Knowledge-Centric Information Systems

🕒 Published (v1): 2026-07-01 15:05 UTC · Source: Arxiv · link

Why this paper was selected

Data engineering governance principles applied to LLM-era systems; architectural framing

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This vision/position paper argues that AI systems which operationalize heterogeneous organizational knowledge (documents, runbooks, code, decisions, etc.) require a new architectural discipline — knowledge architecture — that generalizes data engineering's guarantees (ingestion, lineage, governance, quality) to a richer managed unit: the knowledge artifact. It maps eleven established data-engineering guarantees to semantically-shifted knowledge-level analogues and proposes a layered reference model for ingesting, governing, and serving executable knowledge to humans, agents, models, and workflows.

Problem

Classical data architecture governs structured records; its guarantees (schema validation, CDC, lineage, catalogs) do not transfer unchanged when the managed asset is heterogeneous, meaning-bearing knowledge artifacts consumed operationally by LLMs, agents, and agentic workflows. There is no systematic architectural framework for making organizational knowledge reliably executable — governed, traceable, fresh, and assembly-ready — rather than passively consultable.

Method

This is a conceptual/taxonomic paper, not an empirical study. The core technique is analogical lifting: each mature data-engineering guarantee is taken as a template, its semantics are re-examined under the unit shift from record to knowledge artifact, and a knowledge-level reformulation is stated. Eleven pairs are tabulated (Table 1): e.g., ETL → knowledge ingestion, CDC → knowledge change detection, lineage → knowledge provenance, materialized view → knowledge view, medallion bronze/silver/gold → raw/curated/operational knowledge layers. A five-concern reference model (Sources → Ingestion → Repository → Governance+Provenance → Serving) is proposed, and emerging formats (LLM Wiki, Google OKF) are analyzed as partial evidence of the shift.

Key Contributions

• Paradigm framing: characterizes knowledge architecture as the evolution of data architecture when the managed unit shifts from data records to heterogeneous knowledge artifacts consumed by agents, models, and workflows.
• Guarantee taxonomy (Table 1): eleven data-engineering guarantees redefined with changed semantics at the knowledge level; clarifies why direct transfer fails (e.g., change detection must catch semantic drift, not merely byte-level change).
• Layered reference model (Figure 3): five-concern architecture (ingestion, repository, governance/provenance, serving, consumers) applicable to enterprise AI systems.
• Medallion analogy (Table 2): concrete pattern elevation — raw/curated/operational knowledge layers generalizing bronze/silver/gold.
• Research agenda: six open questions on representation, synchronization, provenance, quality, governance, and evaluation.
• Boundary clarification: distinguishes knowledge architecture from RAG (a serving technique), knowledge management (managerial/sociotechnical), Semantic Web (representation only), and context architecture (ephemeral assembly).

Results

No empirical results; this is a vision/position paper. The paper does not report benchmark numbers, ablations, or experimental comparisons. Industrial signals cited as qualitative evidence: GitHub Copilot Workspace, Microsoft Copilot Studio (integrating SharePoint, Dataverse, Salesforce, ServiceNow, etc.), and Google Cloud OKF.

Limitations

• No empirical validation of the taxonomy or reference model; all claims are conceptual.
• The guarantee mappings are asserted, not formally derived — equivalence conditions are unstated.
• Does not address implementation cost or feasibility of the full architecture stack.
• OKF and LLM Wiki are treated as "early evidence" but are nascent; the gap between the vision and deployable systems is large and not quantified.
• Semantic drift detection and meaning-level change-data capture — central to the proposal — are acknowledged as open problems without proposed solutions.
• Scope is enterprise-centric; how the framework scales to open or federated knowledge ecosystems is not addressed.

Relevance to Harnesses / Meta-Harnesses

Harnesses and meta-harnesses are orchestration layers that assemble context, invoke tools, and coordinate agents — they are precisely the consumers at the bottom of the reference model this paper proposes, and their reliability depends directly on the knowledge-serving layer upstream. The paper's "knowledge views" and "context bundles" concept directly formalizes what harnesses receive as input, while the governance and provenance layers determine which knowledge artifacts an agent is authorized to act on — a concern central to safe harness design. The medallion raw/curated/operational layering offers a concrete architectural pattern for how a meta-harness might stage knowledge before exposing it to sub-agents. Harness builders tracking this topic should treat this paper as a call to treat the knowledge substrate as first-class infrastructure rather than an afterthought in the RAG retrieval step.