The server provides three well-named, verb-based tools (search, ask, ask_more) with comprehensive input schemas and detailed parameter descriptions. All tools declare ToolAnnotations with readOnlyHint=True and idempotentHint=True, showing good security awareness. However, output schemas are not formally documented, only a docstring mentions 'Returns: Searchable results' without specifying structured fields. Error handling exists but is generic (HTTPStatusError caught and returned as plain text). The domain_filter parameter uses a sophisticated design accepting both include/exclude domains via '-' prefix, but the mechanism is underdocumented in descriptions. Parameter descriptions are well above the 20-char minimum and often exceed the 72-char baseline. The recency parameter resolution logic is complex and well-handled server-side, reducing LLM cognitive load. Missing: formal output schema documentation, per-parameter error guidance, and explicit rate-limit or timeout constraints.

• Formally document output schemas in tool descriptions. Example for search(): 'Returns a markdown-formatted list of search results, each with a number, title, URL, and snippet. Format: "1. Title\n URL: https://...\n snippet text\n\n"'
• Implement structured error responses. Instead of returning plain strings, catch exceptions, categorize them (e.g., {"error": "auth_failed", "message": "Invalid API key", "action": "Check PERPLEXITY_API_KEY in environment"}), and guide the LLM on next steps.
• Add validation and error guidance for domain_filter. If a domain is malformed, return a specific error: 'Invalid domain filter: "example..com", domains must be lowercase, alphanumeric with dots, no hyphens at start/end.'
• Document timeout and rate-limit behavior. Add a note in tool descriptions: 'Note: API calls may timeout after 60 seconds if Perplexity is slow. If you receive a timeout, wait 10 seconds before retrying. Calling search() more than 10 times per minute may hit rate limits.'
• Unify enum order with description order for 'sources' parameter. Change enum to ["web", "academic", "sec"] to match the description sequence.
• Add a 'search_depth' or similar parameter to ask() and ask_more() to give LLMs explicit control over API call expense (quick vs thorough), rather than relying on 'scope' and 'thoroughness' which map inconsistently to backend behavior.
• Return structured JSON from tools instead of formatted strings. This enables reliable parsing by downstream agents and reduces hallucination risk. Example: {"results": [{"rank": 1, "title": "...", "url": "...", "snippet": "..."}]}