SPADEs Semantic Data Planner (Zalo AIC)
LLM Agent for Vietnamese Business Question Answering & Executable Query Grounding
An LLM-driven semantic planning and Text-to-SQL/structured query pipeline built for the Zalo AI Challenge Datamate track, translating Vietnamese business analytics questions into schema-grounded, executable relational queries.
The Technical Bottleneck
Translating ambiguous Vietnamese business questions into executable multi-table queries requires resolving implicit temporal expressions, schema linking, complex aggregations, and type-safe joins.
Why Existing Baselines Fail
Direct end-to-end Text-to-SQL prompting with small-to-medium open-weight models frequently produces syntax errors or hallucinated column names; decomposing the task into semantic planning and executable validation exposes exactly where reasoning breaks down.
Core Formulation & Insight
Built the SPADEs pipeline combining schema grounding, temporal normalization, join/aggregation planning, and post-generation query execution validation with Qwen3-8B, paired with granular error logging over a 200-query held-out diagnostic set.
Architectural Design
Multi-step semantic planner and SQL/query validator for Vietnamese business analytics.
1. Vietnamese Query Understanding & Temporal Normalization
Parses business metrics, time windows, and entity filters from natural-language Vietnamese queries.
2. Schema Grounding & Join Path Planning
Maps extracted intents onto relational table schemas, foreign keys, and aggregation functions.
3. Executable Validation & Diagnostic Logging
Validates generated plans for syntactic executability, schema coverage, and strict typed output accuracy.
Technical Stack & Systems Engineering
Implementation Highlights
- →Implemented detailed failure-mode diagnostics separating executable plan generation (24%), schema coverage (20%), and strict typed accuracy (1%) on early Qwen3-8B held-out runs.
Datasets & Protocol
Datasets & Corpora
- 200-Query Held-Out Vietnamese Business Analytics Set (Zalo AI Challenge Datamate)
Evaluation Protocol
Evaluated Qwen3-8B semantic planning across executability, coverage, and strict typed accuracy.
Tracked Evaluation Metrics
Diagnostic held-out evaluation on 200 complex business queries quantified early Qwen3-8B bottlenecks (24% executable plan rate, 20% coverage), guiding schema-grounding error analysis.
Vietnamese business analytics benchmark
Early Qwen3-8B diagnostic run (not final competition score)
Empirical Analysis & Interpretation
Transparently reporting early diagnostic bottlenecks demonstrates how strict typed evaluation exposes the gap between plausible SQL syntax and exact business-logic execution.
What This System Contributes
- 01Built the semantic planning, schema grounding, and held-out error-analysis pipeline.
Boundary Conditions & Future Work
- ·Private competition repository; reported metrics reflect early diagnostic logs rather than a final leaderboard submission.