Definition

  • Any software that uses an LLM inside its functionality 1:06.
  • Two ways to build one 1:06:
    • Add LLM-powered features to an existing product (e.g. let users "talk to their data").
    • Create entirely new LLM-powered applications: chatbots, assistants, coding tools.

Complexity spectrum

From simplest to most complex:

  1. Single prompt — e.g. generating e-commerce item descriptions may need just one well-crafted prompt 1:30.
  2. Prompt chains — multiple prompts chained together 1:30.
  3. RAG (retrieval augmented generation) — giving the LLM access to external data; e.g. a support chatbot pulling help-center articles. This adds components like a database and search logic to the app 1:30 1:48.
  4. AI agents — solve multi-step tasks: plan steps, use tools (e.g. internet search), keep memory; e.g. an AI research assistant that writes a report. Internally these are complex graphs and chains of prompts 1:48.
flowchart LR
    P1[Single prompt] --> P2[Prompt chain] --> P3[RAG<br/>+ database + search logic] --> P4[Agent<br/>planning + tools + memory]