CourseAI System Design: The Complete Course · Module 6: Retrieval Architecture · part 42 of 107
Part 42 · Module 6: Retrieval Architecture

Lesson 2: Retrieval beyond one vector search

30 min read·
Module 6 · Lesson 2 · About 55 minutes · Needs: Module 3 lesson 5 (the first-token budget) and Module 4 lesson 1 (rung 3, prompt plus retrieval). This lesson builds its own small index, so lesson 1's code is not required.

The situation

Week 6 at Northdesk, Monday 7 December 2026. Tara Dunne, in finance at Tallis Retail Group, writes to Ask: "My laptop screen is cracked. How often are laptops replaced, and when is the helpdesk open?" Then: "Am I due a new one?"

Ask embeds the message once and takes the best match: the refresh rule. It misses the helpdesk hours, and Tara's own laptop is in no document. Gareth asks: what do we add beyond one search, and what does each addition cost inside Ask's 1,500 ms first-token budget?

THE SITUATION One search finds one fact Tara's message two asks one embedding top child, k=1 refresh sentence score 0.67 3 years found helpdesk hours missed her own laptop in no document
One embedding of Tara's message (amber, user text) finds the refresh rule (green) and misses the helpdesk hours and her own laptop's date (red).

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside

Or open everything with a plan