Glossary

Definitions of key terminology for RAG / Codex that are not found on other pages:

Retrieval-Augmented Generation: A type of AI application in which an LLM is connected to your company’s data/documents, in order to accomplish tasks that require specific knowledge related to your company. To answer a user query, a RAG system retrieves relevant context from the documents/data, and then the LLM generates a response based on this context and the query. When the LLM is able to call Tools, such applications are sometimes called AI Agents.

Tool Calling: Allowing LLMs to call external APIs like databases, calculators, or the internet. Instead of generating a response directly, the LLM can optionally choose to call the API with arguments generated by the model. The model then generates its response for the user based on the result from the API.

Agentic RAG: In standard RAG, retrieval is performed for every user query before any LLM calls. In agentic RAG, retrieval is treated as a Tool which the LLM can choose to call whenever (or skip if the model already knows an answer).

Knowledge Base: The data sources / documents from which context can be retrieved by a RAG application. This effectively forms the primary knowledge of this AI system. Often the Knowledge Base is a complex Vector Database filled with document chunks and metadata.

IDK response: When the RAG application unhelpfully responds to a user query along the lines of “I don’t know” or “I cannot answer”. This may be due to a Knowledge Gap where the required information is missing from the Knowledge Base. Or it may be due to an ineffective retrieval/search system, or due to the LLM not being intelligent enough to handle a complex request. To reduce LLM hallucinations: it’s common practice to output a hardcoded fallback response for this, whenever your AI is unsure how to correctly answer a query.

Bad response: Outputs from your RAG application that are either: incorrect, IDK responses, or otherwise unhelpful. Codex enables nontechnical SMEs (Subject Matter Experts) at your company to steadily reduce bad responses.

Semantically identical query: A user query which is similar enough to another query that one answer is suitable for both queries. Abbreviated as similar query for short. Codex is more effective for RAG applications that receive more such queries.

Semantic distance: Measures how different two queries are in terms of their underlying meaning, rather than how they happen to be worded. In Codex, this is estimated via a vector distance between neural network embeddings of each query. Queries are considered semantically identical when the distance between their embeddings lies below a small threshold value.

Detect, Prioritize, Resolve, Prevent: Four forms of automation provided by the Codex platform. Codex detects issues in your RAG responses (bad responses). Codex prioritizes which issues SMEs should focus on for maximal ROI. Codex resolves issues via SME-provided answers and task tracking. Codex prevents these issues from occuring in the future.