Psearch

Watch out

Several limits are documented rather than worked around. A relevance score does not establish factual accuracy — the tool returns evidence, never an answer — and the default probability threshold is 0.5. The extractor reads the first 8,000 text characters and at most 24 links per page, Markdown pages supply text but their links are not extracted, and login pages, browser challenges or delayed rendering can defeat retrieval. jev_ms is shared by every page in one batch, so summing per-page values does not measure total Jev time; larger batches cut HTTP requests without guaranteeing lower latency or token use, and two redirects in one batch can fetch the same destination. Your query goes to Parallel unless you pass seeds, and page text, the query and link metadata go to TypeSafe for classification, so private URLs do not belong in seeds. The page budget counts failed attempts, so raising depth buys no extra pages. Exit codes are 1 for error and 2 for bad arguments, so check status and errors in JSON and MCP output. Copy reviewed against the repository README on 2026-10-04, not from running it.