pub struct RecallRequest {Show 18 fields
pub query: Vec<f32>,
pub top_k: usize,
pub memory_types: Option<Vec<MemoryType>>,
pub agent_id: Option<String>,
pub metric: Metric,
pub weights: ScoreWeights,
pub n_probe: Option<usize>,
pub n_rerank: Option<usize>,
pub track_access: bool,
pub filter: Filter,
pub visibility: VisibilitySet,
pub max_bytes: Option<usize>,
pub keywords: Option<String>,
pub text: Option<String>,
pub keyword_weight: f32,
pub text_keyword_weight: f32,
pub bm25_k1: f32,
pub bm25_b: f32,
}Expand description
A retrieval request for Mnemo::recall.
Fields§
§query: Vec<f32>Query embedding.
top_k: usizeMaximum results to return.
memory_types: Option<Vec<MemoryType>>Restrict to these memory types (None = all types).
agent_id: Option<String>Restrict to a single agent’s view: its own memories plus shared ones.
metric: MetricSimilarity metric.
weights: ScoreWeightsMulti-signal score weights.
n_probe: Option<usize>Index override: partitions to probe (None = index default). Ignored
when no ANN index is present.
n_rerank: Option<usize>Index override: candidates to rerank (None = index default). Ignored
when no ANN index is present.
track_access: boolWhether to update each result’s accessed_at and access_count in
the catalog. Defaults to true, matching pre-v5 behavior.
Set to false for fully read-only recall — useful for batch
scoring, dry-run scoring, or recalls run by introspection tooling
(e.g. mnemo about consumers) where you don’t want the score’s own
observation to perturb the database. With false, recall does not
dirty the catalog and the next flush is a no-op.
filter: FilterPhase 12.6 pre-scoring filter: session_id + AND-ed metadata
conditions applied before the candidate is scored. Default is
Filter::default (no constraints). When the ANN index is on
and this filter is selective enough to starve the shortlist
below top_k, Phase 12.1’s widen-then-exact-scan logic kicks
in so the result still carries up to top_k correct hits.
visibility: VisibilitySet13a (13.5): the visibility levels recall returns: public and team by
default; private only when asked (RecallRequest::visibility).
max_bytes: Option<usize>Phase 12.8 post-scoring budget: stop adding hits once the UTF-8
bytes of their content fields (not the serialized record
body) would exceed this many bytes. None (default) means no
budget and every hit through top_k is returned. If the top-1
hit is already over budget it is still returned — callers that
need an absolute cap should also pre-size top_k.
Counts only the UTF-8 content bytes so the number is stable
across the Rust / Python / CLI / MCP surfaces; the JSON or
MessagePack envelope is excluded. Roughly 4 bytes per BPE
token for English text — a 4 KB budget comfortably holds ~1K
tokens.
Observable through RecallResponse::truncated
(Mnemo::recall_budgeted): true iff at least one scored
hit was dropped by this budget, false otherwise.
keywords: Option<String>FMT-9d (Phase 2, v9 files): identifiers or exact terms to match
(a version, a PR number, a function name) for hybrid recall. BM25
scores, divided by this query’s best, are added to each candidate’s
similarity with weight RecallRequest::keyword_weight before the
usual scoring. With an empty query vector
(RecallRequest::keywords_only), recall is keyword-only.
Since v0.15 (12.3 step 4), a recall by text is hybrid by default:
with keywords None, the request’s RecallRequest::text is
used for BM25, so give keywords only for identifiers the question
doesn’t contain. Keywords given replace the text. Empty keywords
(.keywords("")) opt out: vector-only recall. None on a request
with no text is plain vector recall.
text: Option<String>12.3 step 4 (v0.15): the query text a recall by text was built from
(RecallRequest::from_text sets it). With keywords None it’s
also scored with BM25 on a v9 file (hybrid by default); a v8 file,
which has no keyword index, ignores it.
keyword_weight: f32FMT-9d: weight ε of the normalised BM25 score in hybrid recall
(default DEFAULT_KEYWORD_WEIGHT), for keywords given.
text_keyword_weight: f3212.3 step 4 (v0.15): the weight when BM25 scores the request’s
RecallRequest::text (no keywords given); default
DEFAULT_TEXT_KEYWORD_WEIGHT.
bm25_k1: f32FMT-9d: BM25 term-frequency saturation k1 (default 1.2).
bm25_b: f32FMT-9d: BM25 length normalisation b (default 0.75).
Implementations§
Source§impl RecallRequest
impl RecallRequest
Sourcepub fn visibility(self, v: VisibilitySet) -> Self
pub fn visibility(self, v: VisibilitySet) -> Self
13a (13.5): the visibility levels to return (private records only when the set includes them).
Sourcepub fn keywords_only(keywords: impl Into<String>) -> Self
pub fn keywords_only(keywords: impl Into<String>) -> Self
FMT-9d (v9 files): a keyword-only request: BM25 over keywords
(identifiers or exact terms), no query vector, so it works without
an embedder.
Sourcepub fn keywords(self, text: impl Into<String>) -> Self
pub fn keywords(self, text: impl Into<String>) -> Self
FMT-9d (v9 files): add identifiers or exact terms to match (hybrid
recall). They replace the request’s text for BM25; "" opts out of
keyword scoring. See the keywords field.
Sourcepub fn text_keyword_weight(self, w: f32) -> Self
pub fn text_keyword_weight(self, w: f32) -> Self
12.3 step 4: the weight of the text’s keyword score when no keywords
are given (see RecallRequest::text_keyword_weight).
Sourcepub fn keyword_weight(self, w: f32) -> Self
pub fn keyword_weight(self, w: f32) -> Self
FMT-9d: the weight ε of the keyword score in hybrid recall.
Sourcepub fn filter(self, filter: Filter) -> Self
pub fn filter(self, filter: Filter) -> Self
Phase 12.6: attach a pre-scoring filter. See Filter and
FilterTerm for term shapes.
Sourcepub fn max_bytes(self, n: usize) -> Self
pub fn max_bytes(self, n: usize) -> Self
Phase 12.8: cap the total UTF-8 bytes of returned content
fields. See RecallRequest::max_bytes.
Sourcepub fn from_text(embedder: &dyn Embedder, query_text: &str) -> Result<Self>
pub fn from_text(embedder: &dyn Embedder, query_text: &str) -> Result<Self>
Build a request by embedding query_text with the supplied
embedder (Phase 3). Since v0.15 it embeds with
Embedder::embed_query and keeps the text in
RecallRequest::text, so the recall is hybrid by default (see
RecallRequest::keywords; .keywords("") opts out). Otherwise the
same defaults as new. Fails if the embedder returns an empty vector
list or an embedding failure.
Sourcepub fn text(self, text: impl Into<String>) -> Self
pub fn text(self, text: impl Into<String>) -> Self
12.3 step 4: set the query text (see RecallRequest::text), for
a caller that embedded it itself.
Sourcepub fn types(self, t: Vec<MemoryType>) -> Self
pub fn types(self, t: Vec<MemoryType>) -> Self
Restrict to specific memory types.
Sourcepub fn weights(self, weights: ScoreWeights) -> Self
pub fn weights(self, weights: ScoreWeights) -> Self
Replace the multi-signal score weights.
Sourcepub fn n_probe(self, n: usize) -> Self
pub fn n_probe(self, n: usize) -> Self
Override the number of IVF partitions probed (accuracy/speed dial).
Sourcepub fn n_rerank(self, n: usize) -> Self
pub fn n_rerank(self, n: usize) -> Self
Override the number of candidates reranked exactly (accuracy dial).
Sourcepub fn track_access(self, track: bool) -> Self
pub fn track_access(self, track: bool) -> Self
Toggle whether recall updates accessed_at / access_count on
the returned memories’ catalog entries (default true). Pass
false for a fully read-only recall — see the field doc.