Deconstructing AI: from Semantic Search to LLMs & RAG

In this two-part presentation, “Deconstructing AI: from Semantic Search to LLMs & RAG,” we trace the evolution of modern AI by exploring semantic search as a foundation for machine learning, the architecture and inner workings of Large Language Models (LLMs), and the powerful technique of Retrieval Augmented Generation (RAG) — distilling over a year of intensive learning into two focused one-hour sessions originally delivered to the Leidos team at the Social Security Administration (SSA). Presentation slides are included.

Focus On Content  ~/ click me to toggle the navigation bar to the right

Center Theme Image

Brief Overview of the Two Presentations: “Deconstructing the AI Landscape” (Parts 1 & 2)

These two slide decks by Venkatt Guhesan form a cohesive, narrative-driven explainer on the foundational technologies that built modern AI and Generative AI. Presented as a “Tale of Three Travelers” (three stories, three paths, three journeys), they use metaphors, brain-inspired analogies, simple examples, and visuals to make complex concepts accessible — perfect for a summary blog post that walks readers through the evolution from raw intelligence simulation to practical, reliable GenAI systems.

Part 1 (Jan 22, 2026 – 31 pages/slides): “Pieces of technology that paved the way for AI and Generative AI”

  • Core Theme: The Quest for Intelligence — how we learned to reproduce the mind artificially.
  • Starts with a clear definition of AI → Machine Learning → Deep Learning.
  • Story #1 – “Qwest for Intelligence: Reproducing The Mind”: Traces biological neurons → Artificial Neural Networks (ANNs) → Perceptron (1958) → weights, bias, activation functions, and training. Uses a fun “Should I go to the playground?” example to show how a single neuron works, then scales it up to millions of neurons in Transformer-based LLMs.
  • Story #2 – “In Search of Meaning: Reproducing The Context”: Introduces the semantic gap, embeddings, vectors, cosine similarity, Word2Vec, and the LLM data pipeline (tokenization → vectorization → storage). Explains how LLMs are essentially sophisticated next-token predictors trained on massive text.
  • Story #3 – “In Search of Uruk: Building A City of Knowledge”: Shifts to constructing a shared “city of knowledge” (referencing the ancient city of Uruk as the first true city). Covers vector databases, semantic search, and the limitations of standalone LLMs.
  • Ends with the realization: “We now have a way to represent the mind!” — setting up the need for the next evolution.

Tone & Style: Educational storytelling with brain diagrams, playground analogies, and beautiful AI-themed artwork. Heavy emphasis on how neural nets, training, and embeddings laid the groundwork.

Part 2 (Feb 20, 2026 – 53 pages/slides): Continues the deconstruction with practical application

  • Opening: Extensive recap of Part 1 (AI basics, neural nets, embeddings, LLM pipeline) so it stands alone.
  • Deep Dive into Retrieval-Augmented Generation (RAG): The star of Part 2. Frames RAG as the essential bridge that fixes LLMs’ biggest flaws (outdated knowledge cut-offs, hallucinations, lack of proprietary/domain context).
    • How RAG works: Two-stage flow — Ingestion/Indexing (chunking, embedding, vector DB) + Inference (semantic retrieval → prompt augmentation → grounded generation).
    • Tools & Frameworks: LangChain (for linear chains), LlamaIndex, Haystack, Flowise, CrewAI, plus cloud-native options.
    • Implementation: Code snippets (simple LangChain RAG chains, local LLMs via Ollama), comparisons (RAG vs. fine-tuning), benefits (real-time knowledge updates, lower hallucination risk).
    • Advanced notes: RAG as an agent, streaming pipelines, UI layers (Streamlit/Chainlit).
  • Ties back to the “City of Knowledge” metaphor: RAG turns raw LLMs into reliable systems grounded in your own data.

Overall Narrative Arc Across Both Decks:

  1. Reproduce the Mind (neural nets & transformers) →
  2. Reproduce Meaning/Context (embeddings & semantic understanding) →
  3. Build the City of Knowledge (RAG + vector stores) = trustworthy Generative AI.

Key Takeaways:

  • These decks are not just technical; they’re a journey story that makes AI feel human and evolutionary.
  • Narrative: “Three Travelers, One Destination: How Neural Nets, Embeddings, and RAG Built Today’s Generative AI.”
  • Visuals (brain-hand illustrations, traveler artwork, clean diagrams) and simple analogies (playground decision, ancient Uruk) translate beautifully to blog images or embeds.
  • Audience: Beginners to intermediate — executives, developers, or curious readers who want “how we got here” without drowning in math.

I hope this article and the attached slides has been useful for you. If you like this article, please consider sharing this article in your favorite social post.

The blog I run can be found here: DigitalRiver.blog

Thanks for checking out. Cheers.