Release Notes: v0.1.5¶

Date: 2026-04-26

This release is a major memory-system refactor. The biggest change is that Lint-AI now supports single-document insert/update through a mutable IndexStore, while the query path is separated into a frozen MemoryIndex snapshot. In practice, that means callers can incrementally upsert one source document at a time instead of treating the whole corpus as a monolithic rebuild.

The release also adds optional claim extraction, temporal context support, temporal fact storage, refreshed lexical data, and aggregation scaffolding.

Highlights¶

• Added single-document insert/update support through IndexStore.
• Refactored the retrieval system into:
• a mutable ingestion/store layer
• an immutable query snapshot layer
• Added SourceDocument as the primary artifact ingestion type.
• Added stable chunk identifiers and chunk lifecycle tracking.
• Added optional claim extraction, disabled by default.
• Added temporal parsing, temporal fact storage, and timeline reasoning scaffolding.
• Refreshed lexical expansion data from WordNet and ConceptNet subsets.
• Added aggregation scaffolding for count/sum style queries.
• Bumped crate version to 0.1.5.

Memory Index Refactor¶

The core architectural change in this release is the split between ingest and query:

• IndexStore
• mutable artifact-facing store
• owns source documents, cached records, tombstones, and lifecycle metadata
• supports document-level upsert and remove
• rebuilds query state from the current store contents
• MemoryIndex
• immutable query snapshot
• built from the current store state
• optimized for retrieval and reranking

This is a more practical model for artifact-driven workflows because:

• a single document can be inserted or updated independently
• the store can keep lifecycle/history data separate from retrieval state
• queries run against a stable snapshot instead of a moving ingest target

Single-Document Insert and Update¶

Before this release, the memory path was primarily batch-oriented. This release adds the ability to ingest and update one source document at a time.

New primary input¶

SourceDocument is now the main artifact input type. It captures:

• doc_id
• source
• content
• concept
• headings
• links
• timestamp
• doc_length
• author_agent

Ingestion behavior¶

• each source document is normalized into a DocRecord
• the record is chunked and enriched
• the record is inserted into the mutable IndexStore
• the snapshot is rebuilt when needed

Why this matters¶

• callers no longer need to reconstruct the entire corpus for a single change
• the model fits incremental artifact systems better
• doc lifecycle and chunk lifecycle can be tracked explicitly

Chunking and Lifecycle¶

This release formalizes chunk-level structure more clearly.

Chunk model¶

• SectionChunk now carries stable identifiers.
• Chunk timestamps inherit from the parent source document.
• Chunks retain headings, line ranges, entities, and important terms.

Lifecycle model¶

Lifecycle metadata is now tracked at both chunk and document levels:

• ChunkLifecycleMeta
• version
• latest status
• update timestamp
• superseded chunk reference
• lineage key
• DocumentLifecycleMeta
• derived from chunk lifecycle state

This makes it possible to reason about:

• the latest version of a chunk
• the latest version of a document
• how a document changed over time

Optional Claim Extraction¶

Claim extraction is now optional and off by default.

Behavior¶

• top_claims remains part of the record schema.
• Claim extraction only runs when the option is enabled.
• Claim-based retrieval plumbing only activates when claim scoring is enabled.

Why this matters¶

• claim metadata is preserved for workflows that want it
• the default query path stays cleaner and less opinionated
• claim support is available without becoming part of every index build

Temporal Support¶

This release adds the first full temporal layer for timestamped chunks.

Query-side temporal context¶

• temporal queries can be evaluated against a range or anchor
• relative expressions can be interpreted with a time context
• chunk timestamps are used directly in ranking decisions

Document-side temporal data¶

• source document timestamps are propagated into chunks
• temporal terms are stored on records for date-oriented retrieval
• temporal metadata is available without rewriting the core query API

Timeline reasoning¶

The release also adds a temporal timeline layer for ordered event analysis:

• timeline windows over timestamped facts
• adjacent-pair detection
• support for sequence/order style temporal questions

Temporal Facts and State Over Time¶

This release introduces a separate temporal fact layer on top of document and chunk lifecycle.

What it is for¶

• storing structured temporal assertions
• tracking validity windows
• answering “as of” and timeline-style questions
• keeping semantic state separate from raw text

Data model¶

• facts are stored separately from chunks
• facts can be versioned over time
• facts can retain provenance back to source chunk/document ids

Why this matters¶

This is the first step toward state-over-time reasoning rather than simple timestamp-based ranking.

Lexical Expansion Refresh¶

Lexical expansion data was rebuilt and checked into the repository.

Changes¶

• refreshed WordNet subset JSON
• refreshed ConceptNet subset JSON
• added a generator script for offline subset rebuilds
• kept lexical expansion data local and inspectable

Policy¶

• lexical expansion data should be regenerated from upstream sources
• domain vocabulary should be added through seed terms and regeneration
• hardcoded Rust synonym lists should be avoided

Query and Aggregation¶

This release also adds early reasoning scaffolding for aggregation-style questions.

Added support for¶

• count-style aggregation
• sum-style aggregation
• reasoning over retrieved evidence with citations

This is still a generic first pass, but it establishes the infrastructure for questions that need retrieval plus computation.

Documentation¶

This release also adds or expands documentation for:

• artifact-oriented indexing
• chunk strategy behavior
• lexical subset generation
• release notes for the current version family

Compatibility Notes¶

• Existing queries still work through the same public API surface.
• Claim metadata remains on the record schema but is not part of default retrieval behavior.
• Persistent store layouts now validate metadata more strictly, so old incompatible indexes may need to be rebuilt.

Version Bump¶

• Crate version updated from 0.1.4 to 0.1.5