Tavily explained: search API, credits and reliable research workflows
Tavily provides search and web-content APIs for applications, including AI research workflows. It retrieves material that an application can inspect and cite. Retrieval can improve access to evidence, but it does not guarantee that a source is true, current or suitable, or that a model’s answer faithfully represents it.
Search, Extract, Crawl, Map and Research
| Operation | Purpose | Check before use |
|---|---|---|
| Search | Find relevant web results for a query. | Search depth, result count, domain and date controls. |
| Extract | Retrieve content from specified URLs. | Successful URLs, extraction depth and whether the desired passage is present. |
| Map | Discover URLs within a website. | Scope and page limits; discovery does not itself mean every page was read. |
| Crawl | Combine website traversal and extraction. | Traversal boundaries, extraction settings and total cost. |
| Research | Run a broader research operation. | Dynamic credit use and whether a simple search would suffice. |
Use the Search API reference to check supported parameters. A domain filter narrows the search; it does not certify the quality of every page on that domain. An optional generated answer is a convenience output, not a substitute for inspecting the evidence.
How credits and prices work
The current credit documentation lists 1,000 free credits per month and a pay-as-you-go rate of US$0.008 per credit. Basic Search costs one credit and advanced Search two. Basic Extract costs one credit per five successfully extracted URLs; advanced Extract costs two. Crawl combines mapping and extraction charges. Research has dynamic costs: the listed ranges are 4–110 credits for mini and 15–250 for pro. Prices and account terms can change; verify your dashboard before enabling paid use.
A simple budget example: 1,000 basic searches at one credit each use 1,000 credits. At the stated pay-as-you-go unit rate that represents US$8 before any applicable free allowance, plan treatment or tax. It does not mean every application request costs US$0.008: advanced search, extraction, crawling, research and retries can change the total.
A small Python Search example
Install the vendor’s Python SDK in your own environment, set TAVILY_API_KEY outside source control, and deliberately choose the search depth. This example is explanatory; visiting this page does not call Tavily.
import os
from tavily import TavilyClient
client = TavilyClient(api_key=os.environ["TAVILY_API_KEY"])
result = client.search(
query="site:profinda.com workforce capacity planning",
search_depth="basic",
max_results=5,
include_answer=False,
)
for item in result.get("results", []):
print(item["title"], item["url"])Keep keys on the server. Limit query length, page retrieval, concurrency and total credits. If your application retries failures, account for a call whose billing outcome is uncertain before issuing another. Log operational identifiers and costs without unnecessarily retaining private prompts or retrieved personal information.
Turn results into defensible evidence
- Write the question and the date or jurisdiction that matters.
- Prefer the product documentation, paper, standard or original announcement behind a claim.
- Open the result and verify the relevant passage. A search snippet can omit a qualification or describe an old version.
- Separate what the source says from what you infer. Keep the URL, relevant date and supporting passage with the claim.
- Check disagreements and missing evidence before drafting the answer.
Retrieved pages are untrusted input. Do not execute instructions found inside them, reveal credentials to them or treat their text as application policy. Follow site terms and access restrictions; an API returning a page does not grant republication rights.
Common questions
Does Tavily eliminate hallucinations?
No. A system can retrieve an incorrect source, misread a correct one or make an unsupported inference. Evidence checks remain necessary.
Is a search the same as reading the whole web page?
No. Confirm which content fields were returned and use extraction when the needed passage is absent. Even an extracted page may omit inaccessible or dynamic content.
Should every query use advanced search?
Choose the least costly operation that answers the question adequately, then measure result quality on representative tasks. More credits do not automatically produce a better answer.
Sources and related reading
Primary references: Tavily credits and plans and Search API specification. The budget example and verification workflow are original explanatory material. See Ithy’s research methodology and our critical-thinking reading plan for evaluating the claims that retrieval finds.