Strengths and Limitations

This overview page covers the strengths and limitations of Lazarus' Large Language Models (LLMs), a type of AI model specialized in understanding and predicting language.

Model Strengths

Three key features of Lazarus’ model design set them apart from virtually all other models.

Spatial Understanding of Documents

The documents you run through the model, or the model’s input, is understood by the model in a two-dimensional way. This means the model ingests each part of your document both as the text it represents and where in the document the text is located. This allows for more complex questions and more effective answers.

Very Low Hallucination Rate

Hallucination rate is understood as what percent of the time a given model “makes stuff up” in its output. Lazarus’ RikAI-2 has a hallucination rate of under 1%. So long as you upload sensible documents and ask effective questions, Lazarus’ models will almost always return the right answer.

Primarily Extractive

Lazarus’ models are primarily extractive, so their first-order process is attempting to find a given answer from the input dataset. If the prompt’s answer cannot be directly found in the input dataset, our models will use their reasoning capability to generate a most probable answer or decline to answer (i.e. “Not in context” or “N/A”) based on the input data, unless prompted otherwise.

Model Limitations

There are a few key constraints and considerations to ensure effective and informed use of the models in various applications.

Math Calculations

Lazarus’ models are not effective at performing algebraic calculations. Since they are trained on producing language and document understanding, they lack advanced mathematical ability.

Time-based Understanding

Lazarus’ models only have a high-level understanding of time, thus, whether a given event happened before or after a specific date may pose a challenge at times. The effectiveness of this capability depends on how complex the question is and the readability of the document.