QilbeeDB 0.14.0

Status: released. QilbeeDB 0.14.0 adds evidence-bound knowledge, company learning discovery and graph consolidation contracts. Check /health and /openapi.json on your server before using these capabilities. Self-hosted installations require an upgrade; console availability is separate from API availability.

Capabilities in this release#

• Canonical graph property indexing finds equal maps and signed-zero values. Existing property indexes are rebuilt before readiness. This changes the durable index format: never open the upgraded directory with an older binary. Rollback requires a complete pre-upgrade backup and reconciliation of later writes. See property index upgrades.

• Rust graph transactions validate their observed entities before committing. Stale point reads or first-write observations now abort without partial index or entity changes. Callers must handle TransactionAborted; the database does not replay application actions. See atomic graph transactions for compatibility and isolation limits.

• Company agent display profiles provide human-readable names with revisioned updates and retained history. The application still owns agent identifiers.

• Company learning discovery and evidence history let administrators inspect authorized resources without manually entering resource IDs or issuing scoped credentials. Retained history does not establish current reuse eligibility.
• Consumer diagnostics expose scoped progress, pending events and incompatible history while preserving private subject boundaries and current administrative authorization.
• External graph consolidation provides bounded, leased work, exact source references, publication fencing and explicit handling of unknown outcomes. The external worker owns inference; the database validates publication and retains its evidence.
• Evidence-bound knowledge binds new procedural revisions to current memory evidence and external tool identity declarations. Qualification and current applicability remain distinct. Tool code, maintenance, publication and execution belong to the consuming application.

The knowledge endpoints use request/success contract version 2 under the existing /api/v1 prefix. Errors retain the platform version 1 envelope. Existing v1 procedures are not silently upgraded, and v1 selection cannot return v2 knowledge. The database does not infer an executable implementation from a tool schema.

Existing cosine, lexical, hybrid and graph-assisted score meanings and opt-in behavior remain unchanged. This release does not promote a retrieval method or claim an improvement in agent reasoning. Embeddings remain externally supplied.

Upgrade your installation#

1. Read the storage compatibility notes and take an application-consistent backup. Test restoration and the new image against a separate copy before replacing your live service.
2. Preserve company, account, credential, memory, relation and checkpoint state. Verify that existing clients retain their permissions and score semantics.
3. Coordinate a maintenance window with your application's operators. Stop writers and automatic retries before the storage transition, then resume them with their original operation identities and checkpoints.
4. After the update, check /health, the served OpenAPI and authenticated reads and writes in a dedicated validation scope. Investigate uncertain outcomes before retrying operations.

The property-index format changes in this release. Do not open the upgraded storage directory with an older binary. Once new writes have been accepted, restoring an earlier backup requires reconciling those writes. See property index upgrades and backup and recovery.

Compatibility and limits#

Existing cosine, lexical, hybrid and graph-assisted search retain their score meanings and opt-in behavior. Embeddings remain externally supplied. Hybrid and graph-assisted retrieval remain experimental; choose a method using evidence from your own workload rather than assuming that a new release improves ranking.

Procedural observations are time-bounded, not execution reservations. Selection scans at most 1,000 procedure candidates and has no continuation cursor; an incomplete candidate scan returns the baseline. External execution, cancellation and business authorization remain application responsibilities.

Process-restart recovery does not establish storage-hardware or host power-loss guarantees. Assess those properties for your deployment and storage provider. Functional validation does not demonstrate better agent reasoning or formal accessibility conformance.