Templatized Query Prompts for Faster Natural-Language Query Generation

The patent segments the query generation process into two distinct parts: (1) a template matching stage that handles structured query templates efficiently, and (2) an LLM-based stage that only processes unstructured or complex queries. This segmentation allows most routine queries to be handled by the fast template matching mechanism while reserving LLM resources for more challenging cases, thereby reducing overall processing time while maintaining query generation accuracy.

The patent implements preliminary action by pre-defining query templates with placeholders for common query structures. Before invoking the LLM, the system performs template matching to see if the natural language query fits any predefined patterns. This preliminary step allows the system to quickly resolve many queries without involving the computationally expensive LLM, thus reducing processing time while preserving accuracy for templatable queries.

The patent divides the query processing workload between a low-cost template matching component and a high-cost LLM component. By segmenting queries into templatable and non-templatable categories, the system ensures that only necessary queries incur LLM costs, thereby reducing overall computational expenditure while maintaining understanding accuracy for the subset of queries that require LLM processing.

The patent uses template copying as a cheaper alternative to LLM processing for routine queries. Instead of invoking the expensive LLM for every query, the system copies and fills predefined query templates that have been previously validated for accuracy. This copying approach dramatically reduces computational cost while maintaining query understanding accuracy for standard query patterns.

The patent segments the query generation task by identifying and extracting entities that match predefined templates. For queries involving large subject entities, the system segments the processing into: (1) entity recognition and classification, (2) template selection based on entity type, and (3) template filling. This segmentation allows the system to handle complex queries with multiple entities efficiently by reusing templates rather than generating entire queries from scratch with the LLM, thereby reducing token requirements while maintaining comprehensive query coverage.

The patent implements universality through reusable query templates that can handle multiple entity types and query patterns. A single template can serve multiple purposes by accepting different entity names and attributes as parameters. This multi-functionality allows the system to cover a wide range of queries involving large subject entities without requiring separate LLM processing for each specific query structure, thus reducing token consumption while maintaining adaptability.

The patent segments the data processing pipeline into distinct stages: query execution, raw data retrieval, and formatted presentation. By separating these functions, the system can apply appropriate processing methods at each stage. The template matching component handles structured data retrieval efficiently, while summary and formatting operations are applied only to the retrieved results, not to the entire query processing flow. This segmentation reduces overall processing time while maintaining data presentation quality.

The patent applies preliminary action by pre-defining output templates and formatting structures for common query results. Instead of using LLM to generate formatting from scratch for each query, the system has pre-prepared templates that specify the desired output structure. This preliminary preparation allows for rapid formatting of query results without requiring expensive LLM processing at runtime, thus reducing summarization time while preserving data presentation quality.