Struct TieredRetrievalConfig 

pub struct TieredRetrievalConfig {

    pub enabled: bool,
    pub classifier_provider: ProviderName,
    pub validator_provider: ProviderName,
    pub token_budget: usize,
    pub validation_enabled: bool,
    pub validation_threshold: f32,
    pub max_escalations: u8,
    pub classifier_timeout_secs: u64,
    pub validator_timeout_secs: u64,
    pub similarity_weight: f64,
    pub recency_weight: f64,
    pub recency_half_life_days: u32,
    pub tfidf_weight: f64,
    pub cognitive_signal_weight: f64,
    pub tier_boost_weight: f64,
    pub semantic_tier_boost: f64,
    pub deep_reasoning_query_conditioned: bool,
}

MemFlow tiered intent-driven retrieval configuration.

Classifies each recall query into one of three intent tiers (ProfileLookup, TargetedRetrieval, DeepReasoning) and dispatches to the cheapest sufficient backend. An optional validation step can escalate to a heavier tier when evidence confidence is low.

Example (TOML)

[memory.tiered_retrieval]
enabled = false
classifier_provider = ""
validator_provider = ""
token_budget = 4096
validation_enabled = false
validation_threshold = 0.6
max_escalations = 1
classifier_timeout_secs = 5
validator_timeout_secs = 5

# Signal weights (all default to 0.0; set to activate each signal)
similarity_weight = 1.0
recency_weight = 0.0
recency_half_life_days = 7
tfidf_weight = 0.0
cognitive_signal_weight = 0.0
tier_boost_weight = 0.0
semantic_tier_boost = 1.0

Fields

enabled: bool

Enable MemFlow tiered retrieval. Default: false.

classifier_provider: ProviderName

Provider name from [[llm.providers]] for intent classification.

When empty, the HeuristicRouter is used (no LLM call). When a provider is set but the call fails, falls back to the heuristic (fail-open).

validator_provider: ProviderName

Provider name from [[llm.providers]] for evidence validation.

When empty or when validation_enabled = false, no validation call is made.

token_budget: usize

Maximum tokens to gather for evidence per query. Default: 4096.

validation_enabled: bool

Enable evidence validation and tier escalation. Default: false.

validation_threshold: f32

Confidence threshold below which validation triggers tier escalation. Default: 0.6.

max_escalations: u8

Maximum tier escalations per query. Default: 1.

classifier_timeout_secs: u64

Timeout in seconds for the classifier LLM call. Default: 5.

On timeout the pipeline falls back to the HeuristicRouter (fail-open).

validator_timeout_secs: u64

Timeout in seconds for the validator LLM call. Default: 5.

On timeout the validator is treated as sufficient (fail-open).

similarity_weight: f64

Weight applied to the raw similarity score from vector/keyword recall. Default: 1.0.

Set to 1.0 and all other weights to 0.0 to reproduce pre-signal behaviour.

recency_weight: f64

Weight applied to the recency decay signal. Default: 0.0 (disabled).

recency_half_life_days: u32

Half-life for recency decay in days. Default: 7.

A message that is recency_half_life_days old receives a recency score of 0.5. Set recency_weight = 0.0 to disable recency scoring entirely.

tfidf_weight: f64

Weight applied to the TF-IDF signal. Default: 0.0 (disabled).

cognitive_signal_weight: f64

Weight applied to the cognitive signal (message access frequency). Default: 0.0 (disabled).

tier_boost_weight: f64

Weight applied to the tier boost signal for consolidated/semantic entries. Default: 0.0 (disabled).

semantic_tier_boost: f64

Additive score awarded to entries in the semantic tier when tier_boost_weight > 0. Default: 1.0.

The final contribution is tier_boost_weight * semantic_tier_boost for semantic entries and 0.0 for episodic entries.

deep_reasoning_query_conditioned: bool

Route the DeepReasoning tier graph step through query-conditioned recall (#3994).

When true, the graph recall step for IntentClass::DeepReasoning uses recall_graph_hela (HELA spreading activation) instead of static-weight BFS, producing query-aligned results. Requires an embedding store. Default: false (opt-in).

Trait Implementations

impl Clone for TieredRetrievalConfig

fn clone(&self) -> TieredRetrievalConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TieredRetrievalConfig

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

Formats the value using the given formatter.

impl Default for TieredRetrievalConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for TieredRetrievalConfig

where TieredRetrievalConfig: 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 TieredRetrievalConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.