External Knowledge Systems

Introduction

Agents often need to query external knowledge beyond a vector DB: Wikipedia, internal wiki, SQL DBs, search engines. Wrap each as a tool with consistent return shape.

Beginner analogy: a researcher's bookshelves — encyclopaedia, journals, web — all accessible with the same library card.

Understanding the topic

Core concepts:

• Wikipedia / DBpedia / Knowledge Graphs.
• Internal wiki (Confluence, Notion, Linear).
• SQL / NoSQL databases via secure read tools.
• Web search APIs (Tavily, Brave, Bing).
• Normalize all return shapes (text + url + score).

Syntax reference

Visual workflow / architecture:

Question
     │
     ▼
  Router ─► Wikipedia
          ─► Wiki search
          ─► SQL query
          ─► Web search
                │
                ▼
            Aggregator → LLM

Real-world use

Perplexity routes across web + internal sources; enterprise Glean indexes 50+ SaaS apps.

Best practices

• Normalize return shape across sources.
• Cache hot queries to reduce cost.

Common mistakes

• Each tool returns a different shape — LLM gets confused.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why normalize return shapes?
• Q2. When use a knowledge graph over vector search?
• Q3. Scenario: SQL queries leak PII into prompts. Mitigation?