Grounding Language Model with Chunking-Free In-Context Retrieval