Skip to content

How does RAPTOR perform hierarchical tree-organized retrieval?

Advanced · 2 min read · RAG & Vector Databases

Open in study mode to save this question and track progress

Short answer: RAPTOR (Recursive Abstractive Processing for Tree-Organized Retrieval) recursively clusters text chunks, generates abstractive LLM summaries for each cluster, and indexes both raw chunks and multi-level summary nodes into a hierarchical tree structure, enabling RAG systems to answer high-level holistic questions alongside low-level detail queries.

Detail

Standard vector retrieval indexes isolated text chunks, failing on high-level thematic queries (e.g. "What are the key overarching themes across the entire 300-page report?").

Level 2 (Global Summary):                  [ Root Summary Node ]
                                            /                  \
Level 1 (Cluster Summaries):     [ Cluster Summary A ]    [ Cluster Summary B ]
                                  /         |         \     /         \
Level 0 (Raw Text Chunks):   [Chunk 1]  [Chunk 2]  [Chunk 3] [Chunk 4] [Chunk 5]

RAPTOR Tree Construction Algorithm

  1. Leaf Chunks (Level 0): Segment document into standard text chunks (e.g. 100 tokens).
  2. Soft Clustering: Cluster chunk embeddings using Gaussian Mixture Models (GMMs) or UMAP dimensionality reduction.
  3. Summarization: Use an LLM to generate a summary for each cluster.
  4. Recursive Indexing: Embed summary nodes and repeat clustering recursively until a root summary is reached.
  5. Collapsed Tree Retrieval: Search across all levels (leaf chunks + multi-level summaries) simultaneously during query execution.

Example

Conceptual Python data model for a RAPTOR hierarchical node:

python
class RAPTORNode:
    def __init__(self, node_id: str, text: str, level: int, children: list[str] = None):
        self.node_id = node_id
        self.text = text
        self.level = level # Level 0 = raw chunk, Level 1+ = summary
        self.children = children or []

# Level 1 summary node linking to 2 raw Level 0 leaf chunks
summary_node = RAPTORNode(
    node_id="summary_lvl1_04",
    text="Executive summary of Q3 financial performance across departments...",
    level=1,
    children=["chunk_001", "chunk_002"]
)

Interview tips

  • Contrast RAPTOR with GraphRAG: RAPTOR builds hierarchical text summary trees, while GraphRAG extracts entity-relationship knowledge graphs.
  • Highlight RAPTOR's benchmark performance gains on long-context QA benchmarks (QASPER, QuALITY).

Related Concepts

⬅ Back to RAG and Vector Databases · All topics