Video cover

Evidently AI · 6:04 · Watch on YouTube

Educational explainer on how to evaluate complex AI systems — RAG pipelines, chatbots, and AI agents — that go beyond the simple input/output shape of tasks like summarization. It defines each system type, then develops three general evaluation strategies: splitting the system into smaller testable components, black-box testing of final outcomes and trajectories, and testing the complete interaction through simulated scenarios and conversation-level online evaluations.

Topics

  1. Introduction: Evaluating Complex AI Systems
  2. Key Concepts: RAG, Chatbots, and AI Agents
  3. Testing RAG Quality
  4. Testing AI Agents by Component
  5. Testing Agent Outcomes and Trajectories
  6. Scenario Testing and Online Evaluation

Overview

flowchart TD
    A[Complex AI systems:<br/>RAG, chatbots, agents] --> B[1. Split into testable components]
    A --> C[2. Test the final outcome]
    A --> D[3. Test the complete interaction]
    B --> B1[RAG: retrieval + generation]
    B --> B2[Agents: tool choice, data unit tests,<br/>intent classification, first response,<br/>single steps in scenarios]
    C --> C1[Black-box result check]
    C --> C2[Trajectory: correct tool/API sequence]
    D --> D1[Manual replay or<br/>user-simulating agent]
    D --> D2[Online: LLM-as-a-judge<br/>on full conversations]

Key takeaways

  • Complex systems are evaluated with the same machinery as simple ones (evaluators + test cases), applied to parts rather than the whole: break RAG and agents into smaller testable components. 1:46
  • RAG splits into two testable components: retrieval, tested as a ranking problem (does it find the known correct document?), and generation, tested for correctness, relevance, and hallucinations. 2:04 2:22
  • RAG needs dedicated edge-case tests: with no document retrieved the system must deny rather than invent an answer, and contradictory context must be resolved deliberately. 2:22
  • Agents decompose into five component tests — correct tool choice, data unit tests, intent classification (a pure classification problem), first response quality against ground truth, and single-step checks given a fixed conversation history. 2:54 3:34
  • Beyond components, test agents black-box (was the booking actually completed?) and by trajectory (was the expected tool/API sequence followed?), and test whole interactions with a user-simulating agent — e.g. a user changing hotel requirements mid-conversation. 4:32 5:00
  • Online evaluation applies LLM-as-a-judge at the conversation level (frustration, politeness, task resolution) plus hard metrics like function call usage — same principles as simple systems, at a bigger scale with more complex checks. 5:24 5:40