FRAIM uses Elasticsearch to build a knowledge search platform for the AI era

LAWGUE, a cloud document workspace product provided by FRAIM Inc., has been widely adopted by a diverse range of users, including business enterprises, law firms, government agencies, and local government authorities, as a tool that enables efficient document compilation. Starting with contracts, the aim has been to transform a wide variety of documents, including rules and regulations, manuals, and disclosure documents, etc., into reusable knowledge.

The basic thinking behind LAWGUE is to structure documents as combinations of semantic units (units that encapsulate meaning), to facilitate search/retrieval and reuse. In today’s era of advanced artificial intelligence, including AI agents, there is an increasingly vital need for a search platform that enables users to access the right knowledge and context with precision through the interlinking of multiple layers of data, including natural language query (NLQ), document structure, metadata, and business process context. To meet these requirements, FRAIM adopted the Elasticsearch Platform, which provides a high level of flexibility and extensibility.

Business enterprises usually have large quantities of business documents but often struggle to make full use of them due to poor searchability and reusability. While the documents themselves may be preserved, the document structure is typically not designed to facilitate searching across them and reusing them, meaning multiple steps may be required to access and act on the information. Even when relevant reference documents are found, users often face significant manual effort to adjust formatting details such as indentation, numbering consistency, section structure, and terminology.

FRAIM’s vision is to address this issue by effectively reinventing document compilation. In line with this vision, FRAIM has developed and launched LAWGUE, a service which restructures documents so that they are made up of semantic units in a way that makes it easier to search in and reuse the documents.

LAWGUE has AI-enabled editing support and proofreading assistance functions. Besides substantially reducing the time and effort needed for revising and proofreading, the system also provides value as a platform for making documents a form of knowledge that has practical utility, rather than just something to be kept on file.

By adopting this approach, which breaks down documents into their component parts, it is possible to collate and accumulate not just documents and fragments of documents that were compiled in the past, but also the related metadata, context, and usage trends, so that all of this can be treated as knowledge assets with multifaceted potential for reuse. In the future, the scope of document usage will be further expanded through more systematic modeling of the relationship between different items of information, and the development of knowledge graphs and ontological structures.

Takao Mizuno, VP of Machine Learning at FRAIM, previously worked as a machine learning engineer at several other companies and has accumulated a wealth of experience working on a diverse range of projects that involved natural language processing, audio processing, and image processing. Through his exposure to these diverse fields, he came to appreciate how image and sound interface technologies can generate value and the practical impact of natural language processing, which is directly linked to thinking and knowledge, as well as the practical benefits of search and recommendation technology that can guide users toward accessing useful knowledge from huge volumes of information.

With this background, he joined FRAIM in 2021 as a part-time and contract contributor, formally joining the company in 2022. He currently leads the development of machine learning models, search technologies, and systems using LLMs and AI agents at FRAIM.

After joining FRAIM, Mizuno was asked to improve text search and vector search across different frameworks. Vector search in particular had been implemented through a unique structure that affected search performance and led to higher costs when scaling. So, he embarked on a comprehensive overhaul of the vector search platform. After comparing several alternative technologies, Mizuno realized that integrating vector search with the Elasticsearch cluster, which FRAIM was already using for text search, would enable a substantial reduction in management and development costs. Performance tests confirmed that k-nearest neighbor (kNN) search could be performed at a practical speed, so Mizuno proceeded to explore effective strategies for integrating vector search with Elasticsearch.

Additional comparative research showed it was possible to make search logic and scoring design consistent with the same query DSL and to implement flexible index design in line with scaling and other requirements. After comprehensive evaluation that recognized the platform’s flexibility to expand the system’s capabilities through analytical processing plug-ins that included a Japanese language tokenizer, the final decision was made to adopt a strategy based on using Elasticsearch for integration that would include vector search.

Through this integration, FRAIM has succeeded in building an environment that enables optimization with a unified platform for all search functions, reducing costs by over 50% compared with the separate platforms that were used before.

FRAIM is not just using Elasticsearch on its own; adopting Elastic Cloud makes search platform operations less burdensome. Utilizing managed services minimizes the costs associated with server management and infrastructure maintenance, and it is relatively easy to scale out or scale up when needed. In this way, the development team is able to allocate more time to enhancing search quality and developing new functions.

FRAIM is also implementing on-premises service rollout for government agencies and local government authorities. By utilizing Elastic Cloud on Kubernetes (ECK), FRAIM has been able to implement a system that reproduces the same structure and settings as the cloud-based version. This approach ensures a consistent design in both the cloud-based and on-premises environments, so that customers have the flexibility to choose a deployment strategy that matches their security requirements and infrastructure policies.

Mizuno notes that "the search/retrieval function of LAWGUE is an important area that is one of the main reasons why users choose LAWGUE." Based on the document that is being compiled, LAWGUE automatically suggests similar expressions and related knowledge from applicable documents, as well as from semantic units and other document parts. By making potentially useful content immediately available in this way, it makes the actual process of compiling documents more efficient.

This search experience is underpinned by Elastic’s flexible search platform. It supports a wide range of search approaches, including traditional BM25 as well as sparse vector techniques such as SPLADE and ELSER based on Transformer models, in addition to dense vector search. The flexible integration of a diverse range of search signals, including score weighting and scripting, precise conditional search using Boolean queries, metadata-based filtering, and aggregation, makes it possible to implement search logic tailored to user requirements with a consistent architecture. The environment facilitates improvements reflecting user feedback and allows smooth implementation of search while supporting new functions and services.

Mizuno has set up a Search Committee within FRAIM to realize ongoing enhancement of search quality, and the company has continued to make improvements based on customer feedback and log analysis. He explains that “Elastic offers a great deal of flexibility in terms of DSL adjustment, so when we have been discussing a particular issue, we can move straight into implementation. The ability to quickly cycle through prototyping and validation is a major advantage.” FRAIM has pursued these efforts since before generative AI and large language models (LLMs) were widely adopted in practice, and today they form a foundation that supports an advanced search experience in the AI era.

The cumulative impact of these efforts has made a significant contribution toward the development of RAG and AI agents. In many cases, search engines and vector search platforms need to be rebuilt from scratch, but because LAWGUE already has search assets built around Elastic, these can be reused as they are, resulting in a substantial reduction in development effort.

Today, with the widespread adoption of AI, it is increasingly important to have a search platform that enables the rapid, precise extraction of the context that users need. Mizuno goes on to point out that "In the future, there will be a vital need for search capability that is able to handle operational context, based on information that is not necessarily explicitly stated in the document, such as the company's existing corporate context and organizational structure, its operational flow, the record of decision-making as embodied by emails, chat messages, and comments, etc."

Elastic's flexibility and extensibility provide a solid foundation that will support LAWGUE and FRAIM’s overall business operations when undertaking this kind of integration of operational context and the future development of RAG and AI agents.