By default, the agent uses the Sonar model, a lightweight and cost-effective search model built on Meta Llama 3.3 70B. You can also configure Sonar Pro for complex multi-step queries that need deeper reasoning. The model is set in the tool configuration and can be changed anytime through the Tars dashboard.

Standard LLMs generate answers from training data that has a knowledge cutoff. Perplexity AI searches the live web before generating a response, so answers reflect information published today, not months ago. It also provides citations so users can verify the sources. Your agent gets real-time accuracy instead of stale knowledge.

Yes. The temperature parameter controls randomness. Set it close to 0 for factual, deterministic answers ideal for customer support. Set it higher for more creative responses useful in brainstorming or content generation. You can also use system prompts to instruct Perplexity to stay concise, formal, or technical.

Citations can be enabled through the return_citations parameter. When active, every Perplexity response includes links to the web sources that informed the answer. Your agent can present these to the customer for transparency, or use them internally for quality assurance.

No. Tars sends the query to Perplexity AI in real-time and uses the response within the current conversation. Search results and generated text are not stored in a separate database. The conversation log will contain the agent's final response, but raw Perplexity data is not persisted.

Perplexity AI charges per token on a pay-as-you-go basis. Sonar costs $1 per million input and output tokens. Sonar Pro costs $3 per million input tokens and $15 per million output tokens. Each customer query uses tokens based on the question length and response depth. You can control costs with the max_tokens parameter.

Yes. Perplexity AI supports domain filtering. You can configure the agent to search only specific domains like your documentation site, trusted industry publications, or government sources. You can also exclude domains to prevent results from competitors or unreliable sources.

If the web search returns no relevant results, Perplexity still generates a response based on its training knowledge but flags lower confidence. Your agent can be configured to fall back to your internal knowledge base or escalate to a human agent when Perplexity indicates low search relevance.