raptor

Why we included this project

Plain vector search retrieves the most similar chunks and assumes the answer sits inside one of them, which falls apart when a response has to combine material scattered across a long document. RAPTOR builds a tree instead: it clusters the text, summarizes each cluster, and repeats that step upward, so the same content appears at progressively broader levels of abstraction. A query then reads from several levels at once, which is exactly what you want when no single passage holds the whole answer. The package is easy to get into, index your documents, ask a question, get an answer, and the interfaces let you swap in your own summarization and embedding models, with the question-answering step pluggable as well. Advanced configuration docs are still marked as work in progress, and it needs an OpenAI API key to run. If your RAG pipeline keeps missing questions that draw on several sections of a corpus, RAPTOR is a solid option to test, grounded in the original research paper.