CrashDB-QA
A localized, database-schema-based agentic workflow system for querying and analyzing a car-crash database through natural language, comprising modules for question understanding, SQL generation and validation, execution, and result interpretation.
Timeline (last 90 days)
CrashDB-QA is a fully local, database-schema-based agentic workflow designed to analyze crash databases directly through natural language.
CrashDB-QA was evaluated on a frozen snapshot containing 4.4 million Wisconsin crash records.
The research team built a benchmark of 300 questions covering 10 analysis categories and tested 9 local open-source models.
Domain knowledge grounding significantly improved query accuracy.
A code-specialized 7B model achieved the best accuracy-latency balance among the tested models.
Composite analyses (such as between-group or time-period comparisons) were harder to handle than routine retrieval.
CrashDB-QA includes several dedicated components: understanding user questions, generating and validating read-only SQL, executing queries, and interpreting results.
A shared knowledge base anchors query construction in the agency's data dictionary and safety analysis conventions while recording intermediate outputs to support inspection and auditing.
The study only targets data from one state, and its generality remains to be further verified.
All Facts (9)
CrashDB-QA is a fully local, database-schema-based agentic workflow designed to analyze crash databases directly through natural language.
50%UnverifiedCrashDB-QA was evaluated on a frozen snapshot containing 4.4 million Wisconsin crash records.
50%UnverifiedThe research team built a benchmark of 300 questions covering 10 analysis categories and tested 9 local open-source models.
50%UnverifiedDomain knowledge grounding significantly improved query accuracy.
50%UnverifiedA code-specialized 7B model achieved the best accuracy-latency balance among the tested models.
50%UnverifiedComposite analyses (such as between-group or time-period comparisons) were harder to handle than routine retrieval.
50%UnverifiedCrashDB-QA includes several dedicated components: understanding user questions, generating and validating read-only SQL, executing queries, and interpreting results.
50%UnverifiedA shared knowledge base anchors query construction in the agency's data dictionary and safety analysis conventions while recording intermediate outputs to support inspection and auditing.
50%UnverifiedThe study only targets data from one state, and its generality remains to be further verified.
50%