For a docs chatbot, grounding means the model answers from retrieved documentation passages rather than from its training data. The retrieved text is placed in the model's context with instructions to answer only from it.
Grounded systems should link to their source material and decline questions that the available evidence does not support. An assistant that never declines is more likely to answer beyond the retrieved evidence.
Grounding quality depends on retrieval quality. If chunking or indexing misses the relevant section, the model either refuses or answers from a partially relevant passage. Most grounding work is really retrieval work.