HYBRID RECALL - INCREASED PRODUCTION SYSTEM

Abstract

This invention relates to a hybrid recall-augmented generation (RAG) system (1) that uses vector embeddings and knowledge graphics to significantly improve the access to information and response generation of large language models (LLMs).

Description

1 TARIFF HYBRID RECALL - INCREASED PRODUCTION SYSTEM Technical Area This invention significantly improves the access to information and response generation of large language models (LLMs). a hybrid backend that significantly improves performance by using vector embeddings and infographics It is related to the invoke-augmented production (RAG) system. Previous Tech 10 Large language models (LLMs) are trained on large text datasets. They have acquired impressive natural language comprehension and production abilities. However, this The main disadvantages of these models include the limited timeliness of training data. their ability to produce false or fabricated information, known as "hallucinations" 15 trends and lack of access to information specific to a particular field or institution. They remain. In order to resolve these issues, a Recall-Enhanced Reproduction Ag (RAG) systems have been developed. Traditional RAG systems are a Before responding to the user's query, relevant information is obtained from an external source. It retrieves data from its base (usually a vector database) and presents it as context to the LLM. 20 This helps the LLM program produce more accurate and up-to-date answers. However, known RAG systems have some shortcomings and disadvantages: • Single-Mode Recall Limitations: Most RAG systems only offer single-mode recall. It focuses on vector search based on semantic similarity. This is for text fragments 25 complex relationships or structural information between them (for example, one entity's relationship with another) (connections with entities) can be ignored. This is especially true for multi-step reasoning. Inadequate in queries requiring execution or in-depth contextual knowledge. It can stay. 2 • Lack of Context: Recalling only text snippets is insufficient for an LLM. This can limit the richness of the context presented. Not explicitly stated in the text. However, information that can only be understood through the relationships between entities may be lost. • Sequencing and Fusion Challenges: From different recall methods (e.g., keyword and vector) effectively 5 the obtained results Combining and sorting is a complex problem. The relevance of the retrieved information... Accurately assessing the level and importance of an LLM improves its performance. It has direct effects. • Risk of Hallucinations: When the quality of the recalled information is poor or When insufficient, LLM still carries the risk of producing hallucinations. 10 These shortcomings make current RAG systems incompatible with complex information environments or high-level datasets. reaching their full potential in accuracy-requiring applications This is preventing it. Therefore, we need to eliminate these disadvantages and create a more robust, accurate, and A new system is needed that will produce contextually rich responses. 15 Chinese patent no. CN121233792, which falls under the known state of the art. The document describes an artificial intelligence knowledge based on a large model and RAG technology. This is mentioned in the method of creating the base. Brief Description of the Invention The aim of this invention is to improve the information access and response of large language models (LLMs). using vector embeddings and infographics that significantly improve its production The goal is to implement a hybrid recall-augmented production (RAG) system. 25 Detailed Description of the Invention The “Hybrid Recall-Enhanced” process was implemented to achieve the purpose of this invention. The "Production System" is shown in the attached figure; this figure is 30. 3 Figure 1. Schematic view of the system described in the invention. The parts shown in the figure are individually numbered, and the corresponding numbers correspond to these numbers. given below: 1. System 2. Electronic device 3. Server D. External server 10 Large language models significantly improve information access and response generation. The invention is a hybrid system that improves upon, utilizes vector embeddings and infographics. recall-enhanced production (RAG) system (1); - users can ask questions and receive answers to their questions in order to access information. 15 at least one electronic device (2) that provides -to find the answer to the question asked via electronic device (2) Access to documents that may be included from an external server (D) and this extracting text-based data from documents, and breaking this data down into smaller parts. the division into manageable parts, entity models of fragmented texts 20 extracting key elements from texts and analyzing the relationships between elements creating infographics by identifying relationships, as well as fragmented High-resolution images of each text fragment are created using vector embedding models of the text. converting a digital representation in a three-dimensional vector space into an information graphic. 25 of the data and vector embedding data are recorded in the owned database. when receiving a query via an electronic device (2), this query and data hybrid recall and fusion score of data recorded in the database and its transmission to the enhancement module, in this module both vector for semantic similarity. from the database as well as from the knowledge graph database for structural and relational context. Simultaneous searching and the resulting sequence fusion 30 By using an engine to combine and sort them, thus providing both semantic and analytical results for queries. 4 at least the most structurally relevant response is provided. It includes a server (3). Server (3) located in the system (1) which is the subject of the invention, legal documents, medical records, Text-based data to be processed in the form of academic articles will be received from external servers. 5 (D) is configured to receive it. Server (3) receives (D) from the external server. semantic integrity of the received text, PDF or web page documents smaller, smaller units that the LLM can process and recall in a way that is protected, It is structured to allow for its breakdown into manageable chunks. Server (3) uses 10 documents to reveal the structural information within the text. important elements in the text fragments such as people, places, organizations, and concepts assets and the relationships between these assets, the classification of assets, and Entity models are pre-trained models used for processing. It is configured to extract using. The server (3) extracts each of the fragmented parts. The text fragment is embedded in BERT or Word2Vec-based AI models. a numerical representation in a high-dimensional vector space using models transforming the semantic content of text fragments with vector embeddings by capturing semantically similar parts close together in vector space It is configured to enable its positioning. The server (3), extracted entities and relationships, information through nodes (entities) and edges (relationships) 20 by transforming the text into infographics that represent hidden information. to clearly reveal complex connections and contextual information that may remain then storing the infographics in the infographic database. It is configured to provide the generated vector embeddings. The server (3) It has 25 optimized features for fast and efficient semantic searching. It is configured to allow storage in a vector database. Server (3) stores documents with vector representations of user queries. by calculating the similarity between the vector representations of the particles, the most relevant to use the vector database to retrieve the parts is configured. The server (3) processes both the query from the user and vector data 30 to enable simultaneous searching in both the database and the graphical database It is structured in such a way. The server (3) obtains from both callback mechanisms. The results obtained were from the "Reciprocal Rank Fusion Engine". It is configured to combine and sort using. The server (3), this Reciprocal Rank Fusion assesses the relevance of different recall outcomes with the engine. By combining them using the Mutual Ranking Fusion (RRF) algorithm, the final 5 It is configured to generate a ranking score. Server (3), fused and augmented context as input to a grand language model the provision of this enriched context by the LLM, allowing the user to... to produce a more accurate, consistent and contextually relevant answer to the question, and 10 It is structured to provide this. Industrial Application of the Invention The system of invention (1) provides access to information and response of large language models (LLM) 15 An innovative hybrid recall designed to improve its production. It is an augmented growth (RAG) system. Existing RAG systems generally rely on a single return. call mechanism (e.g., vector-based only or key-based only) While focusing on word-based systems, the system in question (1) has both semantic similarity. 20 It aims to retrieve more comprehensive and accurate information. The system (1), splitting documents into parts, creating vector embeds, and embedding assets. It constructs an infographic by extracting vector data. User queries are processed using both vector data. It is searched using a hybrid approach in both the database and the graphical database, and The results are combined using a "Reciprocal Sequence Fusion Engine" and 25 This fusion enhances the richness and relevance of the context offered to the LLM. By increasing its level, it reduces hallucinations and makes them more accurate and contextual. It enables the production of rich responses. The invention is particularly relevant for complex and extensive information. in areas where they have foundations, the existing systems have a lack of information and contextual By eliminating disadvantages such as limitations, it offers significant advantages. 30 6 Around these basic concepts, the invention is called “Appropriate Position Estimation System (1)” It is possible to develop a wide variety of related applications, and the invention described herein It cannot be limited to examples; it is essentially as stated in the requests.

Claims

7 REQUESTS 1. Large language models significantly improve information access and response generation. improving, using vector embeddings and infographics; - Users can ask questions to access information and receive answers to their questions. 5 at least one electronic device that enables it to receive (2), -to find the answer to the question asked via electronic device (2) Access to documents that may be included from an external server (D) and this extracting text-based data from documents, and breaking this data down into smaller parts. breaking down manageable fragments, entity models of fragmented texts 10 extracting important elements from texts and using assets Identifying the relationships between them and creating infographics, as well as Each text is rendered using vector embedding models of the fragmented texts. a numerical representation of the part in a high-dimensional vector space the transformation of infographic data and vector embedding data 15 keeping records in the database owned, electronic device (2) When a query is received through this system, this query and the information recorded in the database data together with hybrid recall and fusion score & enhancement module This module transmits information from a vector database for semantic similarity. also from the knowledge graph database for structural and relational context. the search should be performed in a timely manner and the results obtained should be arranged in a sequential order. By using a fusion engine to combine and sort the data, thus providing queries in both directions. To ensure that the most relevant answer is given, both semantically and structurally. a hybrid backup characterized by having at least one server (3) configured to do so Call-in augmented production (RAG) system (1). 25 2. Legal documents, medical records, and academic articles will be processed. to enable the retrieval of text-based data from external servers (D) A system like the one in Claim 1 characterized by the configured server (3). (1). 30 8 3. Text, PDF or web page received from external server (D) the LLM can process the documents in a way that preserves their semantic integrity and broken down into smaller, manageable chunks that it can recall Claim 1, characterized by the server (3) configured to provide or A system like the one in 2 (1). 5 4. Text in documents to reveal structural information within the text. important parts in the form of people, places, organizations, and concepts assets and the relationships between these assets, the classification of assets, and Entity models are pre-trained models used for processing. 10 characterized by the server (3) configured to extract using a system like any of the above requests (1).

5. Analyze each fragmented text piece using a BERT or Word2Vec-based artificial intelligence system. High-dimensional vector 15 using embedded models which are intelligence models converting text to a digital representation in the space, using vector embeddings by capturing the semantic content of their parts, they are semantically similar. To ensure that the parts are positioned close together in the vector space. from the above requests characterized by the server (3) configured for a system like any other (1). 20 6. Extracted entities and relationships, nodes (entities) and edges (relationships) transforming information into infographics that represent information through by revealing complex connections and contextual elements that might otherwise remain hidden within the text. clearly present the information, then the information contained in infographics 25 server configured to store the graph in the database (3) like any of the above-mentioned claims characterized by system (1).

7. The generated vector embeddings enable fast and efficient semantic search. 30 in order to do this, it has an optimized vector database 9 characterized by the server (3) configured to enable storage. a system like any of the above-mentioned requests (1).

8. Vector representations of user queries and stored document snippets. By calculating the similarity between vector representations, it retrieves the most relevant parts. 5 server configured to use vector database for calling (3) like any of the above-mentioned claims characterized by system (1).

9. The query from the user is processed in both vector and graphical databases. 10 structured to enable simultaneous searching at the base level in any of the above requests characterized by the server (3) such a system (1).

10. The results obtained from both recall mechanisms are compiled under "Mutual Sequence 15". Combining using a "Reciprocal Rank Fusion Engine" and characterized by the server (3) configured to sort a system like any of the above requests (1).

11. Relevance of different recall outcomes with this engine Reciprocal 20 Using the Rank Fusion (RRF) algorithm structured to bring together and create a final ranking score in any of the above requests characterized by the server (3) such a system (1).

12. Fusion and augmented context as input to a grand language model. the provision of this enriched context by the LLM, allowing the user to... a more accurate, consistent and contextually relevant answer to the question to produce and retrieve the most relevant information, both semantically and structurally. 30 with server (3) configured to enable presentation to LLM. a system like any of the above characterized claims (1). 10 20 30

Images