Ontology-Driven Context Engineering: Turning 10K Objects into Agent-Ready Knowledge with 98.6% Recall
AsiaInfo's telecom case study shows a four-stage ontology-driven pipeline — task understanding, knowledge location, relationship completion, and layered workset delivery — achieves 98.6% object recall, 4.2s retrieval, and 83.5% SQL first-pass rate on 150 real queries, proving task-oriented knowledge organization as a core agent capability.
Enterprise AI agents are moving from concept to production, but adoption lags: Gartner predicts over 40% of agentic AI projects will be canceled by 2027 [1], and McKinsey finds only 23% of companies have scaled agents in any function [2]. The bottleneck is not model reasoning but supplying task-relevant knowledge in an agent-usable form. This has sparked interest in context engineering — defined by Anthropic as finding the minimal high-signal token set within a limited attention budget [3], systematized into retrieval, processing, and management components [4], and shown by the ACE framework to self-improve via execution feedback for a 10.6% gain [5]. Interface standards are converging with Anthropic's Model Context Protocol (MCP) [6] and Google's Agent2Agent (A2A) protocol under Linux Foundation governance [7].
The telecom sector echoes this need. TM Forum's AI-Native Open Digital Architecture (ODA) states that general LLMs are insufficient; agent interoperability requires ontology-driven semantics built on SID information models and eTOM process models [8]. Telstra similarly emphasizes modeling relationships between data, not just data itself [9].
Based on AsiaInfo's practice in telecom data development, where the ontology reaches 10,000+ objects and hundreds of thousands of attributes, the challenge shifts from storage to supply. The core thesis: "organizing knowledge for tasks" must become an independent system capability for enterprise agents.
Three Constraints on Knowledge Supply
Ontology provides business understanding: objects (what exists), attributes (characteristics), relationships (connections). But a map ≠ task completion. A typical data development request involves source models, extended attributes, filter conditions, and join constraints. The agent needs not "what related objects exist" but "which objects jointly support this task and how they connect." At scale, three competing goals emerge:
Completeness — missing any business object can change the result.
Conciseness — flooding the model with similarly named objects increases disambiguation burden; more results ≠ more reliable.
Interactivity — if every query repeats heavy preparation, knowledge capability cannot enter continuous development flows.
Therefore, the evaluation unit must shift from single objects to knowledge worksets — containing necessary objects, attributes, and connection rationales for downstream modeling, SQL generation, or analysis.
Four-Stage Task-Oriented Knowledge Organization
The practice abstracts into four stages (Figure 1):
3.1 Understand Task: Configure Processing by Uncertainty
Users describe goals in business language, rarely listing full source objects and join conditions. The starting point is identifying information roles in the request: existing sources, target artifacts, and semantic filter conditions needing interpretation (e.g., "extend broadband attributes based on existing model and exclude virtual numbers" combines base source, attribute extension, and exclusion rule). Work is divided by uncertainty:
Complex language understanding goes to the LLM; exact table/column names use precise matching; object connections use structured queries — to control latency and error propagation.
3.2 Locate Knowledge: Objects and Attributes as Joint Entry Points
Business expressions often diverge from asset naming. A user may know the "customer billing" object, or only need "uplink speed" without knowing its model. Object-only retrieval misses the latter. Hence both objects and attributes serve as retrieval entries, combining semantic understanding, lexical matching, and identifier lookup. The criterion is requirement coverage : every key business condition must find a knowledge source, and unsupported requirements must be flagged.
3.3 Complete Relationships: Importance Determined by Task Role
After locating objects, we must judge if they jointly support the task. Some objects supply business fields; others play bridging roles — not mentioned in the query but critical for join paths. Graph relationships provide structured evidence: find necessary connection paths around requirement entry points, add bridging objects, and limit irrelevant expansion. Every added object must have a clear purpose. Unlike Microsoft GraphRAG which targets global inductive questions over a corpus [10], this stage aims to "make this task work."
3.4 Deliver Workset: Layered Organization with Provenance
Core objects, necessary bridges, and backup candidates play different roles and should be presented in layers. For each object: which requirement it satisfies, attribute provenance, and why a relationship was chosen — all delivered with the result so generation is traceable and developers can audit omissions or correct misunderstandings.
Case Study: The Unnamed Bridge Object
In a product instance modeling task, developers needed to reuse a base model, extend primary/secondary card info and broadband attributes, and filter out virtual numbers. Four knowledge needs: base source, card attributes, broadband attributes, virtual-number filter.
The system located objects for each need, then checked graph connectivity. A "product service instance" object — never mentioned by the user — entered the core workset because it bridged the base object to the virtual-number object, providing the join path for the filter. Final core workset: 4 directly recommended objects + 1 bridge object , covering all four needs; similar objects returned as backups.
Knowledge importance must be judged by its role in the task. Returning only name-similar objects looks rich but lacks the evidence needed for filtering; relationship completion organizes scattered knowledge into a coherent development context, letting developers focus on reviewing sources and transformation logic.
Evaluation: From Retrieval Hit to Task Usability
Test set: 150 real questions across three scenarios (object lookup, attribute reverse lookup, complex analysis — 50 each) on the 10K-object, hundreds-of-thousands-attribute ontology. Complex analysis includes multiple transformation constraints, further testing table-creation and data-processing SQL generation. Results (Figure 3):
Knowledge location quality: Object recall accuracy 98.6% (150 cases, project aggregate).
Online retrieval efficiency: Average retrieval latency 4.2 seconds (retrieval stage only).
Downstream artifact usability: Table-creation and data-processing SQL first-pass rate 83.5% (50 complex analysis cases, project test verdict).
Requirement processing efficiency: Average development cycle 59.3 seconds (50 complex analysis cases, project test process; not indicative of production delivery cycle).
Enterprise data task difficulty is quantified: Spider 2.0 benchmark contains 632 real enterprise text-to-SQL tasks with schemas averaging 812 columns; best LLM achieves only ~17% task success [11] — the difficulty lies not in language understanding but in missing schema, dialect, and join-path knowledge. This is why knowledge supply is treated as an independent capability, not a retrieval submodule. First-pass rate below 100% confirms generation validation and human review remain necessary.
Four evaluation dimensions are recommended (Table 2). Metric interpretation must consider task requirements: fewer objects may mean more precise organization or omission; faster response may come from sensible division of labor or skipping necessary steps. Retrieval performance cannot be equated directly with business correctness.
Engineering Trade-offs: Online "Lightness" at the Cost of Offline Maintenance
Second-level knowledge supply requires upfront investment. In this practice, objects, attributes, and relationships are pre-organized into queryable indexes and graph structures, synchronized with ontology changes; online requests simply query around the current need. Engineering effort shifts to index maintenance, relationship governance, and evaluation set construction. Reducing per-task usage cost requires continuously bearing knowledge asset maintenance cost.
Three boundaries for production:
Graph paths only prove connectivity; granularity, cardinality, and time scope may affect results — further business validation needed.
Knowledge organization is constrained by ontology quality: missing relationships, ambiguous definitions, stale attributes propagate into task context. Parts lacking evidence should surface unknowns, trigger clarification, or fall back to humans.
Retrieval and permissions are governed separately: findable ≠ authorized to access data or execute operations. Version consistency is equally critical: when knowledge structure changes, the system must confirm which version a task used for reproducibility and traceability.
Judgment & Outlook: Ontology Evolving from Project Deliverable to Agent Production Line
Industry convergence: Palantir positions ontology as the "operating layer of the organization," a core architectural differentiator for its AI platform [12]; ACE shows task experience can distill into evolving contexts that feed back into future tasks [5]; TM Forum embeds ontology-driven semantics into the telecom AI-native architecture foundation [8]. Ontology is shifting from a one-off data governance deliverable to a continuous production line feeding task capabilities to agents.
Three judgments:
The value of large-scale ontology will increasingly lie in whether it can timely supply complete, concise, evidence-backed knowledge around concrete tasks , not in the completeness of the asset catalog itself.
The context supply capability determined by knowledge organization will become a key metric for enterprise agent maturity — it directly decides whether agents enter business processes as demos or as production systems.
Knowledge evolution must maintain governance boundaries : human candidate replacement, relationship confirmation, and failure correction must be reviewed before feeding back into knowledge maintenance and evaluation; model-proposed associations must be verified before entering formal knowledge assets.
Let knowledge enter the task, let the task validate knowledge — a practicable path for turning ontology assets into agent capabilities.
Code example
[1] GARTNER. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027[EB/OL]. Gartner Newsroom, 2025[2026-10-19]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027.
[2] MCKINSEY & COMPANY. The state of AI in 2025: Agents, innovation, and transformation[R/OL]. McKinsey Global Survey, 2025[2026-10-19]. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai.
[3] ANTHROPIC. Effective context engineering for AI agents[EB/OL]. Anthropic Engineering, 2025[2026-10-19]. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents.
[4] MEI L, YAO J, GE Y, et al. A survey of context engineering for large language models[EB/OL]. arXiv: 2507.13334, 2025[2026-10-19]. https://arxiv.org/abs/2507.13334.
[5] ZHANG Q, HU C, UPASANI S, et al. Agentic context engineering: Evolving contexts for self-improving language models[EB/OL]. arXiv: 2510.04618, 2025[2026-10-19]. https://arxiv.org/abs/2510.04618.
[6] ANTHROPIC. Introducing the Model Context Protocol[EB/OL]. Anthropic News, 2024[2026-10-19]. https://www.anthropic.com/news/model-context-protocol.
[7] THE LINUX FOUNDATION. Linux Foundation launches the Agent2Agent Protocol project to enable secure, intelligent communication between AI agents[EB/OL]. 2025[2026-10-19]. https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents.
[8] TILLER A. AI Native ODA: The path to open digital autonomy[EB/OL]. TM Forum Inform, 2026[2026-10-19]. https://inform.tmforum.org/features-and-opinion/ai-native-oda-the-path-to-open-digital-autonomy.
[9] BUSHAUS D. Telstra's autonomous network vision is built on ODA[EB/OL]. TM Forum Inform, 2025[2026-10-19]. https://inform.tmforum.org/features-and-opinion/telstras-autonomous-network-vision-is-built-on-oda.
[10] EDGE D, TRINH H, CHENG N, et al. From local to global: A graph RAG approach to query-focused summarization[EB/OL]. arXiv: 2404.16130, 2024[2026-10-19]. https://arxiv.org/abs/2404.16130.Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
