LLM Zoomcamp 1.5 — Search

Why It Matters

Efficient pre‑filtering reduces LLM inference costs while delivering more accurate, context‑aware answers, a competitive advantage for AI‑driven products.

Key Takeaways

• Index FAQ data to enable efficient search before LLM processing.
• MinSearch offers lightweight alternative to Elasticsearch for small datasets.
• Text vs keyword fields control relevance and exact-match filtering.
• Boosting adjusts importance of fields like question versus answer.
• Integrated search function becomes first step in RAG pipeline.

Summary

The video walks through adding a search layer to the LLM Zoomcamp 1.5 project, showing how to index a 1,100‑document FAQ set so that queries can retrieve relevant passages before invoking a large language model.

Because sending the entire corpus to an LLM is costly and can degrade answer quality, the presenter recommends a retrieval‑augmented generation (RAG) approach. He reviews heavyweight options such as Apache Lucene, Elasticsearch and Solr, then introduces MinSearch—a lightweight, Python‑only library he created for small‑scale use cases.

He explains the distinction between text fields (full‑text searchable) and keyword fields (exact‑match filters), using course identifiers to limit results. A boosting dictionary is also demonstrated, giving the question field twice the weight of answers and down‑weighting sections.

By filtering and ranking documents early, developers can cut API expenses, speed up response times, and improve answer relevance, making the search step a critical foundation for any RAG pipeline.

Original Description

Building a keyword search engine over the FAQ with minsearch: text vs keyword fields, filtering, and boosting.

Lesson notes: https://github.com/DataTalksClub/llm-zoomcamp/blob/main/01-agentic-rag/lessons/05-search.md

LLM Zoomcamp is a free course on building real-world LLM applications: https://github.com/DataTalksClub/llm-zoomcamp

Module 1 (Agentic RAG), Part 1 — lesson 5 of 10.

Chapters:

0:00 Why we need search

1:31 Search libraries

2:16 minsearch

3:42 Text and keyword fields

6:11 Trying a search

8:00 Boosting fields