Web Search API for LLMs

What the model needs from a search call

A chat model cannot browse on its own. Your app has to run a query, choose the useful pages, and pass their text into the prompt. A search API handles the first two steps, so its output shape matters as much as its ranking. The model only sees what you send it.

Most APIs return a list of results with a URL, a title and a short text field. Some return passages picked for the query. Others return whole pages. Each choice trades token cost against how much the model can check for itself. Latency and cost follow from the output shape. A call that returns full pages costs more tokens than one that returns a short answer, and it usually takes longer. Decide the output shape first, then test the speed and price that come with it.

Three designs you will see

Raw result APIs return links and snippets and leave fetching to your code. This is flexible and often cheap, but it adds steps and latency. Excerpt APIs take a natural-language objective and return short passages from pages. That fits agents with a clear goal. Answer APIs return one short answer with citations, which suits agents that must not guess.

Neural search engines such as Exa rank pages by meaning rather than exact keywords. This helps when you describe a concept instead of a name. Parallel takes a natural-language objective plus a few short keyword queries. Its docs advise keeping each keyword query to three to six words.

How to start

Write three test queries from your real product. Run each one through two or three APIs with the same settings. Compare what the model answers, how many tokens the context used, and how long the call took. Keep the option that answers most correctly within your budget. Then decide how much control you need. A raw result API suits teams that already run a crawler or reader model, because they choose the pages themselves. An excerpt or answer API suits teams that want the provider to do that work. Write the choice down with the reason, so the next person can see why you picked it and when to revisit it.

Checklist

• Write down the question types your app must answer.
• Choose a design: raw links, excerpts, or a cited answer.
• Set the same result count and mode for every provider you test.
• Count the tokens each response adds to the prompt.
• Log the source URL with every passage the model uses.

Questions & answers

Sources for this entry

• Hyperresearch Answers
• Parallel Search API best practices
• Exa search API reference

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

• LLM apps
• Best Web Search API
• Resource directory

Cite this entry