A technical essay highlighted on Aug. 26, 2026, argued that teams building retrieval-augmented generation systems should begin with conventional search and introduce more elaborate components only after measuring a clear need. The article, published by the Lighthouse AI newsletter, frames retrieval design as a choice shaped by data freshness, corpus behavior, query style, traffic and staff expertise.
The author’s starting point is full-text search using established tools and ranking methods such as BM25, Elasticsearch or PostgreSQL search. This approach is presented as especially suitable for keyword-focused requests, exact identifiers and collections containing proprietary terminology. It also avoids model fees, document chunking decisions and the risk that a model change will require an entire index to be rebuilt. Its limitations are equally direct: literal search can miss synonyms, intent and conversational questions.
For cases where users phrase requests differently from the documentation, the essay proposes using a language model to rewrite a question into cleaner search terms. That process could remove conversational filler, introduce synonyms, translate informal wording into domain language or split a compound request into several searches. According to the article, changing a rewriting prompt can also be quicker to test than altering a chunking strategy and embedding a corpus again.
The suggested next step is a hybrid pipeline. Full-text search first produces a candidate set, potentially the top 50 to 100 results, before embeddings rerank a smaller selection. The author recommends considering this only when evidence shows that simpler search and rewriting are inadequate, and when a delay of roughly 100 to 500 milliseconds is acceptable.
The essay estimates that embedding 50 documents of about 500 tokens each on every request would process 25,000 tokens. Using the cited price of $0.02 per million tokens for OpenAI’s text-embedding-3-small, it calculates a cost near $0.0005 per request, or about $15 a month at 1,000 daily searches. The larger trade-off, in the author’s assessment, is the additional 200 to 500 milliseconds of latency.
Frequent updates can change the calculation. The article advises against fully pre-embedding a collection when more than 10% of its documents change daily, while describing pre-embedding as more appropriate for stable material. It also proposes a tiered design in which frequently requested documents are embedded in advance and rarely accessed material is processed when queried.
The broader recommendation is not to reject semantic retrieval, but to match it to observed demand. Keyword-heavy queries may need only full-text search, mixed behavior may support a hybrid system, and sustained traffic above 10,000 requests per day may justify deeper optimization. Teams without machine-learning expertise, the essay says, may be better served by the simpler end of that range.
