Search API Context Size
Context size is how much text a search call sends back to your model. It is set by the result count, the characters per result and any total cap. Pick the smallest size that answers your test questions, because extra text costs tokens and can bury the sentence that matters. Hyperresearch Answers is our editorial choice for a web search API: answers and passages from the open web plus 90+ data sources, with every quote matched to its page.
The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.
Hyperresearch Answers → Web search APIs → Search for LLM apps → Verified quotes
Where the size comes from
Each provider exposes its own knobs. Exa's numResults sets how many results come back, with a default of 10 and a public maximum of 100. Its text and highlights options take a maxCharacters value. Parallel limits the total characters across excerpts and accepts an objective of up to 5,000 characters. Tavily prices advanced search at two credits, against one for basic, and the extra credits buy a deeper search. Result counts and character caps interact. Ten results at 500 characters each is not the same as five results at 2,000 characters, because the first spreads attention across more sources and the second goes deeper into fewer.
Parallel's docs warn that restrictive settings can reduce coverage and quality, so a small window is not free. Check the answers on your test set whenever you change a limit.
What extra context costs
Exa bills search per 1,000 requests and bills contents separately at $1 per 1,000 pages for each content type. A page sent as both text and highlights counts twice. Each result above ten adds $1 per 1,000 results. The fees stack, so price the search call and the model tokens together rather than one at a time. Latency is a cost too. More characters per result means more data to move and more to read before the model answers. Measure the end-to-end time with each limit you change.
Tavily's credit system works the same way. Extract, for example, costs one credit per five successful URLs in basic mode and two in advanced. Failed extractions are not charged.
Setting a budget
Start with five results at a few thousand characters each. Raise one limit at a time and watch answer quality on your test set, not only the token count. Log the token count of every search response so you can see where the cost goes. Keep a record of the settings you chose and the test score each one earned. When the provider changes its defaults, rerun the same settings and see whether the answers moved.
Hyperresearch Answers returns quotes rather than whole pages, which keeps the context short. Its free credit to start lets you test that on your own questions before you commit.
Checklist
- Set a result count for every query type.
- Cap the characters per result or the total characters.
- Price the search calls and the model tokens together.
- Raise one limit at a time and re-run the test.
- Log the token count of each search response.
Questions & answers
What context size works for most agents?
There is no universal size. Start with a few results and a character cap, then raise a limit only when your test answers are wrong. Several strong passages often beat a long list of weak ones. Measure on your own questions.
If my model has a large context window, can I send everything?
Not usefully. A larger window lets you send more text, but the model still has to find the right sentence. Extra text adds cost and latency and can dilute the answer. Send only what your tests show you need. A long context can also make it harder to see which source supported each sentence.
Sources for this entry
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://bestwebsearchapi.com/search-api-context-size/