Lesson 5: Under the Hood
Grounding
By default, most LLMs only know what they learned during training which has a cutoff date and no knowledge of current events, private documents, or anything that happened after training ended.
Grounding is what fixes this by connecting the model to real, external information before it generates a response.
Some chatbots can use web search in real time to incorporate other information.
Another way is for users to upload PDFs, spreadsheets, or text files directly into a conversation. ChatGPT reads the document and grounds its responses in the actual content of that file rather than guessing from training data.