Fresh answers with links: how to work with a web-search model

Models like Sonar from Perplexity answer not from memory but from web search results. Web search in Yenisei Chat and Yenisei Agent works the same way. So you need to work with such a model differently: between the question and the answer there is a search, and you can influence it with your phrasing.

Sonar is not in the Yenisei catalog; the techniques below work equally for it, for our search, and for the "model plus search" pairing you assemble through the API.

Phrase it as if you were searching yourself

The model turns your question into search queries. The more specific the question, the more precise the results.

• Bad: "Tell me about the new rules."
• Good: "What changes to dairy product labeling rules in Russia came into force in 2026? Give dates and links to official sources."

Specify the region, the period, the type of source ("official", "industry media"): all of this narrows the search in the right direction.

Always ask for sources

Demand links explicitly: "After every fact, add a link to the source." That way you get a verifiable answer rather than a smooth retelling of who knows what.

Through the API this is set in the system message, and the model calls search as a tool executed by your code:

from openai import OpenAI
client = OpenAI(base_url="https://api.yenisei.ru/v1", api_key="YENISEI_KEY")
resp = client.chat.completions.create(
    model="kimi-k2-thinking",
    messages=[
        {"role": "system", "content": (
            "Answer briefly and to the point, only from the search results. "
            "After every statement, add a link to the source. "
            "If the sources contradict each other, say so explicitly."
        )},
        {"role": "user", "content": "Compare the current terms of subsidized mortgages at two or three of the largest Russian banks."},
    ],
    tools=[WEB_SEARCH_TOOL],  # the search function definition; your code executes it
)

Which model to put behind the search

• A simple factual question, such as an exchange rate, a date, a definition, "what happened": Qwen3.8 Flash (qwen-flash) or DeepSeek V4 Flash (deepseek-flash), fast and inexpensive.
• An overview or comparison with a lot of material found: DeepSeek V4 Pro (deepseek-pro), whose long context holds dozens of pages.
• Contradictory sources, where the job is to judge rather than retell: Kimi K2 Thinking (kimi-k2-thinking), a model with explicit reasoning.

What not to expect from search

• Analyzing an attached document. If you already have the data, search is unnecessary; the model does better and cheaper without it.
• Writing prose. Search helps with facts, not with composition.
• Blind trust. The answer is only as good as the web is on your topic. On mainstream topics it is excellent; in narrow Russian-language niches, open the sources and look for yourself.

The practical takeaway

Treat a search-grounded model as a fast research assistant, not an oracle. Give it a precise question, demand links, double-check what matters against the sources. Then it will handle what a regular model cannot: answers tied to today.