Struct RetrievalConfig 

pub struct RetrievalConfig {
    pub depth: u32,
    pub search_prompt_template: String,
    pub context_format: ContextFormat,
    pub query_bias_correction: bool,
    pub query_bias_profile_weight: f32,
    pub query_bias_centroid_ttl_secs: u64,
}

Retrieval-stage tuning for semantic memory (MemMachine-inspired, #3340).

Controls ANN candidate depth, search-prompt template, and memory snippet rendering. Nested under [memory.retrieval] in TOML. All fields have defaults so existing configs parse unchanged.

Example (TOML)

[memory.retrieval]
# depth = 0          # 0 = legacy (recall_limit * 2); set ≥ 1 to override directly
# search_prompt_template = ""
# context_format = "structured"

Fields

depth: u32

Number of ANN candidates fetched from the vector store before keyword merge, temporal decay, and MMR re-ranking.

• 0 (default): legacy behavior — recall_limit * 2 candidates, byte-identical to pre-#3340 deployments.
• ≥ 1: the configured value is passed directly to qdrant.search / keyword_search. Set to at least recall_limit * 2 to match the legacy pool size, or higher for better MMR diversity.

A value below recall_limit triggers a one-shot WARN because the ANN pool cannot saturate the requested top-k.

search_prompt_template: String

Template applied to the raw user query before embedding.

Supports a single {query} placeholder which is replaced with the raw query string. Empty string (default) = identity: the query is embedded as-is.

Applied only at query-side embedding sites — stored content (summaries, documents) is never wrapped. Use this for asymmetric embedding models (e.g. E5 "query: {query}").

context_format: ContextFormat

Shape of memory snippets injected into agent context.

See ContextFormat for the exact rendering and token-cost implications. Default: Structured.

query_bias_correction: bool

Enable query-bias correction towards the user’s profile centroid (MM-F3, #3341).

When true and the query is classified as first-person, the query embedding is shifted towards the centroid of persona-fact embeddings. This nudges recall results towards persona-relevant content for self-referential queries.

Default: true (low blast-radius: no-op when the persona table is empty).

query_bias_profile_weight: f32

Blend weight for query-bias correction (MM-F3, #3341).

Controls how much the query embedding shifts towards the profile centroid. 0.0 = no shift; 1.0 = full centroid. Clamped to [0.0, 1.0]. Default: 0.25.

query_bias_centroid_ttl_secs: u64

Centroid TTL in seconds (MM-F3, #3341).

The profile centroid computed from persona facts is cached for this many seconds. After expiry it is recomputed on the next first-person query. Default: 300 (5 min).

Trait Implementations

impl Clone for RetrievalConfig

fn clone(&self) -> RetrievalConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RetrievalConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for RetrievalConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for RetrievalConfig

where RetrievalConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for RetrievalConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.