Hybrid RAG for SEC filings: leveraging LLMs and knowledge graphs for hallucination-free financial analysis
Year: 2026
Type: dissertation
Abstract: This thesis presents a hybrid Retrieval-Augmented Generation (RAG) system designed as an accessible and cost-effective solution to support investors and financial specialists in the study of large SEC filings. The system addresses the limitations of standard vector-based RAG architectures in the financial domain — loss of structural context and susceptibility to hallucination — by integrating structured graph traversal and semantic vector search within a unified Neo4j knowledge graph, allowing the system to return traceable and accurate financial data. The knowledge graph is built from six SEC filing types — N-CSR, N-PORT, 497K, 10-K, Form 4, and DEF 14A — extracted with the edgartools library across both fund- and company-level disclosures. Its schema was manually designed to capture the most relevant financial entities and relationships while preserving traceability to each source filing, in a structure that supports incremental ingestion of new filings. At query time, a SetFit classifier routes each query while schema slicing, dynamic fewshot retrieval, and entity resolver augment the context passed into a locally deployed text-to-Cypher model, which generates either a graph traversal or a vector search call. Generated queries are validated before execution, and the retrieved data is enriched with document provenance before a response generation model produces the final grounded answer, delivered through a Streamlit interface. Evaluation of the system demonstrates that each pipeline module contributes to the overall accuracy of the system, which is able to answer multi-hop entity queries beyond the reach of standard semantic search. The proposed system was evaluated against three alternative retrieval paradigms — a standard vector RAG baseline, web-augmented commercial language models, and an agentic pipeline with direct database access — showing consistent advantages in answer accuracy, source traceability, and security. The system operates on consumer-grade hardware with a locally deployed 7B-parameter model that delivered strong text-to-Cypher performance on the constructed benchmark despite its compact size, reducing operational costs compared with fully API-dependent alternatives, while an optional fallback to a larger API-hosted model is available for the most complex queries.
Author: Álvarez Conde, Luis
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213491045