web_search · unlob Docs

Behaviour

The query supports boolean operators: AND, OR, -term to exclude, and "exact phrase". Name a vertical to disambiguate, or omit it and the router picks one from the query.

The index is admission-controlled, so a small result set usually means the index holds little on the topic, not that the query was wrong. One exception: if `partial` comes back true, part of the corpus was unreachable and the answer really is incomplete.

Every filter below is also a query parameter on GET /search, with the same name and the same meaning. IN-list filters accept either an array or a comma-separated string.

Arguments

inputSchema

Name	Type	Description
`authority`	array \| string	Publisher-authority IN-list
`collapse`	"none" \| "host" \| "page" \| "story"	Group results: one per host, per page, or per story (near-dup cluster). collapse=story is how you stop reading the same wire copy five times.
`community`	array \| string	Graph signal: restrict to these graph communities
`content_type`	array \| string	IN-list: article, news, docs, reference, qa, forum, academic, product, code, how-to, opinion, video, document
`exclude_site`	array \| string	Hosts to exclude
`facets`	boolean	Also return host/vertical/source/content-type facet counts and a publish-year histogram
`fields`	array \| string	Project the hit down to these fields, to save context
`from`	integer	Only results fetched at/after this unix-second timestamp
`lang`	string	Single language code, e.g. en
`langs`	array \| string	Language IN-list
`limit`	integer	0 returns only the match count
`max_words`	integer	Passage length ceiling
`min_centrality`	number	Graph signal: only passages this central to the corpus
`min_host_rank`	number	Trust floor (0..1)
`min_independent_sources`	integer	Graph signal: only stories asserted by at least this many distinct hosts. The corroboration filter.
`min_quality`	integer	Quality floor
`min_words`	integer	Passage length floor
`mode`	"keyword" \| "semantic" \| "hybrid"	—
`prefer_authority`	boolean	Soft rerank bias toward higher-authority hosts
`prefer_recent`	boolean	Soft rerank bias toward newer passages — a nudge, not a filter
`published_from`	integer	Content publish date at/after this unix-second timestamp (not crawl time)
`published_to`	integer	Content publish date at/before this
`query`required	string	Boolean-capable: rust AND async, -python, "exact phrase"
`safe`	boolean	true (the default) drops explicit content; false allows everything
`site`	string	Restrict to an exact host, e.g. docs.rs
`sort`	"relevance" \| "recency" \| "host_rank" \| "quality" \| "published" \| "words" \| "centrality"	—
`source`	"cc" \| "delta"	cc = Common Crawl, delta = freshly crawled
`term`	string	Exact-needle: only passages carrying this salient term, e.g. CVE-2024-3094. Use it for identifiers an embedder is blind to.
`tld`	array \| string	Top-level-domain IN-list, e.g. edu, gov
`to`	integer	Only results fetched at/before this unix-second timestamp
`topic`	array \| string	Topic-tag IN-list
`vertical`	string	e.g. code, cooking, science. Omit to auto-route.

Returns

A JSON payload in the text content block, and the same payload as structuredContent for clients that support it.

outputSchema

Name	Type	Description
`facets`	object	Present only when facets=true
`mode`required	"keyword" \| "semantic" \| "hybrid"	The mode actually used
`partial`	boolean	true when at least one shard failed or timed out — the answer is short because a slice of the corpus was unreachable, NOT because that is all there is. Never treat a partial result as complete.
`results`required	object[]	—
`routed`required	boolean	true when the router chose the vertical, false when you named it
`total`required	integer	—
`vertical`	string \| null	The vertical actually queried

Calling it

Over Streamable HTTP, with the required arguments filled in:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "web_search",
    "arguments": {
      "query": "<query>"
    }
  }
}

Most clients do this for you — see Connect a client. To drive it by hand, complete the startup lifecycle first.