Document Chunk Ordering for Complete LLM Retrieval Prompts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models generate incomplete answers due to the limitation of character input for prompts, as documents are divided into small text fragments that only include partial content, leading to inaccurate retrieval responses.
Innovation Solution
A document processing apparatus that divides documents into chunks and blocks, calculates similarity scores, rearranges blocks to maintain chunk order, and generates prompts to ensure coherent and complete answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a document is divided into small text fragments (document blocks) to fit within prompt character limits, then the prompt can be transmitted within character limits, but the answer becomes incomplete because partial content is lost
Solution Approach 1:
The patent divides a document into multiple chunks, where each chunk contains multiple document blocks. This hierarchical segmentation allows the system to manage document content in smaller units that fit within prompt character limits while preserving the complete semantic context within each chunk. The segmentation principle resolves the contradiction by organizing content at different levels: fine-grained blocks for retrieval and coarser-grained chunks for maintaining completeness.
Solution Approach 2:
The patent implements a nested structure where document blocks are contained within document chunks. Each chunk acts as a container that holds multiple related blocks, creating a nested hierarchy. This nesting allows the system to transmit only relevant chunks (containing selected blocks) within prompt limits while ensuring that each transmitted chunk maintains internal completeness, thus resolving the information loss problem.
2Productivity
If only relevant document blocks are extracted based on similarity scores, then the prompt size is reduced, but the coherence and completeness of the answer deteriorates
Solution Approach 1:
The patent performs preliminary extraction of document blocks that are similar to the query before generating the final answer. By pre-identifying and extracting only the most relevant blocks within each chunk, the system prepares a condensed but coherent set of information that maintains answer reliability while improving retrieval efficiency. This preliminary action ensures that the large language model receives focused, high-quality input.
Solution Approach 2:
The patent introduces document chunks as an intermediary layer between the raw document blocks and the large language model. The chunk structure acts as a mediator that organizes scattered relevant blocks into coherent contextual units. This intermediary structure helps the model understand the relationships between blocks and generates more accurate, coherent answers while still benefiting from selective extraction.
3Adaptability or versatility
If document blocks from the same chunk are scattered in the retrieval result, then the prompt can include more diverse content, but the understanding and coherence of the answer decreases
Solution Approach 1:
The patent merges document blocks that belong to the same chunk into a unified contextual unit. By combining related blocks within each chunk before transmission to the large language model, the system preserves the contextual coherence and logical flow of the original document. This merging ensures that scattered blocks are reassembled into meaningful sequences, maintaining answer quality while allowing diverse chunk selection.
Data Source
AI summary
A non-transitory computer-readable storage medium storing a program for causing a computer to divide a document to be retrieved into individual content chunks to generate document chunks, divide each of the document chunks into document blocks having a predetermined number or less of characters, extracts the document blocks for which a similarity score is high from among the document blocks, generates a retrieval result, and rearrange the order of the document blocks included in the retrieval result to make the order of the document blocks that belong to an identical document chunk consecutive, and make the order of the document blocks that belong to the identical document chunk match the order in the document chunk.