Skip to content

Overview

RAG Showcase compares six modern retrieval-augmented-generation (RAG) approaches under identical conditions — same corpus, same embedding model, and the same generation model for the chunk-based approaches — so the dominant variable is the retrieval-and-reasoning approach itself. Two deliberate exceptions: graph-rag generates through LightRAG's QUERY role model, and n8n-adaptive-rag inherits the generator of whichever approach it routes to (see Evaluation Methodology §4).

1. How It Runs

Each approach is an OpenAI-compatible /<name>/v1/chat/completions endpoint in a self-contained plugin package (backend_plugins/rag/) that is bind-mounted into Atlas's FastAPI backend through a generic plugin seam. Each endpoint is registered into Atlas's LiteLLM gateway via its /model/new admin API, so the six approaches appear automatically as selectable models in Open WebUI.

flowchart LR
    U[Open WebUI<br/>multi-model chat] -->|/v1/chat/completions| L[LiteLLM gateway]
    L --> R[RAG plugin seam<br/>backend_plugins/rag]
    R --> V[vanilla-rag]
    R --> H[hybrid-rag]
    R --> C[contextual-rag]
    R --> G[graph-rag]
    R --> A[agentic-rag]
    R --> N[n8n-adaptive-rag]
    V & H & C & A --> W[(Weaviate<br/>+ TEI rerank)]
    G & A --> LR[(LightRAG<br/>knowledge graph)]
    N --> WF[n8n workflow<br/>classify → route]

Open a multi-model chat, select the approaches (or flavors) you want, and one prompt fans out — every answer comes back with a uniform answer, retrieved-context, and metrics footer so they are directly comparable.

2. Flavors

Named tuning flavors such as graph-rag-wide or hybrid-rag-high-recall can also appear as model aliases. They route to the same base approach with reproducible parameter overrides (retrieval depth, rerank on/off, graph query mode, agent step budget, …). One base approach can therefore be benchmarked at several operating points without code changes. See Flavor Tuning.

3. Fair-Comparison Guarantees

  • One corpus, ingested once into both a plain vector collection (RagBase) and a context-prefixed collection (RagContextual), plus a LightRAG knowledge graph.
  • Shared models — the chunk-based approaches generate through the same LiteLLM model and every approach embeds with the same embedding model (graph-rag's generator is LightRAG's QUERY role model by design); LLM roles are local-first (backend_plugins/rag/roles.yaml).
  • Uniform output contract — every approach returns the same answer/context/metrics shape, which the evaluation harness parses and scores.

4. Further Reading