Your A.I. R&D lab

When a document is injected into the Cube knowledge base, it is deeply analyzed and prepared. To summarize the main steps:

1.The document structure is retrieved.

2.A cutting strategy algorithm is run for token chunk optimization

3.The text is divided into snippets related to each other, depending on the structure of the document.

4.The snippets are indexed according to their content.

5.Snippets and index are stored in Cube’s Vectorized Knowledge Base.

Cube can be queried on your own documents as well as on public documents.

Each question is transformed into a "task" which is a more generic term that encompasses questions, summaries, data compilation, etc.

When a task is set, the first thing Cube does is determine the best strategy to accomplish the task. Cube has several strategies at its disposal and selecting the most appropriate is key. Examples of CUBE strategy models:

•"Cube Direct Query Model”: used to solve a simple task.

•"Cube Splitter Model“: a set of questions is defined, whose answers will help accomplish a complex task. See phase 3 for details.

•“Cube Chain Of Thought Model”: follows a logical chain of thought, asking questions one by one by importance, until the information gathered may complete the task

One approach is for Cube to decompose the main task in relevant sub-questions.

Therefore, instead of one complex question, Cube has now to answer N simple questions whose answers will help complete the main task.

Each question is asked by using the main keywords in the question itself and the indices of the text snippets. Cube then automatically retrieves the selection of snippets (as relevant information may be located in several snippets).

For each sub-questions Cube then produces an answer that can be tuned to produce short or verbose answers.

Finally, Cube complete the main task by compiling all the sub-answers and by extracting the relevant information needed to answer.

Once the tasks analysis is done, the most relevant document chunks are cut and prepared, the indexes are ready, the whole set is optimised to get the best summary answer from the LLM AI model. The final answer can be short or verbose depending on the settings.

Then the whole set is sent to a local or SAAS LLM. Cube is using GPT and will connect other models such as AI21, Claude, Mosaic, DeepMind, LLaMA, Salesforce

 

* Cube may be used from an to any languages.

* Any request to/from Cube is fully anonymous (non PII), GDPR compliant