The increased recovery generation finds a new dimension in graphic technology. The point on this approach which already impacts conversational assistants.
The subject is no longer debated. A founding research article dating back to November 2023 demonstrates how the graphs deeply impact the rag or increased generation of recovery. A device which, let us remember, allows you to glean information from an intelligent assistant within a documentary base. Result: the graph allows you to multiply by three the degree of precision of the RAG. “The Graph mode clearly corresponds to the new generation RAG”, loose Didier Gaultier, boss of AI within Orange Business Digital Services. “Unlike the purely textual vector space, the vector space of the graph is closer to the meaning insofar as its semantic force is much more important.”
Basically, the RAG consists in vectorizing the questions asked of the conversational assistant on the one hand, and the files of the documentary base to be questioned on the other hand. Objective: to glean similarities between the request issued and the research elements by comparing the two vectors, in order to ultimately produce the adequate response. The RAG works very well when the questions asked relate to well -defined content, especially when the subject concerns an information element available in one or more documents. In this case, we can find the information quite simply insofar as all the content is vectorized. At the end of the race, it is then translated by the natural language assistant.
“But what happens if the question does not concern content but dependencies or interactions. This is the case for example if we seek the impact of new legislation on internal procedures, or the existing links within a complex logistics chain to anticipate its weaknesses. In these cases, the graph finds all its meaning”, explains Nicolas Rouyer, senior presales consultant at Neo4J, Graph oriented.
An upheaval in knowledge management
Another example: the search for side effects within the prescription notices of drugs. “The question that can arise is whether side effects are present in several instructions or appear with drugs taken at the same time. In these cases, we can also call on a graph,” argues Nicolas Rouyer.
And the NEO4J consultant to specify: “If I stop at vectorized content, I do not capture the links between concepts. I am therefore not able to ask deep questions.” Conclusion: with the graphs, it becomes possible to make Deep Reseach on the data of the company, without having to re-trainer the LLM.
According to Didier Gaultier, the GRAD -oriented RAG promises in particular to upset business skills management. “Via its vector mode, the graph allows you to put your finger on missing skills in CVs by comparing the vitae curriculums between them by play of associations. If an employee presents a skill A and a skill B, he may necessarily have a competence C starting from the observation that these three skills are present in a significant number of other CVs. By questioning the graph via a RAG, That you would not have found otherwise, “explains Didier Gaultier.
Another advantage of the graph within the framework of the RAG: the answer to a question asked can improve as the graphic enrichment in the form of dynamic data enriched over the conversations. The graph will also make it possible to define much finer access rights compared to a simple vector documents base. They can be defined in terms of the relationships between nodes and/or in terms of detail of nodes as such.
“Klarna’s CEO has decided to reduce its information system to a unique graphic oriented knowledge base coupled to an LLM and its associated chatbot”
To summarize, the graph makes it possible to connect concepts within the same documents and between several documents. “We see that this logical layer enriches the increased generation of recovery in a spectacular way. It will allow it to deliver not only an answer but also a whole context which will allow you to better understand it”, summarizes Didier Gaultier. “The fact remains that the results delivered will be de facto very bulky.”
To solve the equation, the idea is to use the graph to identify the distance between the question asked and the possible answers, and thus optimize the volume of data output of the model. Reverse of the medal: This treatment involves large Language Model (LLM) or Small Language Model (SLM) equipped with important context windows. “This type of LLM and SLM should be born this year. What will go through sacrifices, for example integrate less information within the neurons of the model while increasing the RAM for execution”, anticipates Didier Gaultier.
This obstacle did not prevent certain companies from getting started. Among the references of NEO4J in the RAG Figure Klarna. This Swedish fintech has set up a customer relationship management chatbot based on an LLM combined with a graph of knowledge. This allowed him to increase his performance in the field of support. Klarna then declined her chatbot for the internal support of employees. “There followed a phenomenal progression of productivity. Suddenly, Klarna’s CEO has made a radical decision. Note: reducing its information system, which counted hundreds of applications, to a unique graphic knowledge base coupled to an LLM and its associated chatbot. Exit SaaS systems like Salesforce or Workday”, underlines Nicolas Rouyer. “By going through it, Klarna was able to delete almost all of her existing computer that was cut into silos to lead to a simple conversational interface with a unique and coherent database.”
A complex site
“The whole question is to find LLM or SLM adapted to the rag oriented Graph”, warns Didier Gaultier, before recognizing: “Databases No4J but also Elastic are well placed to position itself on this new market.” At Orange, we started a first research project on this field. “This idea of rag oriented graph is targeted by Apple (As part of its Apple Intelligence project, editor’s note), but also by Microsoft with his Copilot assistant. However, these actors have not yet managed to make it happen. Which shows the complexity of this quest, “notes Didier Gaultier.
Where do these difficulties come from? According to the boss of AI at Orange Business Digital Services, they should be seen in the apprehension of the graphs which implies for companies to review their logic of definition and organization of data. But also in the complexity of vector spaces. A technology on which generative AI is largely leaning, but also theories of dimension reduction or quantum physics.
In the United States, the Writer language model platform has already included a graph brick at its offer. “Thanks to the semantic relationships woven by this brick, we are able to carry out robust and very fast analyzes on the basis of our customers’ documentary content,” JDN Kev Chung, Chief Strategy Officer in Writer.
What about the next step? “It will emerge with the possibility of combining the rag oriented graph with information based on any type of media in a multimodal logic”, anticipates Didier Gaultier.
