IKnowFlow: Trustworthy RAG for Sensitive Domains

1 Introduction

Vision. This doctoral research envisions RetrievalAugmented Generation (RAG) architectures for sensitive domains, where each generated claim carries an auditable trail of its retrieval origin, selection rationale, and attribution evidence, enabling domain experts to verify, contest, or refine system outputs before they reach end users.

Goal. The goal of this doctoral research is to formalize trustworthiness in RAG as a system-level property composed of four verifiable dimensions: traceability, interpretability, grounding, and provenance. Each dimension is studied at the RAG stage, where it is most at risk, and the resulting methods are unified under IKnowFlow.

In 2025, a preregistered study found that leading RAGbased legal tools hallucinate between 17% and 33% of the time, fabricating citations even when grounded in curated databases [Magesh et al., 2025]. In healthcare, RAG systems produce unsupported claims by blending retrieved passages with parametric priors over contradictory evidence [Kim et al., 2025]. These are structural failures: RAG was designed for factual grounding but never engineered to be trustworthy.

RAG [Lewis et al., 2020] reduces confabulation by grounding Large Language Model (LLM) responses in external evidence but does not guarantee trustworthy generation. Three bottlenecks persist (see Figure 1). Dense retrievers match on embedding similarity without exposing why a document is selected. Re-rankers optimize scores without generating justifications. Citation mechanisms verify surface correspondence without tracking whether claims causally descend from retrieved evidence. This research addresses these bottlenecks through IKnowFlow, operationalizing trustworthiness as traceability, interpretability, grounding, and provenance. The following four questions guide this thesis:

1. RQ1 (Traceability): How can source-attribution failures be diagnosed across model families and domains?
2. RQ2 (Interpretability): How can dense retrievers expose interpretable query-document matching signals?
3. RQ3 (Grounding): How can RAG systems select explicitly justified evidence to keep generation grounded in relevant and verifiable source context?
4. RQ4 (Provenance): How should attribution preserve provenance structure, capturing how evidence originates, evolves, and is challenged over time?

2 Current Progress and FutureWork

My current research addresses RQ1 and RQ2. To diagnose traceability failures (RQ1), REASONS [Saxena et al., 2025b] benchmarks LLMs on 20K scientific papers across 12 domains, revealing that even state-of-the-art models with retrieval augmentation fabricate citations in scientific text, meaning the attribution problem propagates through reranking and generation. To address interpretability (RQ2), IMRNNs [Saxena et al., 2026b] augment dense retriever with two learned adapters: one conditions document embeddings on the query, the other refines query embeddings using corpus-level feedback. This bidirectional modulation exposes human-readable keyword-level matching signals, yielding +7.14% recall across seven benchmarks. For sensitive applications, IMRNNs gives domain experts the ability to audit why a specific document was retrieved; a prerequisite absent from most existing dense retrievers.

To address RQ3, METEORA [Saxena et al., 2026a] replaces score-based re-ranking with rationale-driven selection, where a preference-tuned LLM generates query-conditioned rationales, an evidence chunk selection engine pairs these with passages using adaptive cutoffs, and a verifier filters poisoned content before generation. Across six datasets spanning scientific, legal, and finance, METEORA achieves 33.34% higher answer accuracy using 50% fewer chunks, and raises adversarial defense F1 from 0.10 to 0.44. Extending grounding further, my work on Neurosymbolic Retrievers [Saxena and Gaur, 2026] integrates symbolic reasoning with neural retrieval. Evaluations with licensed psychologists show that neurosymbolic grounding improves clinical relevance in mental health risk assessment over purely neural approaches. Together, these contributions make RAG auditable in domains where unexplained evidence selection invites legal liability and clinical risk.

The completed work substantially advances RQ1, RQ2, and RQ3, with RQ3 continuing to evolve through ongoing neurosymbolic extensions (see Figure 2). My comparison of generation-time and post-hoc citation [Saxena et al., 2025a], showing post-hoc verification achieves 78% versus 69% human-evaluated correctness, begins addressing RQ4. Two directions remain. First, I will study negation-aware retrieval as a safety-critical blind spot in dense retrievers. In clinical settings, confusing "patient denies suicidal ideation" with "patient reports suicidal ideation" is not just a retrieval error; it can change the meaning of the evidence entirely. I will examine whether IMRNNs can be adapted to capture such negation at the embedding level in an interpretable manner. Second, provenance-aware verification (RQ4) will construct a claim-level ontology with typed relations such as extends, validates, contradicts, and summarizes, enabling verification over dynamic source graphs. In law, healthcare, and science, knowing that a claimwas later contradicted is as important as knowing it was ever stated.