RagLeap v0.13.1 and ragleap-memory 0.3.0: Better Action Auditing and Optional Vector Search

RagLeap has released two updates aimed at enhancing AI system reliability through better action auditing and optional vector search capabilities. The first release, ragleap-core v0.13.1, introduces structured JSON audit log lines for agent actions, providing a detailed record of what agents do without storing the actual text content.

The audit log entries include structured information such as timestamps, action types, channels, roles, targets, approval/rejection details, action IDs, and SHA-256 hashes of content and results. However, certain actions like GET reads, chat routes, and database table controls are not yet covered by the audit logs. Upgrading to this version requires rebuilding and restarting the application.

The second release, ragleap-memory 0.3.0, introduces optional vector and hybrid search functionalities, with keyword search remaining the default. Developers must supply their own embedder function to control the embedding process. The library supports three search modes: keyword, vector (cosine similarity), and hybrid (combining keyword and vector results).

The embed_pending() method allows developers to generate vectors for existing entries, with options for scope selection, batch limits, dry-run mode, and replacing stale vectors. A daily embedder-token cap per tenant has been added to manage resource usage. The library also introduces EmbedderError for handling failed or unusable embedder responses, along with database migration requirements and usage counters for embedded tokens and embedding calls.

However, stored vectors are not encrypted, and vector search scans only a limited number of entries, with quality depending on the selected embedder and not yet measured.