Meta Global AI Developer Hackathon. $1M in prizes.Get notified
LLAMA 3.1 8B, LLAMA 3.3 70B
Contextual AI
#1
Most factual model on Kaggle leaderboard for FACTS Grounding1Leading
Elo scores in head-to-head arena on industry-specific use cases293.5%
Accuracy on RewardBench, Hugging Face leadership for human preferenceBuilding
on Llama 3 allowed Contextual AI to open source LMUnit*All results are self-reported and not identifiably repeatable. Generally expected individual results will differ.
CASE STUDY
Reducing hallucinations with a better context layer for AI agents
At a glance
Industry
TechnologyUse case
Creating a no-code, agent-building solution grounded in enterprise dataGoal
Design better context engineering to reduce hallucinationsLlama versions
Llama 3.1 8B, Llama 3.3 70BDeployment
Google CloudTHEIR STORY
Turning complex enterprise data into reliable AI solutions with context engineering
Founded by some of the industry’s earliest contributors to retrieval-augmented generation (RAG) techniques, Contextual AI specializes in context engineering for enterprise systems. The company’s modular solution is designed to allow anyone to quickly build RAG-based agents that retrieve information from enterprise data sources, with additional functionality for developers and power users.
THEIR GOAL
Enable enterprises to reduce hallucinations by building a more grounded context layer
Even when enabled with RAG capabilities, generative AI models still hallucinate sometimes. Contextual AI wanted to make it easier for enterprises to connect data sources to their solution and get high accuracy right out of the box. Minimizing hallucinations meant improving the context layer with a grounded language model that could check against source data to ensure responses were accurate. A grounded model responds only with relevant information from a dataset and gives inline attributions to the underlying data sources.
THEIR SOLUTION
Fine-tuning Llama to beat foundational models
Initially, the Contextual AI team believed they would have to train their own grounded language model from scratch, which would have been time-consuming and capital-intensive. But after fine-tuning Llama 3.3, the team was able to beat every commercial foundational model in grounding performance. For the team’s purposes, Llama models have been easier than alternative models to fine-tune and achieve excellent accuracy, groundedness, latency and cost performance.In addition to the Contextual AI Grounded Language Model (GLM), the team also used Llama to develop reward models for their RAG solution, including LMUnit and a groundedness scoring model. LMUnit helps developers evaluate natural language models through fine-grained unit tests, which can streamline debugging. The groundedness scoring model reviews all generated answers to rate adherence to the prompt and context, acting as a final layer of protection against hallucination.
THEIR APPROACH
A unified system for RAG components
To build an enterprise’s context layer, Contextual AI starts by ingesting not just structured data but also thousands, or even millions, of multimodal documents. These are often highly technical and can include diagrams, schematics, charts and tables. Once this information is in the datastore, retrievers and rerankers optimize which sources are prioritized and fed into the generative model. The GLM can then respond to a wide range of queries.The entire pipeline runs on Google Cloud, which includes training on GPUs, inference on Vertex AI, and multimodal data ingest using Gemini 2.5 Flash and Pro.With APIs that orchestrate RAG components as a unified system, Contextual AI helps enterprises scale from early projects to more sophisticated use cases. Developers who want more control can also access primitives for document parsing, reranking, grounded generation and LLM evaluation. This unified approach gives enterprises precise control over which data powers their AI agents.
At the core of the Contextual AI solution is a fine-tuned grounded language model that improves the accuracy of results.
THEIR SUCCESS
Faster time to value-building AI agents
Today, Contextual AI’s solution can derive answers from enormous volumes of complex, highly technical data that would otherwise be impossible for a single human to review. Users can build custom agents capable of supporting PhD-level engineers and providing answers within minutes. That makes the solution especially popular with enterprises in finance, legal services and technology and engineering. More than a dozen partners are already building vertical solutions on the platform.Having achieved the highest score for FACTS Grounding, the GLM from Contextual AI is the most hallucination-resistant and factual model currently available. It’s part of an enterprise-grade solution that offers scalability, security and governance.
- #1 most factual model on Kaggle leaderboard for FACTS Grounding1
- Leading Elo scores in head-to-head arena on industry-specific use cases2
- 93.5% accuracy on RewardBench, Hugging Face leadership for human preference
- Building on Llama 3 allowed Contextual AI to open source LMUnit
- Position on Kaggle leaderboard as of August 19, 2025.
- https://contextual.ai/blog/introducing-grounded-language-model
“With Llama, we can take very good components and make them even better. That’s the power of open source.”
Douwe Kiela, CEO and Co-Founder, Contextual AI
“Instead of buying a point solution for every problem, companies can build their own agents running on our fine-tuned Llama model.”
Mike Klaczynski, Head of Partnerships, Contextual AI
“We used Llama to create a solution that delivers high accuracy and a fast time to value, so companies can get very high-quality applications very quickly.”
Douwe Kiela, CEO and Co-Founder, Contextual AI
Previous slide
Next slide
Start building
Explore more
ConsumerBusiness
Shopify | Llama case studiesShopify uses Llama to generate product pages, localize content, and automate support, helping developers scale workflows and save time.Read more
ConsumerTech
Scribd, INC | Llama case studiesDelivering faster, cheaper and more accurate results with a Llama-powered AI content discovery assistant.Read more
Stay up-to-date
Our latest updates delivered to your inbox
Subscribe to our newsletter to keep up with the latest AI updates, releases and more.