Project Background

1. Current Status and Challenges

As an enterprise-level development tool and solution provider, GrapeCity has established a comprehensive knowledge service system, including:

• Standardized Documentation: Product help documents, API documentation
• Community Resources: GCDN technical community (including help center, topic tutorials, technical discussions, and other sub-sections)
• Search Platform: Existing "GrapeCity Search Center" supports cross-platform content retrieval (official website, blogs, community, videos, etc.)

Although the existing system supports fuzzy search and exact search, the accuracy of keyword queries is insufficient, leading to poor user experience. This is specifically manifested in the following typical scenarios:

• Product Learning: Users have difficulty precisely locating function descriptions in help documents
• Solution Search: Users cannot efficiently find solutions that match their problems
• Problem Deduplication: Technical support personnel find it difficult to quickly confirm whether a problem already has a solution

2. Technical Exploration

With the rapid development of artificial intelligence technology, especially large language model technology, we envisioned a Product Intelligent Assistant that could provide instant and accurate product Q&A services to users like an expert. This vision is based on the following technical research results:

• LLM Advantages: Possess powerful natural language understanding and extensive knowledge coverage capabilities
• RAG Value: By introducing external knowledge sources, effectively supplement LLM knowledge limitations and improve Q&A accuracy

Necessity of Self-developed RAG

During the technology selection process, we evaluated multiple open-source and commercial RAG solutions and found that they could not directly adapt to our data structure and business requirements. Additionally, in order to deeply master cutting-edge technology, accumulate R&D experience, and achieve flexible expansion, we decided to independently develop a RAG system. The main reasons include:

1. Significant Content Structure Differences
   Help documents and forum posts have significantly different formats, making it difficult for universal solutions to adapt.
2. Rich RAG Technical Pathways
   Different enhancement strategies need to be verified one by one to select the most suitable technical route.
3. Frequent Dynamic Updates
   Forum content is continuously updated, requiring research into efficient incremental update mechanisms to ensure knowledge base timeliness.
4. Flexible Product Design Adaptation
   Self-developed solutions can better fit actual scenarios and support future feature evolution.
5. Controllable Performance and Maintenance
   Self-built systems have advantages in deployment efficiency and long-term operation and maintenance costs.

Exploration of QA Pre-generation Technology

In traditional RAG solutions, users' original questions are directly matched and retrieved with text paragraphs in documents. This approach has a fundamental problem: user questions and text paragraphs as answers have structural differences in semantic expression, creating misaligned matching situations.

Specifically:

• User questions are usually concise interrogative sentences (such as "How to set responsive layout?")
• Answer paragraphs in documents are often descriptive statements (such as "Find layout setting options in the property panel, supporting three modes...")
• This semantic gap between interrogative and declarative sentences limits the accuracy of direct matching

During our research on large language models, we discovered their excellent information extraction capabilities, which can convert knowledge points in natural language text into structured forms. Therefore, we proposed a new approach: question-to-question matching — using large language models to generate corresponding "preset questions" for each document section, then performing precise matching between user questions and these preset questions.

• Traditional approach: User question ↔ Answer text (semantic misalignment)
• Improved approach: User question ↔ Preset question (semantic consistency)

Furthermore, we process original documents into standard QA pair forms (one question + one answer), where one document can often be broken down into multiple knowledge points, generating multiple QA pairs. After extensive optimization, we successfully achieved this goal and made it adaptable to various types and sizes of documents.

Additionally, we found that when user questions simultaneously match preset questions and corresponding answers, the retrieval effect is better. This is because answers contain more information dimensions, especially user input information. Relevant optimization details will be elaborated in the subsequent "Technical Principles" chapter.

4. Project Decision

Based on preliminary technical validation results, we formulated a phased implementation plan:

Short-term Goals (3-6 months):

• Build intelligent Q&A systems for four core products: Forguncy, Wyn, SpreadJS, and GcExcel
• Cover core content from help documents and technical communities

Long-term Vision:

• Establish a sustainably evolving product knowledge hub system
• Drive continuous system iteration and optimization through user feedback

Appendix - First Batch of Supported Products

This intelligent assistant initially covers GrapeCity's four core product lines, as follows:

Product	Type	Core Capabilities	Official Website
Forguncy	Low-code Development Platform	Enterprise application rapid construction, visual development	Visit
Wyn Enterprise	Embedded Business Intelligence	Self-service data analysis, multi-dimensional report display	Visit
SpreadJS	Spreadsheet Control	Pure front-end spreadsheet processing, Excel compatibility	Visit
GcExcel	Server-side Spreadsheet Component	High-performance batch document generation, cloud Excel processing	Visit