Exploration of Large Model Applications—Enterprise Knowledge Steward

1. Background and challenges of traditional knowledge management

1. The necessity of enterprise knowledge management

In modern enterprises, knowledge Management is a crucial link. It can help enterprises effectively organize and utilize internal and external knowledge resources, thereby improving the efficiency and competitiveness of enterprises. In order to better manage knowledge, many companies have introduced the concept of knowledge stewards. Knowledge steward is a role or system specifically responsible for managing and disseminating enterprise knowledge. Through knowledge stewards, enterprises can better collect and organize

##With the rapid development and Knowledge is growing explosively, and companies are faced with the challenge of sharing knowledge. How to effectively transfer and share knowledge within an enterprise has become an important issue. Through knowledge sharing, companies can not only improve work efficiency, but also avoid duplication of work.

Another way is to adopt a knowledge sharing model to establish a mechanism that can empower enterprises, thereby better optimizing processes and results, and improving enterprise operating efficiency. This model allows employees within the enterprise to share their knowledge and experience so that everyone on the team can benefit. By sharing knowledge, companies can avoid duplication of effort, reduce errors and mistakes, and be better able to respond to challenges and changes. This

In addition, as a knowledge steward, it can also provide key information and data to decision-makers to help them make more informed decisions. Knowledge Butler has powerful information retrieval and analysis capabilities, and can extract useful information from massive data, integrate and analyze it. This information and data can include market trends, competitor analysis, consumer insights, technology development, etc.

In addition, a very key factor is to reduce the workload of corporate employees and prevent information loss, and improve employee work efficiency and customer service levels, thereby achieving the goals of reducing costs and improving efficiency.

2. Enterprise knowledge management challenges

Before there was a large model, the logic of building a knowledge steward was quite complicated. Usually, we use the concept of knowledge base to build a knowledge base with the help of enterprise knowledge graph or internal data of the enterprise. However, there are many challenges faced during this construction process. First, the construction of a knowledge base requires a lot of manpower and time investment. Collecting, organizing and summarizing knowledge and information within an enterprise is a tedious and time-consuming task. A professional team is needed to process and manage this data and ensure its

• Knowledge fragmentation

Knowledge fragmentation is mainly reflected in two aspects. One aspect is that the enterprise's data is very scattered. For example, the data of the OA system has different departments and different teams. On the other hand, these data are basically provided in unstructured forms, such as Word, PDF, pictures, videos, etc. In the process of building knowledge stewards, how to quickly centralize the fragmented information is the first challenge.

• Information overload

##In the rapid development of enterprise business, they are faced with a large amount of information and data How to establish a screening mechanism in massive amounts of data to ensure the accuracy and timeliness of information is also a major challenge under the ever-increasing situation.

• Data security risks

Enterprises generally do not share their private data with Other institutions or organizations generally pay more attention to the data security of corporate private domain data, so they also need to deal with data security risks.

• Difficulty in knowledge sharing and communication

Different companies have different organizational structures, some Some are more technical, some are more business-oriented, and some are a mixture of technology and business. In the process of communication between business and technology, poor communication is a problem that every enterprise will face in knowledge sharing.

2. Knowledge steward solution

1. What is enterprise knowledge steward

Enterprise knowledge steward is similar to a person’s brain to assist in the storage and understanding of the entire knowledge and create knowledge.

Enterprise knowledge stewards are generally divided into three levels: the first level is the functional and technical needs, mainly responsible for the management of enterprise knowledge, including enterprise Data import, automatic classification and archiving of documents, and other basic functional requirements; the middle layer is the requirement of the application side, including providing some intelligent question and answer, intelligent search, summary generation, auxiliary writing and other functions; the upper layer is the requirement of the business side , including contract review, insurance customer service, and industry report generation.

There are generally three modes of interfaces presented by Knowledge Butler: the first interface is similar to a text box, providing knowledge exploration and analysis; the other is to use API tokens to Intelligent Agents involved in different application scenarios are published as API Tokens to integrate with the enterprise's business system; the third method is intelligent Agent, which explores and analyzes knowledge through conversation mode.

2. Enterprise knowledge steward solution

Enterprise knowledge steward is mainly responsible for enterprise-specific knowledge management and creation, including the following business scenarios:

• Intelligent Q&A

Combined with the company’s own private domain data, through After vectorization, it is stored in a vector database, and uses the question and answer mode to create intelligent question and answer scenarios. Through these scenarios, many more specific business needs can be derived.

• Self-service document analysis

## Do some exploration and analysis through documents, such as To explore the paper, you can ask questions about the content of the paper, and you can also conduct independent analysis of the document, providing segmented preview, contextual retrieval, summary summary and other capabilities of the entire document.

• Customized role scenario

Combined with the private domain data of different roles within the enterprise, Coupled with the prompt word mode, it provides the design of some customized scenarios, such as assisted writing of documents, intelligent meeting minutes, etc.

• Contract review

adopts the human-computer dialogue mode to conduct various audits of the enterprise Review the contract information on some key terms to see if the corresponding information is accurate.

The main functions of the Enterprise Knowledge Butler product include:

• Intelligent Q&A : Combining specific questions and obtaining a source-based answer by retrieving the context.
• Multi-role creative Q&A: Build intelligent application scenarios through prompt words and corporate private domain data.
• #Document analysis: Import the entire document for summary or exploratory analysis.
• Knowledge management: Enterprise data is fully automatically managed through the knowledge manager, and the entire process adopts a very simple model.
• #Agent Build: Development platform, i.e. large model IDE functionality.

Functional architecture of Knowledge Butler:

The bottom is the GPU calculation Power includes two categories, one is reasoning computing power, and the other is fine-tuning computing power. The middle layer is a secure and trustworthy enterprise private domain data memory - DingoDB multi-mode vector database.

The next layer is the functional points of the entire technical layer, including model fine-tuning management, knowledge document management, and intelligent application management.

The top one is for business scenario needs. In intelligent Q&A, you can customize some dialogues of roles, standard QA Q&A, and agents for intelligent applications, document-based auxiliary reading, contract review, and insurance. personal assistant.

##3. Exploration of core technology of knowledge steward

1. Knowledge steward construction process

Next, we will introduce the entire knowledge steward construction process through the intelligent question and answer scenario.

First of all, there needs to be a data source. There may be structured and unstructured data. Generally speaking, the construction of knowledge base is based on unstructured data. Mainly, such as Word, PDF, Excel, as well as enterprise systems, Jira, knowledge management platforms, etc.

These data go through the knowledge processing link and are converted into vectors and stored in the database. You need to load the document first, then give the layout information or structure information of the document, do document vector analysis to generate file blocks, and then call the corresponding Embedding model based on the file blocks to convert them into vectors and store the vectors.

The process of intelligent question and answer interaction: after the user raises a question, first use the intelligent assistant to vectorize the question, and then go to the database to perform semantic retrieval to obtain the context of the article with similar semantics. By combining the context with the prompt words and reasoning through the large model, the answer is finally returned.

The overall process is a process of continuous iteration and feedback optimization. Only in this way can we obtain the exclusive intelligent expert role based on the enterprise's private domain data.

#2. Knowledge steward construction core technology exploration

• Unstructured data processing

Unstructured data ETL processing requires the help of some tools. Knowledge Manager provides some special operators from the technical model. These operators can clean the entire Map, Filter, and Window-based changes, and convert data through the entire ETL Pipeline.

By parsing various files (such as PDF parsers), and then passing through the Hub Operators of different application scenarios corresponding to the middle layer, the Pipeline Hub can be quickly constructed, and then After the data is cleaned and converted, it is Embedding and finally stored in the vector database.

• Accuracy and integrity data guarantee-lossless data parsing

To get a good To improve the model debugging effect, it is necessary to ensure accurate and complete data and have good data processing quality.

Constructing a traditional data retrieval is very simple, but the actual knowledge is more complicated. In addition to the information in the text itself, there are also pictures and table data , paragraph information, etc. In this regard, Jiuzhang Yunji DataCanvas provides Layout parsing mode, which can realize the full storage of multi-modal data such as Layout information, tables, and pictures, and comprehensively improves the quality of the data parsing process.

• Strong correlation retrieval-Reranking secondary filtering

After the document is vectorized , after saving to the DingoDB multi-modal vector database, retrieval is performed through Query. The retrieval results will include the results of the retrieval content itself, as well as the correlation results. At this time, it is necessary to perform secondary screening of Reranking on the Chunks recalled by the retrieval.

#During Reranking secondary screening, the Retrieval Chunk and the corresponding Query must be related to each other. The analysis includes finding the closest semantic match, and then re-pushing the retrieval Chunk after secondary screening to the large language model.

• Secure and trusted answer generation-multi-instruction fine-tuning

In order to ensure the security and credibility of the answer generation process, Jiuzhang Yunji DataCanvas is based on the general large speech model, limits the prompt words for the recalled data, and combines the enterprise's private domain data with the large model Fine-tuning vertical knowledge and adding a wind direction control mechanism ensure high accuracy in answer generation.

• Storage and retrieval capabilities-DingoDB multi-mode vector database

DingoDB can provide a variety of The standardized API supports data query through SQL and Python toolkits, and also provides an integrated way to implement structured and unstructured joint queries. For real-time scenarios, DingoDB provides the ability to query in real-time by writing in real-time, and can perform real-time retrieval while importing data.

##DingoDB also provides calculation acceleration capabilities and supports pre- and post-filtering of Meta. , and range search based on similarity. DingoDB also provides multi-copy tools that can perform partial migration and data migration. It also provides diversified operation and maintenance and monitoring tools to reduce operation and maintenance costs. DingoDB can also provide automatic elastic sharding capabilities, which can dynamically balance data to different machines to achieve load balancing on each node.

• Secure and trustworthy exclusive LLM-fine-tuned Pipeline

In enterprise private domain data For general scenarios, fine-tuning is needed to build a large language model exclusive to the enterprise in a certain scenario. The knowledge manager summarizes the pain points in the entire fine-tuning process and provides a tool-based approach in the product. Data on all problems can be obtained by uploading documents. After having the data, fine-tuning can be performed directly on the interface by configuring parameters. At the same time, the product also provides some fine-tuning data indicators to evaluate the results of fine-tuning.

• Quickly build large model applications-Large Model IDE

Traditional large model applications are often complex to build. Knowledge Butler built its own large model IDE based on Jiuzhang Yunji DataCanvas's own FS capabilities, which can provide a wealth of components and tools, and use a concise application construction method to build The template is published as an agent for intelligent applications.

##4. Summary and Outlook

1. Knowledge Summary of the Butler Solution

The technical highlights of Knowledge Butler mainly include the following six aspects: high-precision retrieval, convenient ETL Pipeline, high availability and scalability, security compliance, intelligent data fusion, and rich scenarios .

The core values ​​of Knowledge Butler include: providing the basic capabilities of knowledge management and intelligent inspiration, and providing a safe and trustworthy application private Deployment mode includes all data of the enterprise, enabling knowledge integration and intelligent interaction. As an intelligent base, it provides flexible expansion capabilities and can develop new Agents based on large models on Knowledge Manager.

2. Future Outlook

Knowledge Manager is AIFS based on Jiuzhang Yunji DataCanvas, providing a complete set of GPU computing power and model scheduling from bare metal to above, and realizing model fine-tuning. Pipeline mode. It uses the general language model and the company's private domain data to perform combination and fine-tuning to form the company's own large language model. Based on the scalability of the large language model, combined with the DingoDB multi-modal vector database, it can realize search Q&A, summary generation and other applications in the enterprise, and carry out enterprise knowledge management.

The above is the detailed content of Exploration of Large Model Applications—Enterprise Knowledge Steward. For more information, please follow other related articles on the PHP Chinese website!