# EnterpriseEM: Fine-tuned Embeddings for Enterprise Semantic Search

**Authors:** Kamalkumar Rathinasamy, Jayarama Nettar, Amit Kumar, Vishal Manchanda, Arun Vijayakumar, Ayush Kataria, Venkateshprasanna Manjunath, Chidambaram GS, Jaskirat Singh Sodhi, Shoeb Shaikh, Wasim Akhtar Khan, Prashant Singh, Tanishq Dattatray Ige, Vipin Tiwari, Rajab Ali Mondal, Harshini K, S Reka, Chetana Amancharla, Faiz ur Rahman, Harikrishnan P A  
**arXiv:** [2406.00010](https://arxiv.org/abs/2406.00010)  
**Date:** 2024-05-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.00010v1/

## Key points

- Challenge of handling proprietary unstructured data in enterprise data management
- Solutions emerging to extract insights for effective employee inquiries
- Reliance on pre-trained embeddings and generative models in these solutions
- Suboptimal alignment of pre-trained embeddings with unique characteristics of enterprise data
- Proposal of fine-tuning pre-trained embedding models specifically for enterprise settings
- Dataset used includes diverse range of internal Infosys Ltd. data sources
- Extraction process involving various file formats and specialized tools like PDFMiner and BeautifulSoup
- Preprocessing steps such as masking PII, cleaning undesirable elements, and chunking cleaned data into relevant units for question generation
- Findings suggest that fine-tuning embedding models can enhance information retrieval performance in enterprises

## Summary

In the realm of enterprise data management, the challenge of handling proprietary unstructured data has become a significant obstacle to efficient information retrieval. To address this issue,  have emerged, aiming to extract pertinent insights to meet employee inquiries effectively. These solutions often rely on  and generative models as foundational components. However, while pre-trained embeddings may demonstrate proximity or disparity based on their original training objectives, they may not align perfectly with the unique characteristics of . This can lead to suboptimal alignment with the retrieval goals of enterprise environments. In response to this challenge, a  is proposed in this paper to fine-tune pre-trained embedding models specifically for enterprise settings. The dataset used for this study includes a diverse range of internal Infosys Ltd. data such as technical course contents, internal knowledge base articles, standard operating procedures for technical tasks, a repository of internal technical queries with resolutions, sales data, and employee blogs. Text data was extracted from various formats including PDFs, MS Word documents, Excel sheets, PowerPoint presentations, and web pages. The extraction process utilized Langchain's document loaders and specialized tools like PDFMiner and BeautifulSoup for efficient data extraction from different file formats. The collected data then underwent preprocessing steps including masking personally identifiable information (PII), cleaning undesirable elements like XML tags and non-ASCII characters using Python libraries like lxml and clean-text. Furthermore,the chunking process segmented the cleaned data into contextually relevant units suitable for synthetic question generation. This involved denormalizing structured data into plain text key-value pairs and separating contextually independent paragraphs from unstructured text streams. Overall,the findings suggest that fine-tuning embedding models for enterprise environments can enhance information retrieval performance by improving precision and relevance in search results. By adapting embeddings to better suit retrieval tasks prevalent in enterprises, this methodology holds promise for enhancing .