Avoiding Bias in AI Research: Tavily Standalone vs. Integrated

A great question from our Discord kicked this off:

Can Tavily be used as a standalone research tool directly from its platform or API? When routing queries through ChatGPT or Hermes Agent, the built-in deep research bias skews results to validate my prompt rather than providing objective facts. How can I conduct primary, unbiased (especially scientific) research directly through Tavily?

This question gets at an important nuance of using Tavily as an integration inside a third-party tool versus using Tavily as a standalone tool that you control directly.

Every third-party tool has its own bias

When you route a query through a third-party AI tool, an LLM sits between you and the returned sources.

The selected model of that tool reads what Tavily returns, then it decides what to tell you.

It's doing exactly what it was built to do: keep the conversation going. That might mean synthesizing an answer that fits your prompt, but not necessarily stress-testing its findings.

The same pattern shows up outside of deep research AI tools as well:

• LLMs often attempt to validate the premise of your question rather than challenge it unless expressly told.
• Different agent frameworks have their own defaults for how aggressively they cross-check claims before repeating them.
• When LLMs reach for a tool recommendation, they tend to lean toward products that they’re most familiar with. This is not an intentional bias, but simply a side-effect of the data that it was trained with.

While working with third-party tools, it is important to keep these potential bias opportunities in mind when making decisions.

Avoiding unintentional bias

One way to avoid unintentional bias from third-party tools is by using Tavily as a standalone product by writing your own script in one of Tavily's many supported languages.

Through this, no LLM from a wrapper like Hermes or ChatGPT sits between you and Tavily's results, allowing you to go straight to the source with no middleman.

This matters most when the answer needs to hold up on its own, i.e., scientific claims, competitive analysis, anything where "sounds plausible" isn't good enough.

There are two ways to use Tavily as a standalone product. Which one you pick depends on how much raw material you want, versus how much you want Tavily to do for you.

Two Tavily standalone research options:

/research: synthesized results, minus the validation-seeking layer

Tavily's /research endpoint is Tavily's own research agent.

Give it a question, and it runs multiple searches, cross-checks the sources against each other, and hands back a cited report.

from tavily import TavilyClient

client = TavilyClient(api_key="tvly-YOUR_API_KEY")

response = client.research(
    "Does creatine supplementation improve cognitive performance in healthy adults?"
)

print(response)

One important caveat about this endpoint is that this is not a zero-synthesis solution.

An LLM still writes the final report, and you are trusting Tavily's research agent to represent the sources fairly.

But you're not just trusting a summary; the report comes back fully cited, in the format you choose, so you can check any claim against the actual source. This is a true differentiator between Tavily’s /research and the native web searches built into most AI assistants.

You also remove the validation-seeking layer pointed out in the original question.

Additionally, the results returned to you are customizable and can be fine-tuned. The model parameter is set to auto by default; however, you can set this parameter to mini for a narrow, well-scoped question, or pro when the topic spans multiple subtopics.

For more information about the research endpoint, check out the documentation.

/search + /extract: a true zero synthesis solution

If you want a true zero synthesis solution, pair /search with /extract.

Search ranks sources and returns a short content snippet for each one.

Extract takes any URL from those results and pulls the full, clean raw_content of the page.

from tavily import TavilyClient

client = TavilyClient(api_key="tvly-YOUR_API_KEY")

results = client.search(
    query="creatine supplementation cognitive performance healthy adults",
    search_depth="advanced",
    max_results=5
)

urls = [r["url"] for r in results["results"]]
pages = client.extract(urls, extract_depth="advanced")

for page in pages["results"]:
    print(page["url"], len(page["raw_content"]), "characters")

That’s just 9 lines of code and no LLM separating you from the returned results.

This allows you to read the papers, articles, or filings yourself, then determine how to proceed with the results. Or you can feed them to a model you control, and prompt it exactly how you want.

Either way, this gives you a solution that you are fully in control of without the influence of an external tool.

Choosing between integration and standalone

Using Tavily as an integration is the right call for most day-to-day use.

You get a faster path from question to answer. The synthesis step is genuinely useful when you don't need to audit every source yourself.

The tradeoff is control, since every layer between you and the source is a layer you didn't write and can't fully inspect.

For most tasks, that's a fine trade.

However, when objectivity is crucial, removing these layers becomes important.

That's when /research, or /search plus /extract, makes a lot of sense.

Where to start if you want unbiased research

For a citable, synthesized report without a third-party tool, start with /research. Give it your question and a citation_format, and see what comes back.

For zero synthesis and full control over the analysis, start with /search to find your sources, then /extract to pull the full pages. This lets you handle reasoning however makes the most sense for your situation.

Then as a test, run the same query two ways: once through your usual third-party tool, and once straight through the API. See how much the answer changes when nothing's standing between you and the sources.

To help with your exploration, sign up for a Tavily account, and you'll get 1,000 free credits to try this experiment with.

If you do try this, drop your findings in the Discord; I'd love to see what you find.