Building Auditable Deep-Research Agents: GraphRAG, Multi-Hop Reasoning & Source Attribution

"Most of you have used a deep-research tool in the last month. How many of you actually opened the citations?" The primary failure mode of modern deep-research agents isn't "wrong facts" — it’s unfalsifiable claims: fluent, plausible prose with citations that no reviewer has time to audit. In this session recorded live at MBZUAI for Abu Dhabi Machine Learning (ADML Season 7, Episode 1), Bibin Prathap (Microsoft MVP, IEEE Senior Member, and AI Strategy Leader) breaks down why baseline vector RAG fails multi-hop synthesis, and how GraphRAG provides the deterministic error-correction substrate required for true autonomous research. Using the open-source VeritasGraph framework, Bibin walks through how to link chunks directly to knowledge graph edges, assemble multi-hop evidence chains across siloed documents, and enforce 6 concrete engineering patterns that turn a confident guesser into an agent that reliably knows when to abstain. 📦 Open-Source Code: https://github.com/bibinprathap/VeritasGraph 🌐 Speaker Website: https://bibinprathap.com/ 💼 Speaker LinkedIn: https://www.linkedin.com/in/bibin-prathap-... 🏛️ Hosted by: Abu Dhabi Machine Learning (ADML) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) 📌 Context Note: This video covers the topic of my selected session for Abu Dhabi Machine Learning, Season 7 Episode 1, scheduled for September 21, 2026, at MBZUAI. I was unable to attend in person and recorded this presentation separately. Organizer’s programme: https://www.meetup.com/adml-abu-dhabi-mach... --- ⏱️ TIMESTAMPS 00:00 - The Citation Problem: Unfalsifiable Facts 00:45 - The Premise: Evidence You Can Defend 03:00 - Session Outcomes & Vendor-Neutral Architecture 04:00 - Deep Research vs. AutoResearch: Attribution as an Error Substrate 06:00 - Why Baseline Vector RAG Breaks on Multi-Hop Queries 09:00 - Ingestion to Answer: Assembling Evidence Across 3 Hops 11:30 - Reasoning Traceability: Attribution is a Data Model, Not a Prompt 13:30 - 6 Engineering Patterns to Eliminate Hallucination 15:30 - Reproducibility: Pinning Corpus, Graph, and Traversal 17:00 - Enterprise Reference Architecture (ACLs, Boundaries, Logs) 18:00 - LIVE DEMO: VeritasGraph Knowledge Graph Explorer 19:30 - LIVE DEMO: Clicking Citations & Inspecting Reasoning Paths 20:30 - LIVE DEMO: PII Masking & Jailbreak Guardrails 21:30 - Don't Demo the Happy Path: 4 Failure Scenarios Handled 23:00 - Honest Trade-offs: The Actual Costs of GraphRAG 24:30 - Toward AutoResearch: Requirements Before Agents Run Experiments 26:00 - The "Dependable Agent" Pre-Flight Checklist 27:00 - Closing: Evidence First, Autonomy Second & Live Q&A --- 🔑 KEY ARCHITECTURAL TAKEAWAYS 1. Chunk boundaries sever relationships: Top-k vector similarity is a heuristic, not a reasoning plan. It cannot resolve queries where premise A, mediator B, and conclusion C live in distinct documents. 2. Attribution is a write-time data model: If provenance isn't stored when the edge is created, read-time prompting will only hallucinate plausible citations. A bibliography is not traceability. 3. Deterministic Traversal: LLM synthesis is stochastic, but graph traversal is deterministic. Separating the retrieval trace from prose generation makes answers auditable and reproducible. 4. Abstention as a safety valve: A research agent must explicitly return "insufficient evidence" and identify the missing graph edge rather than confabulating bridge data. --- 👤 ABOUT THE SPEAKER Bibin Prathap is an AI Strategy Leader, Microsoft MVP (Azure AI), and IEEE Senior Member based in Abu Dhabi. He is the creator of the open-source VeritasGraph framework. Over the past decade, he has focused on enterprise-grade, explainable, and trustworthy AI implementations for heavily regulated environments and sovereign entities. #MachineLearning #GraphRAG #AIAgents #RAG #VeritasGraph #DeepResearch #LLMOps #MBZUAI #ADML #ArtificialIntelligence