1. Catastrophic Forgetting and the Limitations of RAG

When AI is introduced into real-world tasks, one of the biggest concerns is the phenomenon of catastrophic forgetting that occurs when fine-tuning with new data on top of previously learned knowledge.

This refers to the phenomenon where neural network-based AI rapidly loses previously acquired knowledge or patterns during the process of learning new information.

For example, if you fine-tune an open-source LLM with specific corporate data, its general common sense or language abilities may deteriorate.

To avoid such issues, RAG (Retrieval-Augmented Generation) technology is widely used, but RAG also has its limitations.

Representative Limitations of RAG

Poor Handling of Structured Data (Structured Data QA): RAG systems are mainly optimized for unstructured text documents and cannot properly understand or utilize the meaning and relationships of structured data (tables, databases, spreadsheets, etc.).
Limitations with Complex PDFs/Unstructured Documents (Complex PDFs): With complex PDF documents (including tables, multi-columns, images, etc.), information may be lost or context distorted during the chunking process. As a result, important data may not be indexed or may be difficult to search.
Lack of Fallback Model(s): When a RAG system cannot find an appropriate answer, it lacks or has inefficient logic to switch to an alternative (fallback) model. This means that when an answer fails, users are not provided with satisfactory alternatives.
Security Vulnerabilities (LLM Security): There is insufficient defense against security vulnerabilities inherent to LLMs (prompt injection, data leakage, etc.), which can lead to exposure of sensitive information.
Lack of Scalability (Data Ingestion Scalability): Scalability issues arise during the indexing and storage of large volumes of data. As data increases, chunking, storage, and search speeds decrease, and system load increases.
Missing Important Information (Missing Content): Important content in documents is often omitted or not indexed during the chunking process. As a result, users cannot search for the information they want.
Missed Top Ranked: The most relevant chunk (top ranked) may be omitted from search results. Causes include limitations of the search algorithm, degraded embedding quality, and ranking errors.
Context Mismatch (Not in Context): The retrieved chunk often does not match the actual context of the question. This is due to the limitations of simple similarity-based search, resulting in a lack of semantic connectivity.
Wrong Format: The format of the retrieved chunk may be unsuitable for LLM processing or may differ from the answer format desired by the user. For example, tables may be converted to text, distorting the information.
Failure to Extract Information (Not Extracted): The required information is not properly extracted from the chunk, or the LLM fails to recognize the information. This often occurs with complex sentence structures, tables, images, etc.
Inappropriate Information Scope/Depth (Incorrect Specificity): The answer may be too broad or, conversely, too specific for the user's actual needs. It is difficult to control the scope and depth of information.
Incomplete Answers (Incomplete): The final generated answer is incomplete or only partially provided, failing to fully meet the user's needs. This is caused by not synthesizing information from multiple chunks or the LLM only utilizing part of the information.

As such, RAG is overly dependent on simple vector similarity search and chunk-based indexing, so its limitations in reliability, scalability, and accuracy are clear in real-world tasks.

2. Clevi's Semantic Database Overcoming the Limitations of RAG

To overcome these limitations, Clevi has developed a next-generation AI knowledge infrastructure, the Clevi Semantic Database, optimized for semantic connection, complex reasoning, and evidence-based responses.

This solution is possible because we possess our own reasoning model, multimodal AI, and agent-type AI models, and it is one of Clevi's powerful technologies that enables AI to be truly helpful in reality.

By analogy, if a database storing information is a library, Clevi's semantic database is like having an excellent librarian. (If you request information from the librarian, you can get accurate and fast responses.)

Users can add new information to the library at any time and retrieve the information they need.

In addition, you can set version control and access permissions for the data as desired.

All actions of the librarian are recorded in logs, so even if a mistake occurs, it can be corrected.

Main Features and Technical Structure

Semantic Data Storage
- Not just a simple chunk vector DB, but stores semantic relationships (context, hierarchy, causality, etc.) between data in a graph DB
- Easy to expand and supplement in the form of a knowledge graph/semantic network
Hybrid Search and Reasoning
- CTA+Context Query: Reflects both the context of the query and CTA
- Hybrid Retrieval: Combines vector similarity and rule/graph-based search
- Intelligent Folder Hits: Automatically explores semantically related folders/categories
Simultaneous Multimodal Processing
- TextReader Pool, VisionReader Pool, etc. can simultaneously interpret various data such as text, images, tables, and audio
- Whereas traditional RAG can mainly process text, the semantic database excels at multimodal processing
Evidence-Based Responses and Reliability Verification
- Automatically generates document summaries, citations, table findings, etc.
- Merge & Verify: Semantically integrates information from multiple sources and verifies reliability
- Evidence Pack: Provides evidence-based response packages
Complex Reasoning and Planner
- Planner/DAG: Step-by-step planning and DAG (Directed Acyclic Graph)-based reasoning for complex queries
Integrated Agent Structure
- cip-5-agent Supervisor: The top-level agent that manages and coordinates the entire search/reasoning/integration process

3. Structure Summary and Comparison

In summary, the Semantic DB receives data in various forms (audio, text, images, video, etc.) and classifies knowledge by connecting not just simple chunks but also semantic relationships (context, hierarchy, causality, etc.) between data.

Based on this, instead of keyword/similarity-based search like RAG, it performs search and reasoning using semantic connections, enabling more accurate information retrieval and answers from AI.

Also, since it is stored in the form of a knowledge graph/semantic network, it is easy to continuously expand and supplement.

We at Clevi have discovered the limitations of RAG systems in this era of AI transformation and, through our semantic database and proprietary AI models, can now provide customers with more accurate and reliable data quickly.

Clevi's AI system can make your valuable data meaningful regardless of the domain or environment.

We will continue to do our best to maximize the value of your data and provide innovative AI experiences.

Experience the new future of AI with Clevi.