Airbus tries generative AI in RAG mode

The aeronautical manufacturer shared its generative AI project in RAG at the AWS Summit 2024 in Paris. The Airbus GenAIR team focuses on the subject, testing first use cases of personalized AI, adapting to its activity and its needs.

AdvertisingTo explore the potential uses of generative AI, Airbus created GenAIR, a restricted transversal structure, organized around four subjects: technological monitoring, responsible AI, management of demand in this area among manufacturers and finally , strategy and implementation. And to move forward quickly on the subject with this still small team, the manufacturer opted for a collaboration with AWS Prototyping, on the co-construction model.

Airbus, present at the AWS Summit 2024 Paris on April 3, carried out a first PoC in 5 weeks in June 2023. We started with a solution based on our chatbot platform, hosted on AWS, for team support operational, says Nasser Guesmia, head of AI and advanced analytics within the digital and information management division of Airbus. And we extended it to do response generation. To ensure that these are relevant, updated and adapted to its specific needs, Airbus wanted a personalized LLM model and turned to the RAG (retrieval augmented generation) option. Compared to other solutions, it would offer, according to the manufacturer, a better ratio between simplicity of execution – no retraining of the model, no modification of the parameters, fewer hallucinations – and a good level of sophistication of the results gnrs.

Open Source LLMs

The industrialist’s choice fell in particular on AI that he considers responsible and open Source LLMs. Airbus already used Kendra, AWS’ machine learning-based semantic search engine, for information retrieval. He completed the RAG environment with a Langchain framework for orchestration (search for similarities in the vector knowledge base) and context augmentation, and finally the open Source Falcon 40B LLM model for content generation. Among the use cases identified by GenAIR and AWS are the querying of a corpus of technical business documentation, assistance with the writing of technical requirements, summaries of technical exchanges, but also assistance with the management of requirements for industrial design, for example.

We worked in a very agile way, particularly with the business lines, insists Nasser Guesmia, head of AI and advanced analytics within the digital and information management team at Airbus. (Photo ED)

For its PoC, Airbus also relied on the Opensearch engine and Amazon’s Comprehend NLP for search anonymization, Langchain for augmentation and Llama 2 and Mistral for generation. The input data, mostly unstructured, is vectorized and indexed to enrich the model. We worked in a very agile way, particularly with the trades, insists Nasser Guesmia. We were very eager for their feedback, and we regularly placed the actual AI aspect of the project in the background to get feedback from users.

AdvertisingAn as-code infrastructure for future RAG projects

Among the first feedback from this pilot project, Nasser Guesmia insists on the essential maintenance of the level of safety in an activity like that of Airbus, but also on the importance of defining an as-coded infrastructure to duplicate the architecture for other RAG projects.

To maintain control of the project, the GenAIR team did not hesitate to follow short training courses on certain aspects of generative AI that it did not sufficiently understand. To date, however, Airbus still highlights some drawbacks in the use of technology, regarding the possibility of evaluating the results or the real capacity to scale up the project. And there are more serious questions about the fine tuning of the solution, its performance or even the development of a more advanced RAG.