Publication: LLM in OSINT — RAG pipeline architecture, quality evaluation, AgentD22 Data solutions cases
OSINT-RAG v3.2 architecture:
Sources → NiFi → cleaning → semantic chunking (512 tokens) → BGE-M3 embeddings → Qdrant HNSW → hybrid search (BM25 plus dense) → Cohere Rerank v3.5 → Llama-3.1-70B or GPT-4o → citation verification (DeBERTa-v3) → output
Key decisions:
- Semantic chunking along logical boundaries (F1 = 0.91)
- Multi-hop retrieval for complex questions (the ReAct pattern)
- A citation verification layer — entailment classification, precision 0.97
- Source reputation weighting in retrieval
- An eval framework: a golden set of 500 QA pairs, accuracy, citation precision and recall, latency p95
Cases: automatic drafting of DD reports (70% of the time saved), adverse media screening (50K articles per day), entity resolution (40+ sources).
Open source: agentd22/osint-rag (Apache 2.0).
← Back to news