文章背景与核心概要

在大型企业文档库中提取准确、扎根（Grounded）的答案一直是一项重大挑战，尤其是当查询涉及专业技术术语、特定供应商缩写或跨部分推理时。本文介绍了一种名为 DocuSearch 的离线多智能体文档智能系统，该系统专为生产级电信网络运维环境设计，旨在弥补这一技术空白。

DocuSearch 并没有依赖单一的检索向量，而是整合了三种互补的证据来源：基于 BGE-Large 嵌入的 Qdrant 向量检索、基于 SQLite FTS5 索引的 BM25 全文检索，以及利用结构化边表进行的知识图谱邻居扩展。系统通过倒数排名融合（RRF）算法对这些排序结果进行合并，并引入了交叉编码器（Cross-Encoder）重新排序以及最大边际相关性（MMR）来优化结果的相关性与多样性。

尤为重要的是，DocuSearch 引入了一种逐块（Per-chunk）评估循环，大语言模型（LLM）将每个分块视为一个微型检索问题，以判断上下文的充分性、查询完整性和文本扎根情况。在电信语料库上的评估表明，DocuSearch 的 Precision@10 达到了 0.69，Recall@10 达到了 0.79，扎根率高达 89.6%，分别比纯密集向量 RAG 基线高出 15、16 和 18.4 个百分点。

Hybrid Retrieval-Augmented Generation with Knowledge Graph Expansion, RRF Fusion, and Per-Chunk Grounded Evaluation for Enterprise Document Search

Authors: Harish Saragadam, Sudhanshu Sharma, Meghana Pujari
Submitted: 30 Jun 2026
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv: 2609.01617 [cs.IR]

📋 Summary

从大型企业文档库中提取准确、扎根的答案仍然是一个巨大的挑战，特别是当查询涉及技术术语、特定供应商的缩写或跨部门的推理时。DocuSearch 是一个离线的、多智能体的文档智能系统，专为生产电信网络运维环境设计，旨在弥补这一差距。

DocuSearch 没有依赖单一的检索向量，而是集成了三个互补的证据来源： 1. 语义搜索： 利用 BGE-Large 嵌入的 Qdrant 向量存储。 2. 全文搜索： 通过 SQLite FTS5 索引的 BM25。 3. 图扩展： 使用结构化边表的知识图谱邻居扩展。

这些排好序的结果通过倒数排名融合（Reciprocal Rank Fusion, RRF）进行合并，信号权重分别为 0.50（向量搜索）、0.35（BM25）和 0.15（知识图谱），并应用平滑常数 60 来稳定分数。随后，交叉编码器对融合后的列表进行重新排序，同时最大边际相关性（平衡因子 0.65）对结果进行裁剪，以实现最佳的相关性和多样性。

至关重要的是，DocuSearch 引入了一个逐块评估循环，其中 LLM 将每个分块视为一个微型检索问题，以确定上下文的充分性、查询完成度和文本扎根情况。未扎根的回答将被丢弃，取而代之的是多块回退合并。在电信语料库上评估时，DocuSearch 的 Precision@10 达到 0.69，Recall@10 达到 0.79，扎根率为 89.6%——分别比纯密集 RAG 基线高出 15、16 和 18.4 个百分点。

Extracting accurate, grounded answers from large enterprise document repositories remains a difficult challenge, especially when queries involve technical terminology, vendor-specific acronyms, or cross-sectional reasoning. DocuSearch is an offline, multi-agent document intelligence system designed for production telecom network operations environments to bridge this gap.

Rather than relying on a single retrieval vector, DocuSearch integrates three complementary evidence sources: 1. Semantic Search: Qdrant vector store utilizing BGE-Large embeddings. 2. Full-Text Search: BM25 via an SQLite FTS5 index. 3. Graph Expansion: Knowledge graph neighbor expansion using a structured edge table.

These ranked results are merged via Reciprocal Rank Fusion (RRF) with signal weights of 0.50 (vector search), 0.35 (BM25), and 0.15 (knowledge graph), applying a smoothing constant of 60 to stabilize scores. A cross-encoder then reranks the fused list, while Maximal Marginal Relevance (balance factor 0.65) prunes results for optimal relevance and diversity.

Crucially, DocuSearch introduces a per-chunk evaluation loop where an LLM treats each chunk as a mini-retrieval problem to determine contextual adequacy, query completion, and text grounding. Ungrounded responses are dropped in favor of a multi-chunk fallback merge. Evaluated on a telecom corpus, DocuSearch achieves a Precision@10 of 0.69, Recall@10 of 0.79, and a grounding rate of 89.6%—outperforming a dense-only RAG baseline by 15, 16, and 18.4 percentage points, respectively.

🔍 Abstract

从大型企业文档库中获取准确、扎根的答案是一个难题。对于混合了技术术语、特定供应商缩写或需要跨多个非相邻部分进行推理的查询，仅靠密集向量检索往往表现不佳。DocuSearch 的构建正是为了解决这一空白——这是一个在生产电信网络运维环境中开发和评估的离线多智能体文档智能系统。DocuSearch 没有依赖单一的检索信号，而是汇集了三个互补的证据来源：基于 BGE-Large 嵌入的 Qdrant 向量存储语义搜索、基于 SQLite FTS5 索引的 BM25 全文搜索，以及来自结构化边表的知识图谱邻居扩展。这三个排好序的列表通过倒数排名融合进行合并，向量搜索的信号权重为 0.50，BM25 为 0.35，知识图谱为 0.15，使用平滑常数 60 来稳定分数。然后，交叉编码器对融合后的列表进行重新排序，平衡因子为 0.65 的最大边际相关性用于裁剪结果以确保相关性和多样性。DocuSearch 的独特之处在于一个逐块评估循环，将每个块视为自己的微型检索问题：LLM 决定块是否需要更多上下文、是否完全回答了查询，以及答案是否扎根于检索到的文本中。未扎根的答案不会被返回；系统转而采用多块合并。在电信语料库上，DocuSearch 的 Precision@10 达到 0.69，Recall@10 达到 0.79，扎根率为 89.6%——比纯密集 RAG 基线分别提高了 15、16 和 18.4 个百分点。

Getting accurate, grounded answers out of large enterprise document repositories is a difficult problem. Dense vector retrieval alone frequently performs poorly on queries that mix technical terminology, vendor-specific acronyms, or require reasoning across several non-adjacent sections. DocuSearch was built to address exactly this gap - an offline, multi-agent document intelligence system developed and evaluated in a production telecom network operations environment. Rather than relying on a single retrieval signal, DocuSearch pulls together three complementary sources of evidence: semantic search over a Qdrant vector store using BGE-Large embeddings, BM25 full text search over an SQLite FTS5 index, and Knowledge Graph neighbour expansion from a structured edge table. These three ranked lists are merged through Reciprocal Rank Fusion with signal weights of 0.50 for vector search, 0.35 for BM25, and 0.15 for the knowledge graph, using a smoothing constant of 60 to stabilize scores. A cross-encoder then reranks the fused list, and Maximal Marginal Relevance with a balance factor of 0.65 prunes results for relevance and diversity. What makes DocuSearch distinctive is a per-chunk evaluation loop treating each chunk as its own mini-retrieval problem: an LLM decides whether the chunk needs more context, whether it fully answers the query, and whether the answer is grounded in retrieved text. Ungrounded answers are not returned; the system falls back to a multi-chunk merge instead. On a telecom corpus, DocuSearch reaches Precision@10 of 0.69, Recall@10 of 0.79, and a grounding rate of 89.6% - gains of 15, 16, and 18.4 percentage points over a dense-only RAG baseline.

🛠️ Key Technical Components

检索信号： Qdrant（稠密向量 / BGE-Large）+ SQLite FTS5（BM25 全文）+ 结构化边表（知识图谱）。
融合与重排： 带有加权参数的倒数排名融合（RRF） \(\rightarrow\) 交叉编码器重排 \(\rightarrow\) 最大边际相关性（MMR，\(\lambda = 0.65\)）。
评估框架： 使用 LangGraph 的智能体逐块扎根和验证循环。
领域上下文： 电信网络运维，本地企业环境。

Retrieval Signals: Qdrant (Dense Vectors / BGE-Large) + SQLite FTS5 (BM25 Full Text) + Structured Edge Table (Knowledge Graph).

Fusion & Reranking: Reciprocal Rank Fusion (RRF) with weighted parameters \(\rightarrow\) Cross-Encoder Reranking \(\rightarrow\) Maximal Marginal Relevance (MMR, \(\lambda = 0.65\)).

Evaluation Framework: Agentic per-chunk grounding and validation loop using LangGraph.

Domain Context: Telecom network operations, on-premise enterprise environments.