Teaching AI Domain Knowledge: The Story of Two Rebuilds — LLM Wiki + RAG Hybrid Engine - Kurly Tech Blog

A company engineer spent a month testing three ways to supply internal domain knowledge to an LLM: an LLM Wiki, keyword search, and vector-based semantic retrieval.

The wiki approach was simple but became expensive and less reliable as documents grew and context windows filled up.

An inverted-index layer improved speed but missed synonyms and returned noisy matches for domain terms.

SQLite vector search with multilingual-e5-small improved recall, but mixed-topic docs and short internal identifiers were still hard to retrieve well.

The experiment highlights a core enterprise LLM tradeoff: keyword search is brittle, full-context injection is costly, and embeddings need extra help for domain-specific language.