Denodo on Deepseek R1- Opportunities & Considerations for GenAI Initiatives
Reading Time: 4 minutes

The release of Deepseek R1 on 20 January 2025 was a significant event in the ongoing growth of the generative AI (GenAI) market. This document covers Denodo’s position on Deepseek.

Denodo applauds the release of Deepseek R1 and the ingenuity of the team that developed it. While GenAI has broad transformative potential, its training and operating costs have often stood in the way for many organizations. We consider any technical advancement that lowers costs and simplifies GenAI adoption beneficial for organizations and society. Deepseek has innovated a novel Mixture-of-Experts reasoning model that significantly reduces training and operating costs, which helps lower this adoption hurdle. Additionally, by open-sourcing the R1 LLM code, Deepseek has enabled further improvements by AI engineers and scientists worldwide. For more information on Deepseek’s innovation, please refer to their published research paper: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.

While GenAI has transformative potential for most organizations, getting corporate data into an AI-ready state has been a challenge for some. Via retrieval augmented generation (RAG) the Denodo Platform provides a comprehensive solution to the problem of quickly getting the right data AI-ready, and ensuring the rapid and successful deployment of GenAI initiatives. Some of Denodo’s own customers, such as Perkins Coie, Festo, and Alex Forbes, have already deployed GenAI agents and applications to production, using the Denodo Platform to address all the underlying data management challenges.  For more information on these customers’ use cases, please refer to this webinar recording, in which all three customers’ data leaders provide their testimonials.  

Currently, the Denodo Platform supports using the Deepseek R1 LLM for RAG implementations. The Denodo Platform 9.1 has been tested, and Denodo supports the use of Deepseek R1 as an LLM for RAG, via the OpenAI API. Customers interested in using this LLM with Denodo can confidently do so, as long as they are on version 9.1 or higher.

When deciding whether to use Deepseek R1 versus other LLMs alongside the Denodo Platform, customers should note how Denodo’s RAG differs from most RAG implementations supported by other vendors. Most RAG implementations involve aggregating all relevant enterprise data and storing it in a vectorized format, which LLMs then access when processing prompts that require enterprise data for accurate responses. However, this approach is costly and time-consuming, and also results in responses that are up-to-date only as of the most recent loading of data into the vector database. Denodo’s approach is different: Metadata, not the data itself, is vectorized, and the LLM subsequently accesses this metadata to generate queries (in languages such as SQL) to retrieve the correct live data from original data sources. The AI application then executes this query as needed, through the Denodo Platform. Known as “Query RAG”, this approach ensures real-time updated and accurate responses, without significant data replication and vectorization overhead. Additionally, the generation of queries completes very quickly, within seconds or even sub-seconds depending on metadata complexity and the compute power available to the LLM. Thus, by using Query RAG, an AI application can retrieve correct real-time data and generate responses in a time that is appropriate for operational use cases. To learn more about Query RAG, visit our blog introducing Query RAG: Query RAG – A New Way to Ground LLMs with Facts and create Powerful Data Agents.

As mentioned, Denodo supports Deepseek R1 LLM for Query RAG, though Denodo has noted some differences in performance and accuracy between Deepseek and other models. When Deepseek R1 LLM is used with Query RAG, it applies its Mixture-of-Experts reasoning model against vectorized metadata to generate queries. On average, these queries retrieve more relevant and precise datasets compared to other models we have tested, resulting in more accurate responses. However, the generation of those queries has, on average, taken longer than other models by several seconds. This is not an inherent limitation of Deepseek R1. Instead, it is inherent in applying reasoning models in general against the relatively small amount of vectorized metadata the LLM accesses, compared to much larger datasets if vectorized data were what the LLM accessed. As vectorized dataset sizes increase, reasoning models generally start performing better. Customers should consider this latency versus accuracy trade-off when choosing LLMs for their specific use cases involving Denodo’s Query RAG and are welcome to contact Denodo for assistance.

Finally, Denodo has noted discussions in public media regarding the security and privacy risks of using Deepseek client applications, given those applications connect to Deepseek LLMs running on servers operated by Deepseek AI Basic Technology Research Co, Ltd, and other Chinese entities. Some organizations, such as USA government agencies, have forbidden the use of these applications for this reason. The Denodo Platform does not use Deepseek client applications and has no reliance on servers operated by such entities. Only the underlying R1 LLM model is used. The R1 LLM model’s source code has been open-sourced, and it’s already available from several other providers beyond Deepseek itself, including AWS, Azure and others. Additionally, organizations can opt for reproducing the model completely within their own secured execution environments. Regardless of which LLM a customer chooses, Denodo encourages ensuring that an LLM’s execution environment meets their security and privacy compliance requirements, as they would with any other software or SaaS service accessing their sensitive data or metadata.

Visit our website to learn more about Denodo’s support for GenAI initiatives.

To see if the Denodo Platform is right for your GenAI initiatives, try Denodo Express! This free downloadable version includes the Denodo AI SDK and all other features needed to implement AI agents and applications using Query RAG, along with tutorials and examples. Download Denodo Express here.  

Initially published on InsightJam on February 25th, 2025. Login required.

Dominic Sartorio
Latest posts by Dominic Sartorio (see all)