SearchPipe

Run an AI-powered web search for a query: multi-engine retrieval → page fetching → LLM reranking → optional AI answer. Uses the same commercial pipeline and accepts the same parameters as the HTTP POST /search endpoint: authentication → rate limiting → input moderation → credit charge → search → output moderation + AI-generated marker → refund on failure → usage logging. Credits are charged per call. search_depth: "basic" (default, 1 credit) / "advanced" (2 credits) / "fast" (0.5 credit — skips all LLM steps: results keep engine order with engine-provided relevance scores, no answer is generated). include_domains / exclude_domains: restrict or exclude domains ("example.com" matches the domain and any subdomain; "*.example.com" matches subdomains only; "example.com/docs" also requires the URL path to start with /docs; exclude wins over include). time_range: recency window "day" / "week" / "month" / "year"; start_date / end_date: ISO YYYY-MM-DD bounds (results without a known publish date are not filtered). topic: "general" (default) or "news" (always adds a dedicated news channel to retrieval). country: ISO 3166-1 alpha-2 region bias (e.g. "US"); language: upstream market language (e.g. "en", "zh-CN"); combined into the upstream market locale. safe_search: 0 off (default) / 1 moderate / 2 strict. include_images: also run an image search and return an images array alongside results. output_type: "search_results" (default) / "answer" (same as include_answer=true) / "structured" (grounded JSON object per output_schema — requires output_schema; on failure output is null). objective: optional one-line search intent from the calling agent; improves reranking only. stream is REST-only and not available over MCP. api_key: an API key starting with `sp-`. It may also be provided through the SEARCHPIPE_API_KEY environment variable (the parameter takes precedence). Returns Tavily-style structured results (query / answer / results[] / images[] / output). Returns a tool error on authentication failure, insufficient credits, content violation, invalid parameters (422), or retrieval failure.

Verify a factual claim against retrieved evidence: claim in, verdict out — a four-label judgment (supported / partial / unsupported / inconclusive), not generated text. Every result ships with verifiable evidence — span-level quotes, claim verdicts, and source authority — not just links. Verdict labels: `supported` = evidence text clearly supports the claim; `partial` = partly supported or weak evidence; `unsupported` = reachable evidence does not support the claim (misattributed citation detected); `inconclusive` = the claim could not be judged (fetch failure, paywall, or insufficient evidence) — the tool reports that honestly instead of guessing. Verification fetches source pages and judges claim-evidence consistency, so it typically takes 5-15 seconds. Uses the same commercial pipeline as the HTTP POST /verify endpoint: authentication → email verification → rate limiting → input moderation → per-request credit charge → verification → refund on failure → usage logging. Credits are charged per call; an inconclusive verdict is still a delivered result and is not refunded. mode: evidence route — "paper" / "web" / "auto" (default; a DOI or arXiv signal decides). evidence_urls: up to 5 caller-supplied citations to verify. max_evidence: maximum evidence items to return (1-10, default 5). api_key: an API key starting with `sp-`. It may also be provided through the SEARCHPIPE_API_KEY environment variable (the parameter takes precedence). Returns VerifyResponse (verdict / confidence / reason_code / evidences[] / timing). Returns a tool error on authentication failure, insufficient credits, content violation, verification failure, or when the verify endpoint is disabled ("Verify endpoint is not available").