Local computation
Inference is the execution of a model to produce an answer. A framework such as MLX, developed by Apple, supports machine learning computation on Apple Silicon.
A framework alone does not guarantee performance: the model, configuration, hardware and task need to be specified.
Documents and context
Retrieval-augmented generation, or RAG, retrieves relevant passages from documents and supplies them to a model. The paper by Lewis et al. (2020) describes this approach.
Finding a source and generating an answer are different steps. Their quality needs to be evaluated separately; a citation does not automatically make an answer accurate.
Model adaptation
LoRA, by Hu et al. (2021), is an adaptation method that reduces the number of parameters to train. It is a technical reference, not evidence that a particular personalisation has been validated by EVA-I.
Evaluation · In preparation
A result needs context: the task, model, hardware configuration, protocol and measurements used. EVA-I’s own documentation remains to be published with these details and their limitations.
No EVA-I benchmark or independent audit is presented on this page.
External references
- MLX — framework documentation and code
- Lewis et al. — Retrieval-Augmented Generation (2020)
- Hu et al. — LoRA (2021)
These works belong to their respective authors. Citing them does not imply a partnership or endorsement of EVA-I’s developments.