Skip to main content

Key concepts

Key Technologies​

  • Large Language Models (LLMs): Advanced capabilities for text generation and comprehension.
  • Embeddings: Vector representations of data for semantic search and contextual classification.
  • Prompting: Guiding models through specific instructions.
  • Retrieval-Augmented Generation (RAG): Enhancing responses with external information.

Key Components​

Assistant​

An assistant is the main entity that interacts with the user. It is the entry point for the user to interact with the system. After creating an assistant, the administrator can configure it to use specific tools, search for information, and more.

Tool​

A tool is a resource that the assistant can use to perform specific tasks. Tools can be configured to perform various complex tasks. These tools can include fully configurable scripts, Azure functions, or to retrieve information from documents.

Search is a specific type of tool. Information search is performed semantically, using the metadata of the documents to enhance the search.

Topic​

Documents can be organized in a tree-like structure, allowing the administration user to specify topics in a hierarchical way. The tree definition is configurable through the implementation portal.

Document​

A document is a piece of information that the assistant can use to answer questions. It can be a PDF file, a Website or a video.

Block​

Documents are divided into blocks of information. Each block is a section of the document that the assistant can use to answer questions. Each block will have its own embeddings, which are used to search for information within the block. Using the right number of tokens is key to get the best results.

FAQs​

FAQs are a set of questions and answers that the assistant can use to generate answers.

Fixed Questions​

Fixed questions are a set of questions and answers that the assistant can use to answer questions. In contrast to FAQs, fixed questions are not used to generate answers, but rather they are used to provide a specific answer to a question. A common use case is to provide answers to sensitive questions.

Lessons​

Lessons are used to provide context to the assistant. They can be used to provide information that is not present in the documents but that is relevant to answer the questions. For example, a lesson can be used to add jargon or terminology that is specific to a company.

Metadata​

Metadata is used to provide additional information about a document, block, FAQ, or fixed question. It can be used to add other information that can help the assistant to understand the content. The difference between metadata and lessons is the scope, being metadata more specific to the content and lessons more general.

Evaluation​

Evaluation is a component used to measure the assistant's performance. By using metrics based on LLMs, metrics based on embeddings, and traditional NLP metrics, we generate reports for the client on the assistant's performance at a given time. After the implementation stage, it is periodically executed as regression tests.