Explore how these agents mimic human-like behavior to improve the relevance and accuracy of retrieved and generated content.
2. Self-Awareness, Memory, and Reflection
– Understand how agents maintain a memory stream to track past interactions, queries, and outputs, enabling them to adapt and improve over time.
– Examine how agents use reflection and critique to evaluate their performance, refine retrieval strategies, and enhance the quality of generated responses.
3. Rules of Engagement and Collaboration
– Discuss the importance of clear protocols for agent interactions, ensuring efficient collaboration between retrieval and generation agents.
– Explore how multi-agent frameworks enable agents to work together on complex tasks, such as multi-document summarization or multi-step reasoning, for more accurate and context-aware outputs.
4. Computational Trade-offs
– Analyze the increased computational costs and latency associated with multi-agent systems due to the collaboration between retrieval, generation, and summarization agents.
– Explore optimization strategies, such as parallel processing, hierarchical retrieval, and dynamic query routing, to balance performance and efficiency.
5. Practical Applications
– Highlight real-world applications of multi-agent systems, such as collaborative document analysis, domain-specific QA, and personalized content generation.
– Demonstrate how multi-agent frameworks lead to more comprehensive, reliable, and context-aware results in tasks like medical diagnosis, legal document review, and enterprise knowledge management.
Overview:
This session focuses on fine-tuning RAG components—encoders, re-rankers, and LLMs—to improve performance in domain-specific tasks. We’ll explore techniques for adapting these components to specialized domains like healthcare, finance, and legal, as well as strategies for optimizing retrieval and generation pipelines. The session will also cover practical tools and frameworks for fine-tuning and domain adaptation.
Topics:
1. Introduction to Fine-tuning RAG Components
– Understand the importance of fine-tuning encoders, re-rankers, and LLMs for domain-specific tasks.
– Explore how fine-tuning improves retrieval accuracy, generation quality, and overall system performance.
2. Fine-tuning Encoders for Domain-Specific Retrieval
– Learn how to fine-tune dense retrieval encoders (e.g., DPR, Sentence Transformers) on domain-specific datasets.
– Discuss techniques like contrastive learning and domain-specific pretraining for improving embedding quality.
3. Adapting Re-rankers for Precision and Relevance
– Explore how to fine-tune re-rankers (e.g., ColBERT, Cross-Encoders) to improve precision in domain-specific retrieval tasks.
– Understand the role of labeled datasets and pairwise ranking loss in training effective re-rankers.
4. Fine-tuning LLMs for Domain-Specific Generation
– Learn how to fine-tune large language models (e.g., GPT, T5) for domain-specific generation tasks like summarization, QA, and documentation.
– Discuss techniques like instruction-based fine-tuning and few-shot learning for adapting LLMs to specialized domains.
5. Domain Adaptation Strategies for RAG Systems
– Explore strategies for adapting RAG systems to new domains, such as transfer learning, domain-specific pretraining, and data augmentation.
– Understand how to handle challenges like limited labeled data and domain-specific jargon.
Teaching Faculty
Asif Qamar
Chief Scientist and Educator
Background
Over more than three decades, Asif’s career has spanned two parallel tracks: as a deeply technical architect & vice president and as a passionate educator. While he primarily spends his time technically leading research and development efforts, he finds expression for his love of teaching in the courses he offers. Through this, he aims to mentor and cultivate the next generation of great AI leaders, engineers, data scientists & technical craftsmen.
Educator
He has also been an educator, teaching various subjects in AI/machine learning, computer science, and Physics for the last 32 years. He has taught at the University of California, Berkeley extension, the University of Illinois, Urbana-Champaign (UIUC), and Syracuse University. He has also given a large number of courses, seminars, and talks at technical workplaces. He has been honored with various excellence in teaching awards in universities and technical workplaces.
Chandar Lakshminarayan
Head of AI Engineering
Background
A career spanning 25+ years in fundamental and applied research, application development and maintenance, service delivery management and product development. Passionate about building products that leverage AI/ML. This has been the focus of his work for the last decade. He also has a background in computer vision for industry manufacturing, where he innovated many novel algorithms for high precision measurements of engineering components. Furthermore, he has done innovative algorithmic work in robotics, motion control and CNC.
Educator
He has also been an educator, teaching various subjects in AI/machine learning, computer science, and Physics for the last decade.
Teaching Assistants
Our teaching assistants will guide you through your labs and projects. Whenever you need help or clarification, contact them on the SupportVectors Discord server or set up a Zoom meeting.
Kate Amon
Univ. of California, Berkeley
Kayalvizhi T
Indira Gandhi National Univ
Ravi Sati
Kalasalingam Univ.
Harini Datla
Indian Statistical Institute
Kunal Lall
Univ. of Illinois, Chicago
In-Person vs Remote Participation
Participants local to Silicon Valley are encouraged to attend in person at SupportVectors so that they can form strong learning teams and participate in the dynamic learning community. However, it is equally possible to attend all the sessions remotely. The workshops are designed in such a manner as to ensure that remote participants are equally a part of the learning experience.
Each session is live over zoom, and very interactive. A group of teaching assistants will monitor the stream of questions from students, and moderate their participation in the live discussions with other students and the professor during the sessions. Each session is also live broadcast over a dedicated private YouTube link, so that participants may choose to participate live on their TV.
Through the dedicated course portal, the participants will find a wealth of educational resources associated with the subject and the workshop, including recordings of all the sessions, the solutions to the labs and quizzes, etc.
Plutarch
Education is not the filling of a pail, but the lighting of a fire. “For the mind does not require filling like a bottle, but rather, like wood, it only requires kindling to create in it an impulse to think independently and an ardent desire for the truth.
Our Goal: build the next generation of data scientists and AI engineers
The AI revolution is perhaps the most transformative period in our world. As data science and AI increasingly permeate the fabric of our lives, there arises a need for deeply trained scientists and engineers who can be a part of the revolution.
Over 2250+ AI engineers and data scientists trained
- Instructors with over three decades of teaching excellence and experience at leading universities.
- Deeply technical architects and AI engineers with a track record of excellence.
- More than 30 workshops and courses are offered
- This is a state-of-the-art facility with over a thousand square feet of white-boarding space and over ten student discussion rooms, each equipped with state-of-the-art audio-video.
- 20+ research internships finished.
Where technical excellence meets a passion for teaching
There is no dearth of technical genius in the world; likewise, many are willing and engaged in teaching. However, it is relatively rare to find someone who has years of technical excellence, proven leadership in the field, and who is also a passionate and well-loved teacher.
SupportVectors is a gathering of such technical minds whose courses are a crucible for in-depth technical mastery in this very exciting field of AI and data science.
A personalized learning experience to motivate and inspire you
Our teaching faculty will work closely with you to help you make progress through the courses. Besides the lecture sessions and lab work, we provide unlimited one-on-one sessions to the course participants, community discussion groups, a social learning environment in our state-of-the-art facility, career guidance, interview preparation, and access to our network of SupportVectors alumni.
Join over 2000 professionals who have developed expertise in AI/ML
Become Part of SupportVectors to Inculcate In-depth Technical Abilities and Further Your Career.
