%% MemoryBear 在线检索时序图（Query Pipeline）
%% 起点：用户 Query；终点：LLM 生成的回答

sequenceDiagram
    autonumber
    participant User as 用户/API
    participant WF as Workflow Engine<br/>(workflow/nodes/knowledge/node.py)
    participant Config as config.py<br/>KnowledgeRetrievalNodeConfig
    participant Retriever as nlp/search.py<br/>knowledge_retrieval()
    participant Dealer as nlp/search.py:349<br/>Dealer.search()
    participant Qryr as nlp/query.py<br/>Query理解
    participant ESVec as ESVector<br/>elasticsearch_vector.py
    participant Graph as graphrag/search.py<br/>KGSearch.retrieval()
    participant Rerank as models/rerank.py<br/>RedBearRerank
    participant Prompt as prompts/generator.py
    participant LLM as llm/chat_model.py<br/>Base.chat()
    participant Cache as utils/redis_conn.py

    Note over User,Cache: === 阶段 1：Query 准备 ===
    User->>WF: 用户输入 Query
    WF->>WF: _render_template(query, variable_pool)
    WF->>Config: 读取 knowledge_bases[]<br/>reranker_id / retrieve_type

    Note over Retriever,ESVec: === 阶段 2：多知识库检索 ===
    loop 每个 Knowledge Base
        WF->>Retriever: knowledge_retrieval(query, config)
        Retriever->>DB: 验证 KB 状态 (chunk_num>0, status=1)

        alt RetrieveType == PARTICIPLE
            Retriever->>ESVec: search_by_full_text(query, top_k)
            ESVec->>ESVec: match on page_content (ik_max_word)
            ESVec-->>Retriever: List[DocumentChunk]

        else RetrieveType == SEMANTIC
            Retriever->>ESVec: search_by_vector(query, top_k)
            ESVec->>ESVec: script_score cosineSimilarity
            ESVec-->>Retriever: List[DocumentChunk]

        else RetrieveType == HYBRID
            par
                Retriever->>ESVec: search_by_vector()
                ESVec-->>Retriever: rs1
            and
                Retriever->>ESVec: search_by_full_text()
                ESVec-->>Retriever: rs2
            end
            Retriever->>Retriever: _deduplicate_docs(rs1, rs2)
            Retriever->>Rerank: rerank(query, docs, top_k)
            Rerank->>Rerank: similarity() 交叉编码评分
            Rerank-->>Retriever: sorted docs

        else RetrieveType == GRAPH
            par
                Retriever->>ESVec: search_by_vector()
                ESVec-->>Retriever: rs1
            and
                Retriever->>ESVec: search_by_full_text()
                ESVec-->>Retriever: rs2
            end
            Retriever->>Retriever: dedup + rerank

            Retriever->>Graph: kg_retriever.retrieval(question)
            Graph->>Graph: query_rewrite() → keywords + entities
            Graph->>ESVec: get_relevant_ents_by_keywords()
            Graph->>ESVec: get_relevant_relations_by_txt()
            Graph->>Graph: n_hop_with_weight 路径扩展
            Graph->>Graph: Score = pagerank * sim
            Graph->>Graph: _community_retrieval_()
            Graph-->>Retriever: Entity+Relation+CommunityReport chunk
            Retriever->>Retriever: insert(0, graph_result)
        end

        Retriever-->>WF: List[DocumentChunk]
    end

    WF->>WF: _deduplicate_docs(all_results)

    alt reranker_id 配置
        WF->>Rerank: rerank(query, all_results, reranker_top_k)
        Rerank-->>WF: reranked chunks
    end

    Note over Prompt,Cache: === 阶段 3：Prompt 组装 + LLM 生成 ===
    WF->>WF: 返回 {"chunks": [...], "citations": [...]}
    WF->>Prompt: citation_prompt(chunks)
    Prompt->>Prompt: 组装 System Prompt + 检索上下文

    Prompt->>Cache: get_llm_cache(model, prompt)
    alt cache miss
        Prompt->>LLM: chat(system, history, gen_conf)
        LLM-->>Prompt: answer, tokens
        Prompt->>Cache: set_llm_cache(model, prompt, answer)
    else cache hit
        Cache-->>Prompt: cached answer
    end

    Note over User,Cache: === 阶段 4：后处理 ===
    Prompt->>Dealer: insert_citations(answer, chunks, chunk_v)
    Dealer->>Dealer: pagerank*sim 定位引用位置
    Dealer-->>Prompt: answer_with_citations, cited_ids

    Prompt-->>User: 最终回答（含引用标注）