August 2023 Summaries

In deepset Cloud, a new hallucination detector for retrieval-augmented generation has been implemented to combat the issue of large language models (LLMs) generating false information, known as "hallucinations". The detector uses a fine-tuned DeBERTa model to compare an LLM's output to a ground truth and produces scores that indicate how semantically similar a sentence is to a given source document. This allows users to identify when a model is hallucinating and take steps to mitigate the issue, such as showing only fully supported sentences or including a disclaimer for partially supported claims. The detector has been trained on data from Ohio State University's project on evaluating attribution by large language models and has been shown to outperform other approaches, with scores that beat their best model by ten points.

Large language models (LLMs) are valuable tools but can often generate unreliable outputs, particularly when they lack information about a topic, leading to a phenomenon known as "hallucination." This issue poses significant challenges for their application in industries where accurate information is crucial. To address this, the team behind Haystack has developed a hallucination detector for retrieval-augmented generation (RAG) systems, which evaluates how closely an LLM's output matches the information in a curated database. This detector assigns a support score to model responses, categorizing them into "full support," "partial support," "no support," or "contradiction," based on their similarity to source documents. This innovation aims to enhance the reliability of LLM-based systems, especially in sensitive sectors like law and finance, by allowing developers to manage hallucinations effectively and decide how to present model outputs to users. While this detector is a significant step towards making LLMs more production-ready, ongoing research and development are necessary to improve their reliability further.

We are thrilled to announce that deepset has raised $30 million in fresh funding led by Balderton Capital, with existing investors GV (Google Ventures), Harpoon, System.One, and Lunar participating in the round, as large language models continue to proliferate and vector search is taking off. Our LLM platform, deepset Cloud, is a model-agnostic developer platform for AI teams looking to build customized, flexible, yet robust systems powered by the latest language models. With SOC 2-certification and full data ownership, deepset Cloud allows organizations to unlock the value of LLMs, get to market faster, and gain a competitive advantage over their peers. The new funding will enable us to further develop our offering for customers around the globe, refining deepset Cloud for RAG applications and optimizing it for virtual private cloud setups, while also improving LLM observability. With this investment, we're poised to consolidate our status as the leading LLM platform.

Deepset, a leading AI platform company, has successfully raised $30 million in fresh funding led by Balderton Capital, with participation from existing investors like GV (Google Ventures) and others. This funding will enhance Deepset's model-agnostic platform, designed to integrate large language model (LLM) features into various products across industries. The platform, known for its SOC 2 certification, allows organizations to maintain full data ownership while accelerating their market presence and gaining competitive advantages. Since its inception, Deepset has been pioneering the application of LLMs in practical use cases, with its core technology, Haystack, enabling developers to build sophisticated NLP projects by integrating various components such as LLMs, vector databases, and more. The company aims to further develop its offerings, particularly for retrieval-augmented generation applications, by improving component evaluation and making the platform suitable for customers with stringent privacy needs. With a team of 50 dedicated professionals, Deepset continues to foster a productive and supportive work environment while remaining committed to its open-source framework, Haystack.

Deepset, a prominent AI startup based in Germany, is gaining recognition and expanding its offerings in the AI and MLOps sectors. With a recent $30 million funding round, Deepset aims to enhance its Large Language Model-focused MLOps capabilities. The company collaborates with Meta's Llama Stack to advance the development of domain-specific, sovereign AI solutions tailored for enterprises, the public sector, and defense. Deepset's innovative approach has earned it a spot on WirtschaftsWoche's list of top AI startups in Germany, as well as recognition on Sifted's "Rising 100" for B2B SaaS in 2025. Esteemed organizations, including Airbus, The Economist, and OakNorth, are leveraging Deepset's AI platform to address their unique needs, illustrating the company's growing influence and impact in the industry.

Document search is a technology that helps computers find relevant documents in response to a query, making it an essential tool for many applications such as business, law, and education. It uses various methods including metadata, text embeddings, and vectorization to transform unstructured text into structured vectors, allowing for efficient searching and retrieval of documents. Document search has numerous applications including semantic site search, document retrieval, and file similarity, which can be used to improve user experiences, automate tasks, and enhance research. By leveraging advances in large language models and NLP, document search is becoming increasingly important as data continues to accumulate and grow, enabling the creation of more effective search systems that serve specific use cases.