.zenodo.json

{
  "title": "rag-threat-intel: A sovereign RAG pipeline for vulnerability and threat-intelligence Q&A",
  "description": "An on-premise RAG (retrieval-augmented generation) pipeline that ingests NIST NVD CVE feeds and public security PDFs into pgvector, then answers natural-language questions using Ollama-served LLMs with mandatory source citations. Compares three chunking strategies (fixed-size 512+50 overlap, semantic heading-aware, sentence-window ±2) on real measured MRR@5 and MRR@10 over a 14-query CVE-recall subset, plus a three-axis faithfulness pilot (citation density / validity / refusal honesty) on a 12-query subsample, of a 50-question evaluation set. Embedding done with Ollama all-minilm; an embed_compare harness for nomic-embed-text vs all-minilm is included but its numbers are not in this release.",
  "creators": [
    {
      "name": "Bhutto, Ali Murtaza",
      "orcid": "0009-0007-2787-943X",
      "affiliation": "Independent Researcher"
    }
  ],
  "license": "MIT",
  "keywords": [
    "retrieval-augmented generation",
    "RAG",
    "threat intelligence",
    "sovereign AI",
    "pgvector",
    "NVD",
    "CVE",
    "chunking",
    "MRR",
    "faithfulness"
  ],
  "communities": [
    {"identifier": "ai-safety"}
  ],
  "related_identifiers": [
    {
      "identifier": "https://github.com/thunderstornX/rag-threat-intel",
      "relation": "isSupplementTo",
      "scheme": "url"
    }
  ],
  "upload_type": "software"
}