Mirae Asset Securities Improves AI Search Average Score by 84% with Product Knowledge Ontology Based on Amazon Bedrock

·2026.11.20 12:47

Key point

Leveraging Amazon Bedrock and Neptune, the average score for product searches targeted at PBs improved by 84%, from 0.450 to 0.828.

1 / 13

Details

  • Built a product knowledge ontology and AI search service combining Amazon Bedrock, Neptune, and OpenSearch
  • Introduced a P-E-W (Planner-Executor-Writer) multi-agent structure, improving the average score by 84% from 0.450 to 0.828
  • Reduced first answer token generation time by 25%, from 28.68 seconds to 21.41 seconds
  • Launched as an Innovative Financial Service in August 2026, processing over 2,000 complex queries for more than 450 PBs

Architecture and Data Processing

  • Configured multi-account and on-premises-AWS private networks using AWS Control Tower and Transit Gateway
  • Collected DART, EDGAR, and PDF documents, then performed VLM-based parsing and RDF/LPG conversion
  • Achieved a 98.9% pass rate against human review criteria for chart summaries when applying GPT-5.6 Sol

Ontology and Semantic Layer

  • Defined ontology based on CQ (Competency Question) and mapped classes/properties/relations based on TTL
  • Converted finalized facts from RDB to RDF, and research opinions to LPG with validity periods for loading
  • Improved query interpretation accuracy with a semantic layer defining business rules, property dictionaries, and aliases

Agent Design and Performance

  • Separated into a P-E-W structure to resolve issues with the initial single agent (such as premature judgment and context loss)
  • Provided guidelines for re-querying when evidence is insufficient through Plan verification via Agentic Loop
  • Conducted 5 measurements using an internal evaluation system based on 250 questions across 6 scenarios

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.