raptor-rag

RAPTOR RAG: Hierarchical Summarization Tree for Multi-Level Retrieval

Flat chunk retrieval misses the big picture. RAPTOR builds a tree: raw chunks at the bottom, leaf summaries grouping related chunks in the middle, cluster summaries at the top. The tree search retrieves context at all 3 levels, giving the LLM both granular details and high-level themes.

Workflow

documentText, question
       │
       ▼
┌──────────────────────┐
│ rp_chunk_docs        │  Split into raw chunks
└───────────┬──────────┘
            ▼
┌──────────────────────┐
│ rp_leaf_summaries    │  Summarize groups of chunks
└───────────┬──────────┘
            ▼
┌──────────────────────┐
│ rp_cluster_summaries │  Summarize groups of leaves
└───────────┬──────────┘
            ▼
┌──────────────────────┐
│ rp_tree_search       │  Search all 3 tree levels
└───────────┬──────────┘
            ▼
┌──────────────────────┐
│ rp_generate          │  Generate from multi-level context
└──────────────────────┘

Workers

ChunkDocsWorker (rp_chunk_docs) -- Returns raw chunks: {id: "chunk-1", text: "..."}, {id: "chunk-2", text: "..."}.

LeafSummariesWorker (rp_leaf_summaries) -- Produces leaf-level summaries with chunkIds: ["chunk-1", "chunk-2"] linking back to source chunks.

ClusterSummariesWorker (rp_cluster_summaries) -- Produces cluster summaries with leafIds: ["leaf-1", "leaf-2"] linking to leaves.

TreeSearchWorker (rp_tree_search) -- Returns context at 3 levels: root (broadest summary), cluster (mid-level), and leaf (most specific), each with relevance scores.

GenerateWorker (rp_generate) -- Generates from the multi-level context using system prompt.

Tests

32 tests cover chunk creation, leaf summarization, cluster summarization, tree search, and generation.

Further Reading

• RUNNING.md -- how to build and run this example