Initial situation

Cloud-optimised search platform with advanced language understanding

Search is a key feature in many applications. For this client, this was particularly true: an electronic tendering and procurement system is doomed to fail without an excellent search function.

The search must create an optimal match between buyers and suppliers in professional markets. Only then can suppliers find all the tenders they are interested in and only then can buyers receive an optimally representative set of offers.

Our client was in the process of programming a new integrated tendering platform that would unify a range of acquired products. In the process, the development team encountered serious relevance and performance issues with one of the most important functions - the tender search.

In 2020, we at NEOMO had started researching how Large Language Models could improve the way we consume information and acquire knowledge.

Together with our client, we were able to put these technologies to the test in practice.

Objective

Fast, reliable, redundant and optimised for European languages - this is what the future search component should look like.

The AWS cloud was chosen as the deployment target: The system should be able to scale automatically according to demand and it should also be possible to deploy it flexibly and automatically in test and staging environments.

In addition, the specific linguistic phenomena of Norwegian, Danish, Swedish, Dutch and English were to be taken into account - to an extent that went far beyond the standard capabilities of Solr and Elasticsearch.

Deployment and scaling

AWS and Terraform were chosen by the client. They also wanted Elasticsearch to be used.
We expanded the deployment so that it used AWS' auto-scaling tools: it automatically adapted to high-load phases, but was operated with a minimum number of nodes during periods of low query volume and in the test system.
Through extensive load testing, we proved to ourselves and the client that this approach worked flawlessly and indeed provided the best balance of cost and performance.
We integrated the finished system into the client's "data highway" so that all updates were reflected in the search results in real time.

Performance optimisations

In almost every search scenario, users expect response times of less than one second. In many cases, the actual response time is significantly slower.
In applications such as this one, in which many programmatic, complex queries are executed, performance is usually even more unsatisfactory.
Based on the scalable architecture, we optimised the Elastic index structure, as well as the queries and alerts, until the average and maximum response times matched user expectations.
We used automatic regression tests to ensure that the latency would remain within limits even with future extensions.

Language support

Search is invariably also "translation". The language in which the user formulates the query is always slightly different from the language of the documents they are searching for.
In the first step, the customer definitely wanted to ensure optimal matching for all Scandinavian languages and English.
Elasticsearch is a fantastic search engine, but the natural language modules included as standard are not great.

To fix this, we rolled out a series of bespoke language modules, which included the following functions:

Pre-processing of tender documents
Extraction of relevant key data from the documents
Support for language phenomena such as inflection, compound words, etc.
Error tolerance in the search input

This brought the language of user enquiries on the one hand and tender documents on the other closer together.

Result

We redesigned the search platform from the ground up and achieved a performance increase of several orders of magnitude, while at the same time significantly increasing the relevance of the results - making the search function faster, better and, as a result of the scaling changes, cheaper.

In the course of the project, we modernised the entire search backend and the integration into the data flow. Our own cross-functional team was seamlessly integrated into the client's agile development process. This enabled us to achieve fast communication channels and an efficient exchange of ideas.

Haavard Melby, Group Product Director, Mercell International

"NEOMO helped us move our search from a database lookup to a state of the art search engine. Solving complex user problems and increasing our customer satisfaction. Their knowledge and communication was instrumental when creating our product vision and implementing it into the company."

Start a conversation with us.

We are always interested in a good conversation, whether with a specific reason or simply to exchange ideas.

You can reach us via the contact form or you can book a meeting directly.