LexisNexis builds a time-aware AI stack beyond standard RAG with MongoDB

The wild west of data: how standard RAG falls short

At LexisNexis Risk Solutions, navigating financial and commodities markets means dealing with decisions worth millions of pounds, making data accuracy non-negotiable for the company’s clients. Operating at an immense scale, it publishes 2,000 articles daily and tracks over 2,200 global commodities. LexisNexis knows that while enterprises increasingly look to AI to synthesize dense data, typical search engines can fail because they cannot understand how facts connect and change over time.

Tackling this challenge head-on, the company’s ‘Maverick’ team—its lean, three-person innovation unit—set a clear mission: move past the limitations of basic semantic search and build a time-aware reasoning system capable of powering reliable enterprise decision-making. They realized early on that traditional Retrieval-Augmented Generation (RAG) architectures lack structural grounding, treating information as independent text chunks and stripping away vital context.

"Typical RAG implementation, at the end of the day, it's really just semantics. And the problem with semantics is, it doesn't understand the relationship of data,” explained James Melvin, Principal Innovation Software Engineer at LexisNexis Risk Solutions.

This blind spot is especially obvious in what Melvin calls the "wild west" of user inputs, where someone might search for the "United States" in nine different ways. Beyond being fed messy, inconsistent search terms, standard AI also completely fails to understand time, and in fast-moving financial markets, facts change from day to day.

"Today the share price is 100, tomorrow it's 110. Neither of those facts are false," said Melvin. Yet, passing a disorganized jumble of historical dates to a Large Language Model (LLM) invites hallucinations, making a structurally grounded, time-aware approach the only viable path for delivering reliable answers in high-stakes environments.

Bypassing size limits with scalable database collections

To solve this, the Maverick team built a time-aware reasoning platform. As they mapped out the architecture, they realized the sheer scale of their data quickly outgrew out-of-the-box solutions. For instance, MongoDB's native knowledge graph feature, optimized for limits up to 100MB, couldn't support an enterprise tracking complex logistics such as 184,000 maritime voyages.

But instead of switching databases, Melvin’s team leveraged MongoDB's core flexibility. They turned the problem on its head, bypassing the native graph feature entirely and treating the immense graph infrastructure as scalable, native database collections.

"We set out to say: can I make the nodes a collection? Can I make the relationships a collection? Can I make the vectors a collection? Can I have the ontology as a collection? The idea was can I have a single store of truth?" said Melvin.

By organizing the graph into schemaless collections, LexisNexis unlocked limitless scaling and the ability to seamlessly search historical timelines. This was crucial for tracking fast-moving financial markets where ownership constantly shifts. Melvin highlighted a common scenario: if Shell sells an asset to BP two years later, a standard knowledge graph returns both "Shell" and "BP" as the owner. This confuses the LLM because technically, neither statement is historically false.

To fix this, LexisNexis pioneered a custom time-stamped architecture that attaches strict start and end dates to every single fact. Now, if Shell sells an asset, the platform records the precise window of ownership. Rather than relying on the LLM to perform maths across conflicting dates, the database delivers a pre-filtered, chronologically accurate timeline.

Overcoming the LLM bottleneck with MongoDB and Rust

To process this massive scale of data and save valuable processing time for the LLM, the team paired MongoDB with the programming language Rust. In their performance tests, Rust ran 14 times faster than Python.

Melvin admitted that while he loves Python, the language simply isn't fast enough to handle millions of enterprise transactions. Because waiting on an LLM response is already a natural bottleneck in these systems, the platform's underlying code has to execute almost instantaneously to avoid wasting precious processing time.

The team wrote the heavy computational algorithms in native Rust but cleverly packaged them so data scientists could still work in their preferred language (Python) while relying on a high-performance backbone.

But compute speed was only half the battle; the team also needed extreme data flexibility. Melvin initially started with PostgreSQL, a traditional relational database, but its rigid structures slowed down development when dealing with unpredictable data formats.

"I quickly had to go: 'I don't know what the schema is... I need a schemaless instance,'" Melvin recalled. And MongoDB provided that exact operational freedom.

Eradicating the synchronization tax with a unified AI stack

When designing its advanced AI search, the Maverick team initially evaluated standard point solutions—like Neo4j for graphs, dedicated vector databases like Qdrant, and Snowflake for transactional records. However, they realized this fragmented approach would force them to pay heavy "synchronization tax" to keep everything in sync.

By unifying its entire AI stack on MongoDB Atlas, LexisNexis avoided this complexity, running its ontology, knowledge graph, vector search, and operational data within a single, secure environment.

This unified foundation dramatically improved query precision, particularly when interrogating massive tabular datasets. Standard "Text-to-SQL" tools are notorious for failing in production, yielding just 60% accuracy because they lack domain context. The Maverick team bypassed this by using their MongoDB-powered GraphRAG framework to supply high-fidelity context, automatically generating highly accurate MongoDB Query Language (MQL) queries.

"MongoDB helps us to run one central data store that holds everything. It holds our ontologies, it holds our knowledge graph, it holds our tabular data, it holds our textual data, it is our graph rag engine which helps us with local and global searches. It's literally the store of our AI,” said Melvin.

Today, the Maverick team's five-month rewrite proves that the future of enterprise AI lies in simplifying, not bolting on. By consolidating on MongoDB Atlas, LexisNexis has created a fast, time-aware platform that gives clients absolute decision-making confidence.