Semantic Search Embeddings vs Keyword Search for a SaaS Help Center

Short summary

A practical guide to building semantic search for a SaaS help center, recommending embeddings-based retrieval over document chunks with keyword search as fallback and reranking added later. The author shares a costly lesson: an $8,742 token bill from sending whole articles to the answering model instead of retrieving small, attributable chunks first. Includes a runnable Go chunking utility and operational guidance on relevance SLOs, source traceability, and ingestion hygiene.

• •Start with embeddings-based semantic retrieval, keep keyword search as fallback, add reranking only when measurable
• •An $8,742 token bill taught the author to retrieve small chunks first rather than sending whole articles to the chat model
• •Includes a Go chunking utility and guidance on relevance SLOs, source traceability, and ingestion hygiene

Generated with AI, which can make mistakes.