Crate veclayer 

VecLayer

Hierarchical vector indexing for documents with perspective-based memory.

VecLayer provides semantic search over documents with hierarchical context, a persistent memory layer for AI agents, and automatic knowledge aging — important knowledge stays, unused knowledge fades.

Core Concepts

• Hierarchical chunks: Documents are split into chunks organized at different levels (document, section, paragraph, fragment)
• Perspectives: Different views of the same knowledge store (e.g., decisions, learnings, intentions)
• Salience scoring: Retrieval blends semantic similarity, recency, access frequency, and reinforcement
• Aging: Entries naturally decay unless accessed or reinforced

Feature Flags

• llm (default): Enables LLM-powered summarization and clustering
• sync: Enables cross-store synchronization

Re-exports

pub use blob_store::BlobStore;

pub use chunk::content_hash;

pub use chunk::relation;

pub use chunk::short_id;

pub use chunk::visibility;

pub use chunk::AccessProfile;

pub use chunk::ChunkLevel;

pub use chunk::ChunkRelation;

pub use chunk::ClusterMembership;

pub use chunk::EntryType;

pub use chunk::HierarchicalChunk;

pub use chunk::RecencyWindow;

pub use chunk::STANDARD_VISIBLE;

pub use config::Config;

pub use embedder::Embedder;

pub use entry::EmbeddingCache;

pub use entry::Entry;

pub use entry::StoredBlob;

pub use error::Error;

pub use error::Result;

pub use parser::DocumentParser;

pub use salience::SalienceWeights;

pub use search::HierarchicalSearch;

pub use search::HierarchicalSearchResult;

pub use search::SearchConfig;

pub use store::StoreBackend;

pub use store::VectorStore;

Modules

access_profile

RRD-style access tracking for memory aging.

aging

Agent-configurable aging rules for automatic visibility degradation.

blob_store

Git-style content-addressed blob store.

chunk

Core domain types for VecLayer’s hierarchical chunk model.

config

Configuration with 12-factor layered resolution: ENV > TOML file > Defaults.

embedder

Embedding trait and backend implementations.

entry

Entry: the portable, serializable form of a HierarchicalChunk.

error

Error types for VecLayer operations.

identity

Identity: emergent self-model from memory structure.

parser

Document parsing traits and implementations.

perspective

Perspectives: named lenses over the same knowledge.

relations

Shared relation processing for CLI and MCP paths.

resolve

ID resolution and temporal parsing shared by CLI commands and MCP tools.

salience

Salience scoring: how “important” an entry is beyond access frequency.

search

Hierarchical search engine with blended salience scoring.

store

Vector store trait and backend implementations.

Macros

arc_impl

Macro to implement traits for Arc where T implements the trait. This eliminates boilerplate Arc delegation code across trait modules.

Structs

ClusterPipeline

Pipeline for RAPTOR-style clustering and summarization.

OllamaSummarizer

Summarizer using Ollama for local LLM inference. Free and runs locally - perfect for RAPTOR-style summarization.

SoftClusterer

Soft clusterer using K-Means with distance-based soft assignments.

Enums

LlmBackend

Enum-based dispatch for LLM providers. Avoids trait objects while supporting multiple backends.

Traits

LlmProvider

Trait for LLM text generation.

Summarizer

Trait for generating text summaries. Used to summarize clusters of related chunks.

Functions

default_data_dir

Platform-appropriate default data directory for VecLayer.