Titulo Estágio
Augmenting Large Language Models with Context Retrieval
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
Remote (or OutSystems Linda-a-Velha office)
Enquadramento
OutSystems is a low-code development platform, where developers can build applications fast by dragging and dropping visual components into the development environment. In an effort to further improve the speed of development of applications, OutSystems is infusing Artificial Intelligence across the multiple domains of applications, e.g., automatic suggestion of next-best actions in the development of the UI, or logic flows.
Given the recent advance in the field of Large Language Models (LLMs), and given their potential to be great code assistants, OutSystems is exploring their application to provide development suggestions in the multiple domains of an application. The main limitation to their widespread use is that their performance greatly depends on the context that is passed as input to the LLM. On the one hand, the majority of LLMs have hard constraints on the total size of the context they can handle. On the other hand, there are approaches that are unlimited in the input context they can process, but they suffer from long inference time. On top of this, bigger context sizes increase the GPU memory requirements. Therefore, a careful and smart selection of the relevant context for the task at hand is of extreme importance, and thus the focus of this research proposal.
Objetivo
The main goal of this work is to measure the impact that additional context has in the performance of language models, and to design, implement and validate an approach that automatically selects the context that should be taken into account by the language model.
Plano de Trabalhos - Semestre 1
1 - Review of the literature
2 - Benchmark the impact of the context in the performance of language models
3 - Design the approach for automatic context retrieval
4 - Write the intermediate report
Plano de Trabalhos - Semestre 2
5 - Implement the automatic context retrieval approach
6 - Analyze the performance of the approach
7 - Refine the approach
7 - Write a scientific article with the main results and conclusions
9 - Write the thesis
Condições
The work is to be conducted in the scope of a collaboration between CISUC and OutSystems. A workplace can be provided to the candidate at OutSystems, in the Linda-a-Velha office. To note that reallocation to Linda-a-Velha is not required, and that the thesis can be developed in a remote setup. All required computation resources will be provided to the candidate by OutSystems.
There is the possibility that the candidate is awarded a grant for the duration of the thesis. This is dependent on confirmation, and on the candidate’s performance.
Orientador
Filipe Assunção/Nuno Lourenço
filipe.assuncao@outsystems.com 📩