Struct PlanCache 

pub struct PlanCache { /* private fields */ }

Plan template cache backed by SQLite with in-process cosine similarity search.

Stores embeddings as BLOB columns and computes cosine similarity in-process (same pattern as ResponseCache). Graceful degradation: all failures are logged as WARN and never block the planning critical path.

Implementations

impl PlanCache

pub async fn new( pool: DbPool, config: PlanCacheConfig, current_embedding_model: &str, ) -> Result<Self, PlanCacheError>

Create a new PlanCache and invalidate stale embeddings for the given model.

Errors

Returns PlanCacheError if the stale embedding invalidation query fails.

pub async fn find_similar( &self, goal_embedding: &[f32], embedding_model: &str, ) -> Result<Option<(PlanTemplate, f32)>, PlanCacheError>

Find the most similar cached plan template for the given goal embedding.

Fetches all rows with matching embedding_model, computes cosine similarity in-process, and returns the best match if it meets similarity_threshold.

Also updates last_accessed_at on a hit.

Errors

Returns PlanCacheError::Database on query failure or PlanCacheError::Serialization on template JSON deserialization failure.

pub async fn cache_plan( &self, graph: &TaskGraph, goal_embedding: &[f32], embedding_model: &str, ) -> Result<(), PlanCacheError>

Store a completed plan as a reusable template.

Extracts a PlanTemplate from the TaskGraph, serializes it to JSON, and upserts into SQLite using INSERT OR REPLACE ON CONFLICT(goal_hash). Deduplication is enforced by the UNIQUE constraint on goal_hash.

Errors

Returns PlanCacheError on extraction, serialization, or database failure.

pub async fn evict(&self) -> Result<u32, PlanCacheError>

Run TTL + size-cap LRU eviction.

Phase 1: Delete rows where last_accessed_at < now - ttl_days * 86400. Phase 2: If count exceeds max_templates, delete the least-recently-accessed rows.

Returns the total number of rows deleted.

Errors

Returns PlanCacheError::Database on query failure.