RAG Knowledge Base Retrieval System

Utilizing external databases or knowledge bases for retrieval, and combining retrieval results with large language models to generate more precise and contextually relevant responses.

The core of RAG technology is combining the language capabilities of generative AI with knowledge retrieval capabilities, enabling models to not only rely on internal training data when answering questions, but also dynamically obtain the latest and more specialized information from external databases and incorporate this information into generated responses.

RAG Flow

High-Precision RAG System

While RAG knowledge base retrieval systems can be quickly implemented using Vector Search and released as a basic version, improving their response accuracy further presents challenges. Response accuracy is crucial for user experience as it directly affects users' trust and satisfaction with system responses. If response accuracy is insufficient, users may doubt the system's answers, reducing their willingness to use it.

According to data from the 2023 OpenAI Developer Conference, RAG systems using only simple vector similarity search (Vector Search) and selecting the correct embedding model can achieve 45% accuracy. Adding HyDE Retrieval, FT Embeddings, and Chunk/Embedding Experiments can achieve 65% response accuracy.

MaiAgent RAG not only includes the RAG technologies mentioned at the OpenAI Developer Conference but also combines various classic NLP algorithms and proprietary retrieval technologies. Through internal datasets compared with OpenAI RAG's response accuracy, both can achieve 95% response accuracy.

MaiAgent RAG Response Accuracy

The MaiAgent platform provides two types of RAG: MaiAgent RAG and OpenAI RAG. Here's a comparison table of various aspects:

	MaiAgent RAG	OpenAI RAG
Model Support	Supports all models👍	Only supports OpenAI models
Environment Support	Supports cloud and on-premises👍	Cloud only, data must be sent to OpenAI
Response Accuracy	Very high👍	Very high👍
Supported File Formats	Supports all common formats👍 doc, docx, xls, xlsx, csv, ppt, pptx, pdf, txt, json, jsonl, md	No support for xlsx, csv No support for jsonl No support for legacy Office files (doc, xlsx, ppt)
Image Support in Documents	Yes (currently experimental)👍	No
Table Support in Documents	Yes (currently experimental)👍	No
Attachment Upload Support	Supported👍	Supported👍
Data Slice Transparency	Visualized👍	Black box
Debug Difficulty	Normal👍	Black box, cannot debug
Top K Adjustment	Enterprise version customization👍	None
Embedding Model Switch	Enterprise version customization👍	None