Vector Search As Nearest Neighbor Matching: RAG-based Policy Learning in Causal Inference
The paper frames action-specific vector search in RAG as nearest-neighbor matching for policy learning.
Kato and Kato propose one-step and two-step RAG methods under the potential outcome framework. In the two-step version, retrieval gathers neighboring evidence by action in embedding space, then a generator estimates outcomes or contrasts before a plug-in rule chooses the action. The paper decomposes regret into candidate-generation regret and within-candidate choice regret, with bounds tied to nearest-neighbor and transformer prediction error. The one-step method is evaluated directly as a policy because its internal computation is not observed. ArXiv · AI/CL/LG's note
Kato and Kato propose one-step and two-step RAG methods under the potential outcome framework. In the two-step version, retrieval gathers neighboring evidence by action in embedding space, then a generator estimates outcomes or contrasts before a plug-in rule chooses the action. The paper decomposes regret into candidate-generation regret and within-candidate choice regret, with bounds tied to nearest-neighbor and transformer prediction error. The one-step method is evaluated directly as a policy because its internal computation is not observed. ArXiv · AI/CL/LG's note
score 4