Struct BrainSession 

pub struct BrainSession { /* private fields */ }

Implementations

impl BrainSession

pub fn config(&self) -> &ProjectConfig

pub fn resolve_request_floors(&self, request: &mut ContextRequest)

Resolve explicit overrides before legacy sentinels. The explicit zero survives a second resolution at the injected/production boundary.

pub fn open(start: &Path) -> KimetsuResult<Self>

pub fn open_readonly(start: &Path) -> KimetsuResult<Self>

pub fn retrieve_context( &self, stage: &str, query: &str, budget_tokens: u32, ) -> KimetsuResult<ContextBundle>

pub fn retrieve_context_with_request( &self, request: ContextRequest, ) -> KimetsuResult<ContextBundle>

v0.6: full-request variant used by kimetsu_brain_context MCP tool and retrieve_context_readonly_with_request to expose tags, min_score, max_capsules, and prefer_roles.

W3.1: routes through open_embedder_for so the persistent [embedder] enabled = false config field truly disables the cosine path (FTS-only retrieval) without relying on the env var.

pub fn resolved_abstain_evidence(&self) -> f32

v2.7: resolve the absolute abstention floor for this session’s config. See resolved_abstain_evidence_for.

pub fn retrieve_proactive( &self, request: ContextRequest, ) -> KimetsuResult<ContextBundle>

v0.8: proactive (mid-work) retrieval. Pins [NoopEmbedder] so it stays lexical-FTS-only — NO embedding model is loaded even in --features embeddings builds, keeping the per-tool-call hook cheap. request.kinds should restrict to actionable kinds; the caller sets a high min_score and max_capsules: 1 so recall is rare and confident (the human-brain “it comes to you” model).

pub fn retrieve_context_lexical( &self, request: ContextRequest, ) -> KimetsuResult<ContextBundle>

v1.0.0: lexical (FTS-only) retrieval honoring the full ContextRequest. Like Self::retrieve_context_with_request but pins [NoopEmbedder] so NO embedding model is loaded even in --features embeddings builds. The UserPromptSubmit context-hook uses this: it runs in a throwaway per-prompt process that cannot reuse the long-lived MCP server’s warm model cache, so a cold ONNX load there can blow the host’s 30s hook timeout. Semantic ANN recall stays with the warm MCP kimetsu_brain_context tool.

pub fn retrieve_context_with_injected_embedder( &self, request: ContextRequest, embedder: &dyn Embedder, ) -> KimetsuResult<ContextBundle>

v1.0.0: read-only retrieval honoring the full ContextRequest but with a caller-supplied embedder. The warm embedder daemon uses this to run cosine/ANN retrieval with ONE long-lived model instead of opening a fresh embedder per request. The lexical-coverage floor is applied here too (driven from config unless the caller overrode it).

pub fn repo_root(&self) -> &Path

pub fn user_conn(&self) -> Option<&Connection>

v0.4.1: expose the user-brain connection so callers (e.g. kimetsu brain status) can report counts/paths without re-opening the file. Returns None when the user brain is disabled or unresolvable.