v2.0.3

Asset Catalogs

Kavach records versioned AI assets as governed inputs to experiments, evaluations, replay, and decisions.

Overview

The Assets workspace is a registry-centric control-plane surface for prompt, model, dataset, and evaluation-provider records. It is designed to answer a practical operator question: which exact version was used, where did it come from, and what governed work references it?

Every asset version has the same workflow in Studio: Overview, Versions, References, Lineage, and Audit History. Lineage comes from the backend ontology projection; the browser renders graph results but does not infer dependencies itself.

Ownership and Provenance

Kavach distinguishes asset ownership instead of implying that every record was authored in the platform. Each persisted asset version carries one of these provenance values:

  • Managed — created, stored, or lifecycle-managed by Kavach.
  • Observed — reported by runtime or evaluation evidence.
  • Imported — reserved for future external-registry synchronization.

Observed records can retain a source system and source reference. This makes the catalog appropriate for systems that do not allow prompt text or runtime configuration to leave their execution boundary.

Asset Catalogs

  • Prompt Catalog records observed prompt identities, versions, variables, hashes, optional content, and source metadata.
  • Model Catalog records observed provider, model, revision, context window, and runtime parameters.
  • Dataset Registry records immutable evaluation dataset versions, storage locations, checksums, schemas, and record counts.
  • Evaluation Providers lists Kavach-managed adapters and their supported metrics and execution capabilities.

Dataset Artifacts

Dataset bytes live outside the relational registry. Studio accepts UTF-8 CSV, JSONL, and NDJSON uploads, validates them in a stream, calculates a SHA-256 checksum and record count, then writes immutable content to an S3-compatible object store before registering the dataset version.

The local Docker stack uses SeaweedFS through its S3 API. Production can use AWS S3 or another compatible object store without changing the registry contract. Object keys are deterministic and tenant-scoped; duplicate content and conflicting versions are rejected explicitly.

References and Lineage

References list direct ontology relationships for the selected version. Lineage opens a bounded neighbourhood that may connect an asset to experiment candidates, evaluation runs, replay jobs, policies, and governance decisions. It is traceability context, not a cost, quality, or operational analytics dashboard.

Learn how to submit runtime evidence in the Observed Asset Ingestion tutorial.

OSS Boundary

Versioned asset management, not MLOps

OSS Assets intentionally excludes model serving, dataset labeling, cost analytics, quality rankings, recommendation engines, approval workflows, dependency heat maps, and vendor-specific external-registry connectors.