Universal Doc AI: A Universal Intelligent Document Management System Combining OCR, RAG, and LLMs

It proposes a unified AI document management framework that handles both textual and image-based documents, covering PDFs, scanned documents, Word files, text files, and image formats such as JPG, JPEG, and PNG. Whereas prior work tends to address a single stage such as OCR, retrieval, or conversation, this work places document acquisition, storage, organization, and question answering within one architecture. Presented as a system architecture and a six-phase workflow (document acquisition, text extraction, preprocessing, semantic indexing, retrieval, response generation), i.e., a design-level argument.

It connects OCR into the RAG pipeline, applying noise removal, orientation correction, text-region detection, and character recognition to scans and images before semantic processing. Image-based content becomes searchable text, bringing scans and photographs that keyword search could not reach into the scope of semantic retrieval. Given as a stepwise description in Algorithm 1 (OCR text extraction); the text notes handwriting recognition applies "where supported".

It uses overlapping semantic chunking, dense vector embeddings, and vector-database indexing, with cosine-similarity Top-K semantic retrieval instead of exact keyword matching. Retrieval is driven by semantic relevance rather than literal matching, and document identifiers, metadata, page numbers, and chunk references are retained for traceability. Given as stepwise descriptions in Algorithms 2, 3, 4, and 6, i.e., method description rather than comparative experimental data.

An LLM generates answers from the retrieved context and returns document references; when context is insufficient the system informs the user instead of producing speculative responses, which is how hallucination is reduced. Answers are restricted to the knowledge in the uploaded documents and carry verifiable source references, distinguishing the system from general chatbots that rely only on pretrained knowledge. Given as a stepwise description in Algorithm 5 (context-aware generation) plus qualitative statements in the conclusion.

The framework targets institutional settings that need question answering and semantic retrieval within uploaded documents, including educational institutions, enterprises, healthcare organizations, law firms, financial institutions, and government departments; it is designed to support PDFs, scans, images, and text files alongside secure storage, folder organization, metadata management, version tracking, user authentication, and role-based sharing permissions. The text lists multilingual document understanding, handwriting recognition, document summarization, voice interaction, and multimodal reasoning as future work, indicating these capabilities lie in the intended extension scope rather than the current version.

The loaded text is a version without figures or experimental tables, so the specific metrics behind "retrieval accuracy, response relevance, and user interaction efficiency," the dataset composition, sample size, and comparison baselines cannot be confirmed from the available content; the literature review cites that retrieval accuracy decreases as OCR noise increases, yet how this system behaves under such conditions remains an open question; handwriting recognition is qualified as "where supported," so its coverage is unclear; multilingual understanding, summarization, and voice interaction are listed as future directions, so their current availability is unverified.