Perplexity Sonar and web-search-grounded models: how it works and what Yenisei offers

Perplexity Sonar is a lineup of models that work differently from regular chat models. Sonar answers from web search results rather than from "memory", and attaches links to its sources. So the answer is fresh and can be verified.

There are no Sonar models in the Yenisei catalog. Below is how this approach works and how to get answers grounded in fresh data with us.

The Sonar lineup

• Sonar is the base model, for quick answers to factual questions.
• Sonar Pro is the senior one, better on complex and multi-part queries.
• Sonar Reasoning Pro adds reasoning on top of search: it reconciles what it found and draws conclusions.

How this differs from a regular model

A regular LLM knows the world as of its training date, and on a new question ("what came out last week") it will either honestly give up or make something up. A search-grounded model goes to the web, gathers fresh sources and answers from them, attaching links. For tasks where freshness matters, there is no way around search.

In essence it is a pairing of two parts: search finds the pages, the model reads them and answers. Each part can be chosen separately.

How to get the same in Yenisei

In Yenisei Chat and Yenisei Agent, web search is already built in: the answer is built from fresh data with links to sources. Nothing to configure.

Through the API you assemble the pairing yourself: you give the model a search function, it decides what to search for, gets the results and answers from them. This needs a model with confident tool calling:

• Kimi K2 Thinking (kimi-k2-thinking) reasons and reconciles sources;
• DeepSeek V4 Pro (deepseek-pro) has a long context for many retrieved pages;
• Qwen3.8 Flash (qwen-flash) is fast and inexpensive for simple factual questions.

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k2-thinking",
    "messages": [
      {"role": "system", "content": "Answer only from the search results. After every fact, add a link to the source."},
      {"role": "user", "content": "What changes to product labeling requirements came into force this year?"}
    ],
    "tools": [{
      "type": "function",
      "function": {
        "name": "web_search",
        "description": "Search the web for pages matching the query and return titles, URLs and snippets",
        "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]}
      }
    }]
  }'

The web_search function is executed by your code: it queries the search engine you use and returns the results to the model as a role: tool message.

Where search belongs and where it does not

Search is needed when the answer requires freshness and sources: news and market monitoring, gathering information on a topic, "what is the current state of..." questions.

For tasks where the web is not needed at all, such as rewriting text, analyzing an attached document, or generating code, search is superfluous: take a regular model. And remember that answer quality is bounded by the quality of what was found: on niche Russian-language topics, the sources are worth double-checking.