Psi-RAG
An ICML 2026 hierarchical tree-based RAG framework that builds a corpus-level abstract tree index for cross-document retrieval, answering complex queries with a multi-granular agentic retrieval pipeline.
Psi-RAG builds a hierarchical abstract tree index over an entire corpus rather than only at the passage level, using abstraction instead of named-entity recognition so that a corpus of a million tokens can be organized like a library on two 48-gigabyte GPUs, with retrieval in logarithmic time. Indexing is distribution-adaptive: it does not require an explicit document structure, a preset number of clusters, a chosen initial point or dimensionality reduction, and it is designed to handle imbalanced data automatically.
Querying uses a multi-granular agentic retriever built around a reading-and-answering agent that runs a hybrid retrieval pipeline, combining iterative agentic tree search for cross-document, multi-hop questions with BM25-based hybrid retrieval for fine-grained tree search, and optional reranking. Extra bucketing and HNSW indexing keep similarity search fast on corpora above ten million tokens, and the abstraction stored at each node can be either generated summaries or extracted keywords. It is built entirely on open-source LLMs, so the backbone model can be swapped freely, and was presented at ICML 2026.