Information Processing Apparatus for Accurate LLM Answers

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Large language models often produce inaccurate answers when faced with complex or vague user questions, as they lack the precision needed to disassemble and process such queries effectively.

Innovation Solution

An information processing apparatus that decomposes user questions into multiple specific partial questions, selects relevant databases, and generates prompts for a language model to ensure accurate processing by embedding relevant data as context, allowing for sequential execution and refinement of answers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a large language model directly processes complex or vague user questions, then the system maintains simplicity in structure, but the accuracy of the output answer deteriorates

Engineering Contradiction:
Improvesystem structureVSAvoidanswer accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies segmentation by dividing complex user questions into multiple partial questions through a question decomposer. This allows the system to process each partial question separately with appropriate database selection, improving answer accuracy while maintaining overall system simplicity. The decomposition transforms one complex processing task into multiple manageable sub-tasks.

Inventive Principle:
Principle #1Segmentation

2Device complexity

If the system uses a single database for all questions, then the device complexity is reduced, but the accuracy of answers to specific questions deteriorates

Engineering Contradiction:
Improvedatabase structureVSAvoidanswer accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent implements local quality by selecting different databases based on the specific type of partial question being processed. Each database is optimized for particular data types (e.g., sensor data database for sensor-related queries, knowledge base for general knowledge), ensuring that the most appropriate data source is used for each specific query context, thereby improving answer accuracy.

Inventive Principle:
Principle #3Local quality

3Productivity

If the system processes complete user questions as single prompts, then the processing speed is maintained, but the reliability of the output deteriorates for complex questions

Engineering Contradiction:
Improveprocessing speedVSAvoidoutput accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary action by performing question decomposition and database selection before generating the final prompt to the language model. This preparatory processing organizes the input question into structured partial questions with associated data, enabling the language model to process more reliable and focused prompts, thereby improving output reliability without significantly impacting overall processing speed.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4528537A1Information processing apparatus, information processing method, and computer program
Publication Date: 2025.03.26 KK TOSHIBA
  • EP4528537A1 patent drawingFigure 1
  • EP4528537A1 patent drawingFigure 2
  • EP4528537A1 patent drawingFigure 3

AI summary

According to one approach, an information processing apparatus comprising a processor configured to: select, based on a content of an input question, a database corresponding to the question among a plurality of databases respectively storing different kinds of data; generate a prompt to be input to a language model based on the selected database and the question; and generate an answer based on the prompt and the language model.