Contextual you may have unveiled his grounded voice model . GooglePresent Anthropic And Openai On a very important yardstick for truthfulness.
The startup founded by the pioneers of Repetition generation (RAG) technology, reported that your GLM has achieved an 88% factual assessment value on which Facts benchmarkin comparison with 84.6% for Google's Gemini 2.0 lightning79.4% for anthropics Claude 3.5 Sonett and 78.8% for Openai's GPT-4O.
While large language models have modified the corporate software, factual inaccuracies – sometimes called hallucinations – remain a decisive challenge for economic surveillance. The context -related AI goals to unravel this by making a model that’s specially optimized for company lag applications through which the accuracy is of the utmost importance.
“We knew that a part of the answer could be a method called RAG Retrieval Augmented generation,” said Douwe Kiela, CEO and co-founder of context-related AI, in an exclusive interview with Venturebeat. “And we knew that because RAG was originally my idea. This company is admittedly about making RAG in the correct method to make the following stage of the rag. “
The company's focus is significantly different from general models reminiscent of Chatt or Claudethat are designed in such a way that they edit every part from creative writing to technical documentation. Instead, the context -related AI is directed with high corporate environments through which the factual precision outweighs creative flexibility.
“If you may have a rag problem and are in an organization environment in a heavily regulated industry, you may have no tolerance for hallucination,” said Kiela. “The same general language model that is helpful for the marketing department is just not what you wish in a company environment through which you might be way more sensitive to mistakes.”
How context -related AI “Groundness” makes the brand new gold standard for corporate language models
The concept of “Soilness” -Ensuring that the AI ​​answers are strictly related to information that was expressly provided in context, has proven to be a critical requirement for company -KI systems. In regulated industries reminiscent of finance, healthcare and telecommunications, firms need AI that either provide precise information or explicitly recognizes once they know nothing.
Kiela offered an example of how this strict property works: “If you give a typical language model a recipe or a formula, and somewhere in it you say:“ This only applies to most cases. “Most voice models still provide you with the recipe, provided it’s true. But our voice model says: “Actually, it only says that this is applicable to most cases.” This additional piece of nuance captures. “
The ability to say that “I don't know” is crucial for corporate settings. “This is admittedly a really powerful function in case you give it some thought in a company environment,” added Kiela.
RAG 2.0 The context -ki: An integrated method to process company information
The platform of the contextual AI is predicated on what it calls “RAG 2.0“An approach that goes beyond the easy connection beyond the drive components.
“A typical loapping system uses a frozen model for embedding, a vector database for calling up and a black box language model for the generation, which was put together by requesting or an orchestration frame,” says a declaration of corporate. “This results in a” Frankenstein's Monster “generative AI: the person components work technically, but the entire thing is anything but optimal.”
Instead, the context -related AI together optimizes all components of the system together. “We have this component of the retriever mixture, which is admittedly a method to achieve intelligent calls,” said Kiela. “It deals with the query, after which it essentially thinks, like most of the most recent generation of models, (and) First it plans a method to call up.”
This entire system works in coordination with what Kiela calls the “best repeater of the world”, and helps to prioritize essentially the most relevant information before sending to the grounded voice model.
Beyond the easy text: Contextual AI now reads diagrams and establishes a connection to databases
While the newly announced GLM focuses on text generation, the platform of the context -KI recently added support for multimodal content, including diagrams, diagrams and structured data from popular platforms, BigqueryPresent SnowflakePresent Red shift And Postgres.
“The most difficult problems in firms are on the interface of unstructured and structured data,” said Kiela. “I’m mostly excited, this interface of a structured and unstructured data is admittedly. Most of the really exciting problems in large firms are the overlap of structured and unstructured bang, where they’ve some database data records, some transactions, possibly some guideline documents, possibly quite a few other things. “
According to Kiela, the platform already supports a lot of complex visualizations, including circuit diagrams within the semiconductor industry.
Future plans of the context -ki: Creating more reliable tools for day by day business
Contextual AI plans to publish its specialized new-ranker component shortly after the GLM start, followed by expanded functions for understanding documents. The company also has experimental features for more acting skills in development.
Founded in 2023 by Kiela and Amanpreet SinghThe context -ki previously has a fundamental AI Research (fair) team (fair) and hugs. He has secured customers reminiscent of HSBC, Qualcomm and The Economist. As an organization, the corporate positions itself to finally recognize concrete returns for his or her AI investments.
“This is admittedly a possibility for firms that could be under pressure to deliver the ROI of AI to search for more specialized solutions that really solve their problems,” said Kiela. “And a part of it is admittedly a grounded language model that is probably a bit more boring than a typical language model, however it is admittedly good to make sure that it’s justified within the context and that you may really trust that it does his job.”
