RAPTOR
A hierarchical retrieval framework proposed by a Stanford team. It recursively clusters text chunks and invokes an LLM to generate summaries, constructing a bottom-up summary tree to improve question-answering performance over long documents.
Timeline (last 90 days)
RAPTOR builds a tree: it clusters related text chunks, generates a summary for each cluster, then clusters and summarizes the summaries, all the way up; at query time it retrieves across levels, capturing both fine-grained details and high-level synthesis.
The paper reports that RAPTOR, paired with GPT-4 on the QuALITY benchmark (long-document question answering), improved absolute accuracy by about 20 points, reaching state of the art at the time.
RAPTOR is a hierarchical retrieval framework proposed by the Stanford team in 2024, whose full name is Recursive Abstractive Processing for Tree-Organized Retrieval.
RAPTOR works by splitting documents into text chunks, vectorizing them, performing soft clustering using Gaussian Mixture Models, calling a language model to generate abstract summaries, and recursively constructing a summary tree.
RAPTOR outperforms flat RAG on multiple long-document question-answering benchmarks, but indexing costs grow superlinearly with document scale, and summary quality is highly dependent on the capability of the language model used.
The generative RAPTOR variant, even equipped with high-quality clustering summaries, still loses to NavTree—which has zero index cost—at every multi-chunk budget tier.
All Facts (6)
RAPTOR builds a tree: it clusters related text chunks, generates a summary for each cluster, then clusters and summarizes the summaries, all the way up; at query time it retrieves across levels, capturing both fine-grained details and high-level synthesis.
50%UnverifiedThe paper reports that RAPTOR, paired with GPT-4 on the QuALITY benchmark (long-document question answering), improved absolute accuracy by about 20 points, reaching state of the art at the time.
50%UnverifiedRAPTOR is a hierarchical retrieval framework proposed by the Stanford team in 2024, whose full name is Recursive Abstractive Processing for Tree-Organized Retrieval.
50%UnverifiedRAPTOR works by splitting documents into text chunks, vectorizing them, performing soft clustering using Gaussian Mixture Models, calling a language model to generate abstract summaries, and recursively constructing a summary tree.
50%UnverifiedRAPTOR outperforms flat RAG on multiple long-document question-answering benchmarks, but indexing costs grow superlinearly with document scale, and summary quality is highly dependent on the capability of the language model used.
50%UnverifiedThe generative RAPTOR variant, even equipped with high-quality clustering summaries, still loses to NavTree—which has zero index cost—at every multi-chunk budget tier.
50%