Skip to main content
Candidates come from full-text search over statement and cosine similarity over the stored vectors (the query’s vector is embedding, or the server’s embedding of query). The two lists are fused with reciprocal rank fusion. Candidates are then expanded along edges of the chosen families up to depth hops. The as_of and valid_at filters apply to every hop, not only to the candidates.

Request

string
required
Full-text query. Language-agnostic tokens.
timestamp or date
required
string
Scope to one space. Omit for the whole tenant.
timestamp
integer
default:"2"
Expansion hops.
string[]
default:"[\"causal\", \"temporal\"]"
Edge families to expand along. See edge types.
integer
default:"20"
Caps the candidates, not the expanded hits.
string
Narrows the candidates to one topic label. Expansion hops are not filtered: a causal chain may cross topics, and cutting it at a topic boundary would hide the cause.
string
Narrows the candidates to the blocks that state a value for this fact. Like category, expansion hops are not filtered.
object
The case you are asking about, as {key: value} (see list memories for how conditions are checked). Unlike category and fact_key this filters everything: a block whose applies_if does not hold is not a candidate, is not reached by any hop (so nothing enters through it), and does not flag the block it supersedes as superseded_by. Without facts the search is unchanged. The certificate then carries not_applicable.
boolean
default:"false"
Also make predictions and commitments whose window closed by valid_at candidates: the calls that were already settled or past their deadline. Ask for them when the question is how a view changed or turned out. Anything learned after as_of stays out either way, and a claim whose interval closed is still not a candidate. Ended candidates are returned at depth 0; the walk starts from statements in force.
number[]
The query vector. Must come from the same model as the vectors stored on nodes; a node whose vector has another dimension is not a candidate. When absent, the hosted server embeds query with its own model, so a question in other words still finds statements by meaning; a server with no embedding model uses full-text candidates only.

Response

Hit[]
Each hit: node (NodeRow), score, depth, path (node ids from the candidate to this node), via (the edge type that reached it; null for a candidate), superseded_by (the id of the block that superseded it, through a SUPERSEDES edge known by as_of and in force at valid_at, or null), superseded_since (when that supersession took effect, the edge’s valid_from; null when superseded_by is). score is the candidate’s fused rank score multiplied by 0.5 per hop.
object or null
object
A superseded block is returned and flagged, not dropped. Your agent should see what it used to believe, and that it has been replaced. A supersession announced before it takes effect does not flag the block yet: read with valid_at on the day it takes effect and superseded_by appears, with superseded_since saying when.