Complex Subgraphs¶

complex-subgraphs demonstrates a coordinator graph with specialist subgraphs and one nested grandchild graph. The OpenAI request adapter supplies the final user message as question. The keyword graph keeps its selected terms in graph state and emits an optional status update. The selected specialist writes the final assistant message to the shared messages channel.

LangGraph Topology¶

The native xray view expands the specialist and keyword subgraphs.

graph TD;
  __start__ --> route_question;
  route_question -.-> api_collect["api_contract_graph:collect_contract_checks"];
  route_question -.-> docs_extract["docs_graph:keyword_graph:extract_keywords"];
  api_summary["api_contract_graph:summarize_contract"] --> __end__;
  docs_summary["docs_graph:summarize_docs"] --> __end__;
  subgraph api_contract_graph
    api_collect --> api_summary;
  end
  subgraph docs_graph
    docs_prepare["docs_graph:keyword_graph:prepare_keyword_context"] --> docs_summary;
    subgraph keyword_graph
      docs_extract --> docs_prepare;
    end
  end

Request Flow¶

flowchart TD
  request([OpenAI user message]) --> adapter["request_to_input"]
  adapter --> route["Normalize question and select route"]

  subgraph api["API contract subgraph"]
    checks["Collect contract checks"] --> api_summary["Stream contract summary"]
  end

  subgraph docs["Docs specialist subgraph"]
    subgraph keywords["Keyword grandchild subgraph"]
      extract["Extract keywords"] --> prepare["Build context and emit status"]
    end
    prepare --> docs_summary["Stream docs summary"]
  end

  route -->|"API, OpenAI, adapter, stream, or serve"| checks
  route -->|"other questions"| extract
  api_summary --> output["final assistant message"]
  docs_summary --> output
  output --> response([OpenAI assistant text])

LGOS exposes the answer-producing nested nodes summarize_contract and summarize_docs as assistant text. The keyword node's selected terms are available as graph state and, for clients that opt into client_events, as a status update. The final assistant message is identical for streaming and non-streaming requests. Rich status delivery also requires a streaming request with metadata.langgraph_stream_events="v1".

The graph is deterministic and has no checkpointer or Store. Its routing, keywords, checks, and messages exist only for the current request. Any conversation history must come from messages supplied by the client.

Try It¶