The Essential RAG Book
Context-Aware RAG
Context-Aware Retrieval-Augmented Generation (RAG) introduces adaptive mechanisms that leverage conversational or multi-turn context to reformulate user queries before retrieval. Unlike the baseline pipeline, which treats each input independently, context-aware architectures m...
TL;DR
Context-Aware Retrieval-Augmented Generation (RAG) introduces adaptive mechanisms that leverage conversational or multi-turn context to reformulate user queries before retrieval. Unlike the baseline pipeline, which treats each input independently, context-aware architectures maintain and evolve a running state repre...
Key Takeaways
- Context-Aware Retrieval-Augmented Generation (RAG) introduces adaptive mechanisms that leverage conversational or multi-turn context to reformulate user queries before retrieval.
- The key innovation is the query rewriter, a transformer sub-module trained to compress dialogue history into a concise, self-contained question.
Context-Aware Retrieval-Augmented Generation (RAG) introduces adaptive mechanisms that leverage conversational or multi-turn context to reformulate user queries before retrieval. Unlike the baseline pipeline, which treats each input independently, context-aware architectures maintain and evolve a running state representation. [Conversation History] → [Query Rewriter] → [Retriever] ↓ [Generator]
Figure 4 - Context-Aware RAG introduces dynamic query rewriting.
The key innovation is the query rewriter, a transformer sub-module trained to compress dialogue history into a concise, self-contained question. This rewritten query is passed to the retriever, which then accesses relevant knowledge. The generator conditions on both the retrieved content and latent state embeddings derived from past turns. Context tracking can be implemented using sliding-window encoders, hierarchical attention, or memory tokens. Systems such as ChatGPT-RAG and MemoryGPT employ this design to enable continuity and reasoning across multiple turns without exceeding token limits. Evaluation of Context-Aware RAG often uses metrics like Contextual Recall and Dialogue Faithfulness. These measure how effectively the model integrates prior turns and preserves conversational coherence.
