The architectural roots of hallucinations

When an LLM invents a data point, an entity, or a corporate policy that doesn't actually exist, it is almost always because the retrieval pipeline that fed the prompt was inaccurate, stale, or incomplete. There is a precise name for the space between a system that returns basic answers and a system whose answers can be confidently acted on in production: The trust gap.

An agent reasoning over yesterday's customer snapshot, or an embedding model that was swapped without reprocessing underlying records, creates a major operational hazard. It does not produce a visibly wrong answer; it produces a highly confident hallucination. A hallucination in a single sandbox demo is a curiosity, but the same error applied across thousands of live interactions a minute is an incident.

Introducing contextual integrity

Accuracy is a foundational architectural property best defined as contextual integrity: The discipline of keeping live operational data, vector embeddings, search indexes, and agent state automatically consistent rather than manually reconciled after the fact.

When these four variables live in the same underlying record served by the exact same query path, there is no synchronization step for accuracy to leak through. The retrieval an agent receives is securely grounded in the operational truth of the exact same millisecond it was issued.

Voyage AI’s models lead the Retrieval Embedding Benchmark (RTEB), outperforming comparable models from Gemini, Cohere, and OpenAI on the enterprise RAG tasks agents actually run.

MongoDB has delivered the building blocks required to close the trust gap natively within the data layer:

• Voyage 4 series interoperability: A family of shared embedding space models (voyage-4-large, voyage-4, voyage-4-lite, voyage-4-nano) that produce fully compatible embeddings. The flagship voyage-4-large is the first production-grade embedding model built on a mixture-of-experts (MoE) architecture, delivering state-of-the-art accuracy at roughly 40% lower serving cost than comparable dense models. This eliminates the need to reprocess whole data collections when changing cost or latency profiles.
• Asymmetric retrieval: An industry first where documents are embedded once with voyage-4-large for maximum baseline accuracy while continuous queries are embedded using voyage-4-lite to keep latency and token costs low.

Earn Skill: Voyage AI with MongoDB

Explore the solution

The builder's journey: Core filtering rules

For engineering teams looking to move past conceptual talk, practical implementation requires examining real filtering logic under load. A key example is filtered vector search.

MongoDB handles this by filtering by attributes (such as user ID) first and then running the vector search. In contrast, alternative relational planners often run the vector search first and attempt to apply filters afterward. When relational engines search first, highly relevant results frequently fall out of the candidate evaluation set before they can ever be scored. This minor architectural variation has massive accuracy consequences at enterprise production scale.

Watch: The Builder’s Journey: Building AI Apps With MongoDB

LG U+ balances 3.5M monthly consultations

Conceptual architecture arguments hit a very clear operational ceiling when a business runs a call center taking 3.5 million live customer inquiries a month. Korean telecom giant LG U+ set out to build Agent Assist: An intelligent internal tool giving 4,000 human agents real-time transcription, automated inquiry categorization, and contextual summaries during live consultations.

Their legacy relational database simply could not handle vector embeddings, transactional data, and real-time query tests simultaneously within the same application path. By migrating core operational records off Postgres and consolidating them onto MongoDB Atlas, the team entirely eliminated the latency-prone synchronization layer between disparate systems. $vectorSearch runs semantic similarity algorithms natively against the exact same data collections that human agents query for transaction histories.

The service launched just four months after development began, holding sub-second query latency even during peak morning spikes. Processing time per call dropped 7%, and resource efficiency improved 30%. Initially deployed for 1,200 agents, the system now handles over a million queries per week and is scaling toward 2,300 agents.

The production checklist before you ship

What LG U+ demonstrates is the absolute checklist a data layer must deliver before a production agent can be trusted:

1. Co-located indexes: The AI search index and live operational data must reside in the exact same database cluster so agents never retrieve from one version of truth while acting on another.
2. Sub-second latency: Semantic queries must execute in under a second under heavy loads or the agent response arrives too late to be operationally useful.
3. Auto-scaling infrastructure: The platform must automatically scale to absorb morning call spikes without requiring engineers to manually re-provision nodes.
4. Schema flexibility: The underlying data model must dynamically accommodate varied conversation shapes without forcing rigid schema migrations every time a business iterates.