A Comprehensive Review on Synapse: A GraphRAG-Based Context-Aware Research Assistant

Abstract

Research assistants built on Retrieval-Augmented Generation (RAG) usually depend on vector similarity, which struggles with multi-document synthesis, multi-hop questions, and fragmented context. This review examines how Graph Retrieval-Augmented Generation (GraphRAG) closes these gaps by pairing embedding-based search with knowledge graph traversal, and proposes Synapse, a context-aware research assistant built on that idea. Synapse parses scholarly documents into structure-preserving chunks, indexes them both as vectors and as an entity–relation graph, groups the graph into thematic communities, and produces answers with inline, clickable citations to source passages. We survey recent work on RAG, knowledge graph construction, and trustworthy generation, describe the architecture and its key algorithms, and discuss practical challenges such as indexing cost and entity duplication. The review indicates that hybrid retrieval is a promising route to traceable and reliable research assistance.

Introduction

The text presents Synapse, an AI-powered research assistant designed to help researchers search, understand, and analyze large collections of academic documents using natural-language questions. As scholarly publications continue to grow, manually reading and analyzing research papers becomes increasingly difficult. Although Large Language Models (LLMs) can answer questions about documents, they may generate unsupported or inaccurate information when they lack relevant evidence.

To address this problem, Synapse uses Retrieval-Augmented Generation (RAG) combined with knowledge graphs, an approach known as GraphRAG. This combination enables the system to retrieve relevant text, identify relationships between research concepts, and generate answers supported by source citations.

1. Background and Literature Review

Traditional RAG systems divide documents into text chunks, convert them into numerical embeddings, and retrieve passages that are semantically similar to a user's query. However, this approach has several limitations:

• Important information may be separated across different text chunks.

• Relationships between authors, methods, datasets, and research findings are not explicitly represented.

• Questions about overall research trends and themes are difficult to answer using only a few retrieved passages.

Knowledge graphs address these limitations by representing information as connected entities and relationships. GraphRAG combines knowledge graphs with vector-based retrieval, allowing the system to use both the original text and the relationships between concepts.

2. Proposed System Architecture

Synapse uses a five-layer architecture.

1. User and Application Layer: Provides document uploads, authentication, a chat interface, interactive graph visualizations, and clickable citations.

2. Document Ingestion and Processing: Accepts PDF, DOCX, and PPTX files, converts them into Markdown, and divides them into meaningful sections while preserving page numbers and document metadata.

3. Dual Indexing: Stores document chunks in a vector database and extracts entities and relationships into a knowledge graph.

4. Hybrid Retrieval Engine: Combines semantic search, graph-based expansion, and relevance ranking to retrieve useful evidence.

5. Grounded Answer Generation: Uses an LLM to generate responses based on retrieved evidence and links individual claims to their supporting passages.

The system also provides separate workspaces to protect users' confidential documents from access by other users.

3. Implementation Methodology

Synapse follows several important processing steps:

• Document parsing and chunking: Converts different document formats into a consistent structure and preserves section and page information.

• Entity extraction and normalization: Identifies research concepts, authors, methods, datasets, and organizations while merging duplicate entities and alternative names.

• Knowledge graph construction: Connects extracted entities through relationships that represent facts from the documents.

• Community detection and summarization: Groups related entities into research themes and creates summaries to support questions about the entire document collection.

• Hybrid retrieval and generation: Combines relevant text passages, graph relationships, and community summaries to produce evidence-based answers.

For example, a question about which dataset was used to evaluate a particular method can be answered using relevant passages and graph connections. A broader question about the main research directions in a collection can be answered using summaries of related research communities.

4. Technology Stack

The proposed system uses several technologies:

• Python and FastAPI: Backend processing and API management.

• LangChain and LlamaIndex: Coordination of retrieval, extraction, and generation workflows.

• Qdrant, FAISS, or Chroma: Storage and retrieval of vector embeddings.

• Neo4j and Cypher: Knowledge graph storage and querying.

• Hosted LLMs: Entity extraction, summarization, and answer generation.

• JavaScript-based frontend: Interactive graph visualization and citation navigation.

5. Research Gaps and Challenges

The study identifies four major challenges in existing systems:

• The computational cost of building and maintaining knowledge graphs.

• Errors, missing relationships, and duplicate entities in automatically generated graphs.

• The lack of standardized methods for evaluating retrieval quality and answer accuracy.

• Difficulty tracing generated claims directly to their original sources.

Synapse aims to address these issues through entity normalization, structured document processing, hybrid retrieval, and citation-level traceability. However, citations and grounding can reduce unsupported answers without completely eliminating errors.

Conclusion

This review examined how pairing vector retrieval with knowledge graph traversal addresses the limits of standard RAG for scholarly question answering, and described Synapse, our layered design that preserves document structure, indexes content as both vectors and a graph, summarizes communities for corpus-level questions, and ties each claim to a source passage. The case for the design rests on the surveyed literature. We have not yet evaluated Synapse, so its benefits remain expected rather than measured. Future work includes benchmarking against vector-only baselines with established RAG evaluation protocols [8], incremental graph updates that avoid full rebuilds when documents are added, multimodal parsing of figures and tables [20], and agentic retrieval routing that lets the system plan its own search steps [9].

References

[1] Y. Gao et al., “Retrieval-augmented generation for large language models: A survey,” arXiv preprint arXiv:2312.10997, 2023, doi: 10.48550/arXiv.2312.10997. [Online]. Available: https://arxiv.org/abs/2312.10997. [2] B. Peng et al., “Graph retrieval-augmented generation: A survey,” ACM Trans. Inf. Syst., vol. 44, no. 2, pp. 1–52, 2026, doi: 10.1145/3777378. [3] Q. Zhang et al., “A survey of graph retrieval-augmented generation for customized large language models,” arXiv preprint arXiv:2501.13958, 2025, doi: 10.48550/arXiv.2501.13958. [Online]. Available: https://arxiv.org/abs/2501.13958. [4] P. Jiang, S. Ouyang, Y. Jiao, M. Zhong, R. Tian, and J. Han, “A survey on retrieval and structuring augmented generation with large language models,” in Proc. 31st ACM SIGKDD Conf. Knowledge Discovery and Data Mining, 2025, doi: 10.1145/3711896.3736557. [5] Y. Huang and J. X. Huang, “A survey on retrieval-augmented text generation for large language models,” ACM Comput. Surv., vol. 58, no. 12, Art. no. 300, pp. 1–38, 2026, doi: 10.1145/3805774. [6] C. Sharma, “Retrieval-augmented generation: A comprehensive survey of architectures, enhancements, and robustness frontiers,” arXiv preprint arXiv:2506.00054, 2025, doi: 10.48550/arXiv.2506.00054. [Online]. Available: https://arxiv.org/abs/2506.00054. [7] B. Ni et al., “Towards trustworthy retrieval augmented generation for large language models: A survey,” arXiv preprint arXiv:2502.06872, 2025, doi: 10.48550/arXiv.2502.06872. [Online]. Available: https://arxiv.org/abs/2502.06872. [8] A. Gan et al., “Retrieval augmented generation evaluation in the era of large language models: A comprehensive survey,” arXiv preprint arXiv:2504.14891, 2025, doi: 10.48550/arXiv.2504.14891. [Online]. Available: https://arxiv.org/abs/2504.14891. [9] A. Singh et al., “Agentic retrieval-augmented generation: A survey on agentic RAG,” arXiv preprint arXiv:2501.09136, 2025, doi: 10.48550/arXiv.2501.09136. [Online]. Available: https://arxiv.org/abs/2501.09136. [10] A. Hogan et al., “Knowledge graphs,” ACM Comput. Surv., vol. 54, no. 4, Art. no. 71, pp. 1–37, 2021, doi: 10.1145/3447772. [11] A. Khan, “Knowledge graphs querying,” ACM SIGMOD Rec., vol. 52, no. 1, pp. 6–17, 2023, doi: 10.1145/3604437.3604439. [12] L. Zhong, J. Wu, Q. Li, H. Peng, and X. Wu, “A comprehensive survey on automatic knowledge graph construction,” ACM Comput. Surv., vol. 56, no. 4, Art. no. 94, pp. 1–62, 2024, doi: 10.1145/3618295. [13] H. Bian, “LLM-empowered knowledge graph construction: A survey,” arXiv preprint arXiv:2510.20345, 2025, doi: 10.48550/arXiv.2510.20345. [Online]. Available: https://arxiv.org/abs/2510.20345. [14] X. Jiang et al., “On the evolution of knowledge graphs: A survey and perspective,” arXiv preprint arXiv:2310.04835, 2023, doi: 10.48550/arXiv.2310.04835. [Online]. Available: https://arxiv.org/abs/2310.04835. [15] B. Subagdja, D. Shanthoshigaa, Z. Wang, and A.-H. Tan, “Machine learning for refining knowledge graphs: A survey,” ACM Comput. Surv., vol. 56, no. 6, Art. no. 156, pp. 1–38, 2024, doi: 10.1145/3640313. [16] B. Xue and L. Zou, “Knowledge graph quality management: A comprehensive survey,” IEEE Trans. Knowl. Data Eng., vol. 35, no. 5, pp. 4969–4988, May 2023, doi: 10.1109/TKDE.2022.3150080. [17] V. Mavi, A. Jangra, and A. Jatowt, “Multi-hop question answering,” Found. Trends Inf. Retr., vol. 17, no. 5, pp. 457–586, 2024, doi: 10.1561/1500000102. [18] L. Liu, Z. Wang, and H. Tong, “Neural-symbolic reasoning over knowledge graphs: A survey from a query perspective,” arXiv preprint arXiv:2412.10390, 2024, doi: 10.48550/arXiv.2412.10390. [Online]. Available: https://arxiv.org/abs/2412.10390. [19] C. Ma, Y. Chen, T. Wu, A. Khan, and H. Wang, “Large language models meet knowledge graphs for question answering: Synthesis and opportunities,” in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), Nov. 2025, pp. 24578–24597, doi: 10.18653/v1/2025.emnlp-main.1249. [20] Q. Zhang, B. Wang, V. S.-J. Huang, J. Zhang, Z. Wang, H. Liang, C. He, and W. Zhang, “Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction,” arXiv preprint arXiv:2410.21169, 2024, doi: 10.48550/arXiv.2410.21169. [Online]. Available: https://arxiv.org/abs/2410.21169. [21] N. Subramani et al., “A survey of deep learning approaches for OCR and document understanding,” arXiv preprint arXiv:2011.13534, 2020, doi: 10.48550/arXiv.2011.13534. [Online]. Available: https://arxiv.org/abs/2011.13534. [22] Y. Zhou et al., “Trustworthiness in retrieval-augmented generation systems: A survey,” arXiv preprint arXiv:2409.10102, 2024, doi: 10.48550/arXiv.2409.10102. [Online]. Available: https://arxiv.org/abs/2409.10102. [23] D. Edge et al., “From local to global: A graph RAG approach to query-focused summarization,” arXiv preprint arXiv:2404.16130, 2024, doi: 10.48550/arXiv.2404.16130. [Online]. Available: https://arxiv.org/abs/2404.16130.

Copyright

Copyright © 2026 Prof. Priti Tijare, Prof. Aparna Khairkar, Bhushan Dhok, Krushna Chinchamalatpure, Abhas Gawali, Devang Dhote. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.