RAG Showcase - Seven Approach Flow Phases

Parallel lane view of ingestion, retrieval, augmentation, generation, and scoring behavior

0. Shared setup and corpus preparation 0.1 select Atlas profile 0.2 Atlas parse + chunk 0.3 Atlas embed + base 0.4 contextual post-step 0.5 LightRAG + drain 0.6 record job provenance 1. Entry 2. Retrieval plan 3. Evidence source 4. Augmentation 5. Generation 6. Output + metrics 7. Observed trait vanilla-rag hybrid-rag contextual-rag graph-rag agentic-rag n8n-adaptive-rag lazy-graph-rag 1.1 model=vanilla-ragOpenAI chat request 2.1 embed querynomic-embed-text 3.1 profile basedense top-k=5 4.1 stuff contexttop chunks only 5.1 one LLM calllight_gen role 6.1 answersources + metrics baselinefast/simple 1.1 model=hybrid-ragsame request shape 2.1 embed + textsemantic + BM25 3.1 profile basehybrid retrieve k=20 4.1 TEI reranktop_n=5 5.1 one LLM callstuff reranked hits 6.1 answerrerank cost in latency keyword IDspredictable 1.1 model=contextual-ragsame request shape 2.1 embed + textsemantic + BM25 3.1 profile contextualblurb-prefixed chunks 4.1 TEI reranktop_n=5 5.1 one LLM callcontext-aware hits 6.1 answercontext-grounded strong relevanceingest cost 1.1 model=graph-ragdelegating endpoint 2.1 LightRAG queryAtlas query profile 3.1 KG + vectorentities + relations 4.1 graph contextoptional TEI rerank 5.1 LightRAG LLMQUERY role: qwen3.8 6.1 answerKG source marker profile-sensitivefast when tuned 1.1 model=agentic-ragresearch-agent prompt 2.1 ReAct loopMAX_STEPS=4 3.1 toolsvectors or graph 4.1 observationstool trace builds up 5.1 final answeror MAX_STEPS stop 6.1 answer + traceLLM call count varies flexiblestep-limited 1.1 model=n8n-adaptive-ragthin webhook wrapper 2.1 classifysimple vs complex 3.1 routevanilla or agentic 4.1 selected pathCall Approach 5.1 workflow resultRespond to Webhook 6.1 answer + routeadaptive source marker operatorconfigurable 1.1 model=lazy-graph-ragOpenAI chat request 2.1 seed + conceptsdense top-k + terms 3.1 base + cacheconcept co-occurrence graph 4.1 budgeted expansiondepth + node/token caps 5.1 one LLM callexpanded evidence 6.1 answer + sourcescache metrics experimentallazy first build Shared model layer: LiteLLM routes to the configured Atlas provider; generation uses roles.yaml; LightRAG uses Atlas role inputs; embeddings default to nomic-embed-text.

1. Three direct retrievers

  • vanilla, hybrid, and contextual share the same answer_from_context generation path.
  • Atlas builds the profile base and graph; the showcase derives only contextual chunks.

2. Graph and agent paths

  • graph-rag delegates query answering to LightRAG hybrid mode.
  • agentic-rag loops over vector search and LightRAG tools, then returns a trace and final answer.
  • lazy-graph-rag builds and caches a deterministic concept graph on first use, then expands within query budgets.

3. Adaptive workflow

  • n8n-adaptive-rag is intentionally a workflow bridge, not a new retriever.
  • The Atlas-seeded, showcase-owned n8n graph classifies the query and calls a selected approach.

Each LiteLLM model alias maps to /rag/<approach>/v1 inside the Atlas backend container; Open WebUI and compare/run_matrix.py invoke those routes through LiteLLM /v1/chat/completions.