Li Auto: Production Agent Bottleneck Is Context Quality, Not Model Capability

Core Thesis: Context Quality Over Model Capability

Li Auto's senior algorithm engineer Liang Wei and big data platform lead Qian Han independently converge on the same insight: the bottleneck for production-grade AI Agents is not model capability but context quality. Whether an Agent receives fragmented information scattered across prompts, tools, and logs versus structured, semantically governed, traceable context determines if it can deliver stably in production.

Liang Wei's Retrospective: NL2SQL Post-Training Limits

Liang Wei recounts that post-training can push natural-language-to-SQL accuracy to near 100% within a single business domain. However, this comes at the cost of long knowledge preparation and training cycles. Extending to a new domain requires starting over, and any change in business definitions immediately brings maintenance costs back. Accurate query answering does not equate to an Agent that can actually perform work.

Qian Han's Perspective: Context as the Real Bottleneck

Qian Han frames the same problem from the context supply side: the key is not generation ability but whether the Agent gets fragmented pieces or a structured, understandable, traceable context. Both point to the same conclusion — what's missing is not a stronger model but a layer that Agents can understand, govern, and trace.

Ontology as the Unified Entry Point

The role of ontology shifts: it is no longer just a modeling artifact but becomes the single entry point for Agents to consume enterprise data. Objects, relationships, metrics, and business rules are organized into a semantic structure that Agents can query and reason over.

Session 1 (Qian Han): Semantic Layer as Agent Context — Four-Layer Architecture

Using Spark/Flink fault diagnosis as a concrete scenario, Qian Han describes a four-layer architecture that replaces endless prompt engineering:

Ontology : Defines objects, relationships, and boundaries so the Agent knows "what exists in this world."

OSI (Query Semantics) : Translates "what data is needed" into executable, constrained query intent.

Playbook : Encodes expert experience as reusable handling procedures.

General Execution Layer : Performs cross-source evidence gathering across logs, metrics, and data warehouses.

With this stack, Agent behavior shifts from "guessing answers" to "forensics" — pulling evidence and outputting root-cause judgments with evidence, confidence scores, and answer boundaries. Traceability is what separates a toy from a production system.

Knowledge gaps exposed during diagnosis are fed back as versionable, verifiable, regressible semantic assets. Context becomes a governable, rollbackable engineering object rather than an ever-growing prompt. Trade-offs discussed include lightweight modeling, query constraints, and balancing model exploration against engineering determinism.

Unsolved challenges: ontology, metric relationships, and Playbooks still require heavy manual curation with limited cross-business replication efficiency. For larger-scale deployment, permission control, execution tracing, evidence status, result validation, and multi-runtime adaptation remain open. Next steps aim at semantic asset standardization, auto-generation, and a governable, observable, scalable Agent runtime control plane.

Session 2 (Liang Wei): Ontology-Driven Agent from Query to Sales Decisions — Three Architecture Iterations

Liang Wei details an ontology structure designed for Agents: objects, attributes, relationships, metrics, business rules, plus a semantic query mechanism. The core judgment: the gap between business semantics and underlying data structures must be bridged by this middle layer, not left to the model to cross.

Three architectural iterations and their trade-offs:

General MCP : Highly flexible, but results are unstable when business knowledge is insufficient.

Domain Skill : Improves effectiveness, but knowledge tends to scatter across components.

Built-in Agent : Unifies query and analysis processes, but raises the bar for knowledge operations and evaluation.

Each iteration trades one dimension for another; there is no standard answer, only what matches the team's current capability.

In production, three layers sit atop the ontology: Business Fact Layer, Sales Cognition Layer, Scenario Execution Layer. This chain connects ontology to data applications moving from query, to analysis, to business decisions.

Boundaries acknowledged: no off-the-shelf ontology standard exists. They referenced multiple approaches and evolved the ontology based on data analysis needs; it is still evolving, with ongoing trade-offs among expressiveness, development cost, and long-term maintenance. Knowledge governance and continuous optimization remain unsolved.

Complementary Guidance for Different Stages

If your Agent is still in the "inaccurate" phase or you're pushing NL2SQL accuracy higher: study Liang Wei's retrospective on what happens after reaching near 100% accuracy.

If you have a running Agent but results are unstable or unexplainable: examine Qian Han's four-layer split, evidence chains, and knowledge-gap-to-asset pipeline.

If you're debating architecture selection: Liang Wei's three-iteration trade-off checklist covers the two main forks most teams face.

If you're designing an ontology: both sessions warn that there is no copy-paste standard — ontology is traded off, not simply designed.

These two sessions are presented at DACon 2026 Beijing (October 23-24) as complementary upper and lower halves of the same technical roadmap.