Document360 delivers hyper-accurate AI answers

THE CHALLENGE

Meeting rising demand for accuracy in AI

For enterprise organizations, maintaining a single source of truth for standard operating procedures and software documentation is critical, both for operational consistency and continuity, and increasingly to meet compliance requirements. 

Document360 is a flagship knowledge base platform specifically designed to meet this demand. An independent company that sits under the umbrella corporation, Kovai.co, it offers a centralized repository that helps companies reduce support tickets by allowing end users to access information easily.

“If, for example, you are a software company with a public-facing documentation site that uses Document360, most people won’t have to contact your support if they have an issue,” explained Selvaraaju Murugesan, Senior Director of Data Science at Kovai.co. “They simply go to the documentation site and self-serve. Use cases like that are what make Document360 unique.”

To further enhance the experience, Document360 introduced an AI assistant named Eddy. It was an important step forward, but Document360 quickly encountered some complex hurdles. Users rarely asked simple, single-intent questions; instead, they submitted multi-query prompts, such as asking for a country’s capital and its president in one sentence, which the AI initially struggled to parse.

Because AI is a probabilistic engine—relying on maths-based 'best guesses' rather than predictable yes-or-no rules—even a tiny phrasing tweak could shift the numbers behind the search. “Slight variations in how a user phrased a question also often led to entirely different contexts being retrieved,” added Murugesan. “This resulted in inconsistent answers that eroded customer trust.”

To enable reliable semantic search, Document360 evaluated several standalone vector databases. However, the team realized that moving data out of its primary database, MongoDB, presented significant operational risks. Replicating data to an external vector database would require complex, continuous synchronization. More importantly, it threatened compliance and security, as Document360’s intricate access controls were already deeply baked into MongoDB.

“Using a separate vector database for AI use cases meant we had to replicate all the data and keep it in sync,” said Murugesan. “Also, all our access control for every piece of content is baked into MongoDB. The moment we removed it, we had to take those permissions and transfer them to the vector database. That led to a high level of technical overhead.”

OUR SOLUTION

The power of a unified hybrid search platform

Recognizing the massive challenges of managing disjointed systems, Document360 decided to adopt MongoDB Vector Search on Atlas natively. By unifying its rich text content and vector embeddings within a single ecosystem, the company quickly and effectively resolved its data replication, synchronization, and access-control issues.

To address the AI platform’s struggle with complex prompts, Document360 completely reimagined its query architecture. It implemented “multi-query decomposition” to break down complicated user questions into individual intents, and then deployed a robust hybrid search mechanism to respond to them. By simultaneously utilizing MongoDB Atlas Search for precise keyword matching and MongoDB Vector Search on Atlas for semantic understanding, Document360  now ensures that no relevant context is missed.

“Running both a keyword search and a semantic search at the same time was a unique architectural decision that we took,” said Murugesan.

To refine this retrieved data before it reached the large language model (LLM), Document360 then integrated Voyage AI by MongoDB as a reranking tool. This step filters out noisy content and guarantees that the correct, highly relevant text chunks are passed to the AI.

“Voyage AI is an amazing API. Its unique algorithm picks up relevant content for answering a question, and removes the noisy content,” added Murugesan. “We then have a set of article content that can be passed to a large language model that can produce reliable answers.”

Best of all, relying on a unified platform simplified regulatory compliance. Because data never leaves MongoDB, Document360 can confidently adhere to strict privacy laws like GDPR and the EU AI Act without consulting third-party data management policies.

“We needed a lot more content governance, and our legal and privacy teams needed to know about our data management practices and privacy policies,” said Murugesan. “Having everything in one MongoDB database makes it very easy for us to apply all the key policies. That’s a critical aspect for us.”

OUTCOME

Faster responses, happier customers

The combination of MongoDB Vector Search on Atlas and Voyage AI has driven staggering improvements in performance and customer satisfaction. By eliminating the chaotic variation in how context was pulled, Document360 has achieved an extraordinary 99% content retrieval accuracy rate, and a 97% rate for recall at five, ensuring consistent, reliable answers regardless of how a user phrases their query. “The robust cross-encoder-based Voyage AI reranker helped our chatbot produce consistent responses for a wide range of similar prompts,” said Murugesan.

And because all components are consolidated into a single MongoDB ecosystem, Document360’s developers can avoid complex third-party integrations and leverage highly accessible API documentation to be up to speed and productive quickly.

These architectural enhancements, using Voyage AI’s native reranking capabilities, have propelled Eddy AI’s answering rate from 84% to 91%, while baseline response accuracy has leaped from 85% to 96%. Results are also delivered with lightning speed; closed-loop queries resolve in under two milliseconds, and average retrieval takes just 2-4 milliseconds. 

“Millions of documents can be queried in just 20-30 milliseconds,” added Murugesan. “That’s a highly acceptable range for advanced AI use cases, and it’s a key performance benchmark.”

End users have directly validated the resulting leap in quality, with customer trust restored now that the Voyage AI reranker ensures Eddy AI retrieves the exact context. Positive feedback, in the form of ‘likes’, on Eddy AI's answers has surged by 70%, while the rate of ‘dislikes’ has plummeted by 30%. Use of Eddy AI is growing exponentially as a result; this newfound trust has fueled explosive adoption, with search volumes scaling so fast that the previous month's entire search volume was matched in just the first two weeks of the current month. “Hybrid search strengthened the reliability of our Ask Eddy AI chatbot, leading to increased adoption,” said Saravana Kumar, Founder & CEO, Kovai.co. 

Customers have really started to trust Eddy AI,” added Murugesan. “Because it now picks the exact context every time, the reliability of its answers has risen significantly.”

Document360 now plans to leverage this high-performance foundation to introduce agentic AI capabilities, where the AI doesn’t just answer questions, but autonomously updates and corrects documentation based on user feedback.

“The most important thing about running MongoDB for AI use cases is that it’s a single platform,” concluded Murugesan. “We don’t have to procure stack components independently and then try to integrate them with MongoDB. And everything is part of the package, so we only pay for the compute and storage. It’s fantastic.”