RAG Showcase seven approaches · one Atlas stack · reproducible evaluation OPENAI-COMPATIBLE · LOCAL-CAPABLE · PROVIDER-CONFIGURABLE 1 · ASK ONCE Open WebUI multi-model chat one prompt · parallel columns Evaluation harness dataset ladder base + flavor matrices same OpenAI request shape 2 · ATLAS ROUTING + RAG PLUGIN LiteLLM gateway model alias → backend route configured provider source FastAPI plugin seam seven OpenAI-compatible RAG endpoints base aliases + reproducible tuning flavors vanilla-rag hybrid-rag contextual-rag graph-rag agentic-rag n8n-adaptive lazy-graph-rag 3 · RETRIEVE + REASON Weaviate dense + BM25 chunks LightRAG + Neo4j persistent knowledge graph TEI reranker cross-encoder ordering n8n workflow classify → route Lazy graph cache dedicated JSON volume 4 · PERSIST + SCORE + COMPARE Uniform response envelope answer · available evidence · metrics Append-safe artifacts matrix · evidence · evaluation · judgments Ragas metrics coverage-aware context and answer scores Blinded judge panel + leaderboards dataset-aware rankings · sortable metrics Atlas owns the reusable stack · rag-showcase contributes approaches, tuning, datasets, and evaluation

One comparable entry

Open WebUI and the evaluation harness invoke the same LiteLLM aliases and OpenAI-compatible request contract.

Seven distinct retrieval paths

Vector, graph, agentic, adaptive-workflow, and lazy-graph behavior remains visible behind one shared Atlas lifecycle.

Persisted evaluation

Operational metrics, eligible Ragas evidence, and blinded judgments are stored before reports are generated.