Agentic RAG Document Intelligence Assistant
International Journal of Scientific Research in Science, Engineering and Technology · 12.11.2026 · 13 (6) · 50-62
DOI: 10.32628/ijsrset265414 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Scientific Research in Science, Engineering and Technology, online erschienen am 12.11.2026. Verfasst von Lokin V und 2 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
A response to a question regarding uploaded PDFs is only useful if supporting text can be found and examined. This paper presents the Agentic RAG Document Intelligence Assistant, where a Python controller evaluates retrieval results to either select more search, query rewriting, generation or abstention. The budget consists of three searches, with the eligibility evidence lasting for a given number of attempts. You extract the pages from PyMuPDF each at a time, then split yours recursively on 300 chars with a target overlap of 60 chars. The latter chunk are stored as local SentenceTransformer embeddings which, along with the text and metadata, is persistent in Chroma storage (384d vectors). Query rewriting + answer generation leverages openai/gpt-oss-120b via a compatible API. A cosine-similarity boundary of 0.45 (strong) and 0.15 (weak) can differentiate usable, weak and unusable matches but they need calendarisation to make them more suited for different datasets. To reduce unsupported responses, evidence admission, possible limited number of retries and generation constrained in context. Source identifiers link claims to filenames, and physical PDF pages. These choices are compared with adaptive retrieval, handling of corrective evidence, rewriting, document retrieval and citation assessment in a review of ten 2024-2026 studies. This contribution is not particularly an invention of an algorithm but rather a reproducible integration specification and evaluation protocol. Matched baselines, ablations, threshold analysis, citation checks and latency measurements are planned; no unmeasured outcome is reported as a result.
Zitieren
APA
V, L., L, T., & D, M. (2026). Agentic RAG Document Intelligence Assistant. International Journal of Scientific Research in Science, Engineering and Technology, 13(6), 50-62. https://doi.org/10.32628/ijsrset265414
Vancouver
V L, L T, D M. Agentic RAG Document Intelligence Assistant. International Journal of Scientific Research in Science, Engineering and Technology. 2026;13(6):50-62. doi:10.32628/ijsrset265414
BibTeX
@article{V2026,
author = {V, Lokin and L, Thara and D, Mohanapriya},
title = {Agentic RAG Document Intelligence Assistant},
journal = {International Journal of Scientific Research in Science, Engineering and Technology},
year = {2026},
volume = {13},
number = {6},
pages = {50--62},
doi = {10.32628/ijsrset265414}
}
EUVIMED