Psi-RAG v0.1.0

v0.1.0: The first public release of our $\Psi$-RAG.

✨Key Features

• 🌳 Corpus-Level Tree Index: Generalizes Tree-RAG from passage-level saplings to corpus-level large trees with millions of tokens. Abstration instead of named entity recognition: organize your corpus like a library with <3 hours indexing / >1 million tokens on two 48G RTX4090 GPUs!
• 🎯 Distribution-Adaptive Indexing: No need for explicit document structure. No need for searching for the optimal cluster numbers. No need for handpicking the initial point. No need for dimension reduction steps. No need for worrying your imbalanced data.
• ⚡ Efficient Indexing Techniques: Extra bucketing and HNSW support for sonic-speed similarity ranking on corpora with 10M+ tokens!
• 📚 Flexible Abstraction Mechanism: Choose one that you prefer: summaries or keywords!
• 👨‍👩‍👧‍👦 Multi-granular and Multifunctional Retrieval Pipeline: Iterative agentic retrieval empowers cross-document multi-hop tree search. Hybrid retrieval with BM25 navigates fine-grained tree search. Natural reranker support to maximize structured RAG performance.
• 🧑‍💻 Custom Framework Support: Built entirely with open-source LLMs.