DOCMIND AI

DOC MIND implements a local document-review application that connects validated intake, canonical extraction, configured domain audits and human decision records. Its principal objective is not simply to generate more text about a file, but to make the basis and limits of a review inspectable. Supported documents receive explainable findings and conditional health scores; unsupported or insufficiently assessed inputs retain explicit uncertainty. The architecture combines Next.js and React with FastAPI, SQLAlchemy, SQLite, extraction libraries and optional Ollama generation. These are established technologies; the project contribution lies in their orchestration with a check ledger, audit snapshots, evidence identities and conservative revision states. Regression and acceptance results demonstrate concrete implementation behaviours including tenant isolation, role-aware controls, persistence and encrypted fixture recovery. The corrected synthetic cases provide useful diagnostics for currency handling and clause negation. However, successful tests do not prove complete legal understanding, financial authenticity, recruitment effectiveness or independently established real-document accuracy. The system remains an assistive workflow with human responsibility for consequential decisions. The integrated Finding to Evidence to Decision to Revision to Comparison workflow is the project's strongest practical distinction. It allows a reviewer to investigate a concern, record a response and check a related revision without treating missing analysis as success. Further evaluation should extend this traceability to consented real documents, original-page evidence visualization and carefully measured long-running local operation.

[1] P. Lewis et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems, 33, 2020, pp. 9459-9474. https://arxiv.org/abs/2005.11401 [2] N. Reimers and I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP-IJCNLP, 2019, pp. 3982-3992. https://doi.org/10.18653/v1/D19-1410 [3] J. Johnson, M. Douze and H. Jegou. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3), 2021, pp. 535-547; preprint 2017. https://arxiv.org/abs/1702.08734 [4] Tesseract OCR contributors. Tesseract User Manual. Official documentation. https://tesseract-ocr.github.io/tessdoc/ [5] Microsoft. Document layout analysis with Document Intelligence. Official documentation. https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/prebuilt/layout [6] Adobe. Use AI Assistant in Acrobat. Official product documentation. https://helpx.adobe.com/acrobat/desktop/use-acrobat-ai/generative-ai-features/ai-get-answers.html [7] Qwen Team. Qwen3 Think Deeper Act Faster. Model release and local inference overview, 2025. https://qwenlm.github.io/blog/qwen3/ [8] Ollama. API Introduction. Official local inference API documentation. https://docs.ollama.com/api/introduction [9] FastAPI contributors. FastAPI Documentation. Official API framework documentation. https://fastapi.tiangolo.com/ [10] Vercel. Next.js Documentation. Official web framework documentation. https://nextjs.org/docs