Rag.NET.Raptor 1.0.0

Rag.NET.Raptor

RAPTOR for Rag.NET — Recursive Abstractive Processing for Tree-Organized Retrieval: chunks are clustered (UMAP + GMM), each cluster is summarised, and the summaries are clustered again, producing a tree whose upper levels answer broad questions that leaf-only retrieval misses.

Install

dotnet add package Rag.NET.Raptor

Setup

RAPTOR adds one behavior to each pipeline, and UseRaptor places both. TreeScope defaults to Corpus — clustering over the whole corpus, not one document at a time — which needs somewhere to keep leaf chunks between ingests, so leafStorePath is required unless you opt back into PerDocument:

using Rag.NET.DependencyInjection;
using Rag.NET.Raptor;

services.AddRagNet(rag => rag.UseRaptor(leafStorePath: "raptor-leaves.db"));

RaptorIngestionBehavior lands directly after EmbeddingBehavior (it needs the embeddings) and RaptorRetrievalBehavior directly before RerankingBehavior (score adjustments settle before reranking sees them). To choose other positions, name them yourself — Add is idempotent and the pipeline delegates run first, so your placement wins and each behavior appears exactly once:

using Rag.NET.DependencyInjection;
using Rag.NET.Ingestion.Behaviors;
using Rag.NET.Raptor;
using Rag.NET.Retrieval.Behaviors;

services.AddRagNet(
    configure: rag => rag.UseRaptor(leafStorePath: "raptor-leaves.db"),
    ingestion: pipeline => pipeline
        .Add<RaptorIngestionBehavior>(after: typeof(EmbeddingBehavior)),
    retrieval: pipeline => pipeline
        .Add<RaptorRetrievalBehavior>(before: typeof(RerankingBehavior)));

Prefer isolated per-document trees instead — no cross-document summaries, no leaf store required:

services.AddRagNet(rag => rag.UseRaptor(o => o.TreeScope = RaptorTreeScope.PerDocument));

Example

Tune tree construction and how summaries compete with leaf chunks at query time:

rag.UseRaptor(
    options =>
    {
        options.MinChunksForRaptor = 5;      // skip small documents
        options.StoreLeafChunks    = true;   // keep originals alongside summaries
    },
    retrieval: options =>
    {
        options.Mode               = RaptorRetrievalMode.Boost;
        options.SummaryBoostFactor = 1.5;    // score multiplier for summaries
    },
    leafStorePath: "raptor-leaves.db");

RaptorRetrievalMode.Filter with MinRaptorLevel/MaxRaptorLevel restricts results to specific tree levels instead of boosting.

Full guide

• RAPTOR