$τ$-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge

Authors	Quan Shi et al.
Year	2026
Field	AI / Agents
arXiv	2603.04370
PDF	Download
Categories	cs.AI, cs.CL, cs.IR

Abstract

Conversational agents are increasingly deployed in knowledge-intensive settings, where correct behavior depends on retrieving and applying domain-specific knowledge from large, proprietary, and unstructured corpora during live interactions with users. Yet most existing benchmarks evaluate retrieval or tool use independently of each other, creating a gap in realistic, fully agentic evaluation over unstructured data in long-horizon interactions. We introduce $τ$-Knowledge, an extension of $τ$-Bench for evaluating agents in environments where success depends on coordinating external, natural-language knowledge with tool outputs to produce verifiable, policy-compliant state changes. Our new domain, $τ$-Banking, models realistic fintech customer support workflows in which agents must navigate roughly 700 interconnected knowledge documents while executing tool-mediated account updates. Across embedding-based retrieval and terminal-based search, even frontier models with high reasoning budgets achieve only $\sim$25.5% pass^1, with reliability degrading sharply over repeated trials. Agents struggle to retrieve the correct documents from densely interlinked knowledge bases and to reason accurately over complex internal policies. Overall, $τ$-Knowledge provides a realistic testbed for developing agents that integrate unstructured knowledge in human-facing deployments.

---

Engineering Breakdown

Plain English

τ-Knowledge extends τ-Bench to evaluate conversational agents on their ability to retrieve and apply domain-specific knowledge from unstructured, proprietary corpora while interacting with users in long-horizon tasks. The paper identifies a critical gap: most existing benchmarks test retrieval or tool use in isolation, but real-world agents must coordinate both simultaneously to produce policy-compliant state changes. The authors introduce τ-Banking, a new fintech domain simulating realistic customer support workflows where agents must navigate unstructured knowledge bases, apply retrieved information correctly, and execute valid tool operations that modify system state. This work directly measures end-to-end agentic reasoning over messy, real-world data rather than synthetic or structured benchmarks.

Core Technical Contribution

The core novelty is τ-Knowledge, a benchmark framework that moves beyond isolated retrieval or tool-use evaluation to measure agents' ability to coordinate unstructured knowledge retrieval with external tool execution in a single agentic loop. Unlike prior work that treats knowledge retrieval (RAG) and tool use (ReAct-style agents) as separate problems, τ-Knowledge enforces verifiable, policy-compliant outcomes—meaning retrieved knowledge must be correctly applied through tools to produce the right state changes in a banking system. The τ-Banking domain concretely instantiates this: agents must parse customer requests, search proprietary knowledge bases for domain rules, retrieve relevant policies, and execute banking operations (transfers, account modifications) that must both follow retrieved policy constraints and pass compliance checks. This creates a realistic evaluation setting where neither retrieval nor tool use alone suffices; failure at any stage breaks the entire interaction.

How It Works

The τ-Knowledge evaluation framework operates as follows: (1) A user issue is presented to the agent in natural language, typically a customer support query in banking context. (2) The agent must search an unstructured, domain-specific knowledge corpus (simulating proprietary internal documentation) to retrieve relevant policies, procedures, or constraints applicable to that issue. (3) Simultaneously or subsequently, the agent must formulate and execute tool calls (simulating actual banking system operations) that interact with a stateful environment—for example, approving a loan, modifying account settings, or processing a transaction. (4) The system validates that the final state change is both compliant with the retrieved policies and correctly executed (no malformed operations, no unauthorized changes). (5) Success is measured only if all three conditions hold: correct knowledge was retrieved, the tool was called appropriately, and the resulting state is verifiable and policy-compliant. This differs from prior work by creating dependencies between retrieval and action—you cannot succeed by getting the tool right if you ignored the policy, and vice versa.

Production Impact

For engineers building conversational agents in regulated domains (fintech, healthcare, legal), this benchmark directly mirrors production constraints: your agent must reliably ground decisions in unstructured documentation while executing real system changes that regulators can audit. Adopting τ-Knowledge evaluation means (1) testing agents against realistic, proprietary-like corpora rather than clean Q&A pairs, exposing failure modes in knowledge extraction and disambiguation that lab settings hide; (2) validating compliance by checking state changes against retrieved policies, forcing you to instrument your knowledge base and tool outputs in ways that generate explainability logs; (3) measuring latency end-to-end through the retrieval-reasoning-execution loop, which often reveals that naive RAG+tool stacking introduces unacceptable latency in customer-facing systems. The trade-off is setup cost: you must maintain a properly curated, policy-annotated knowledge base and a stateful simulator that mimics your actual backend, plus define verifiable success criteria for each task type. This is non-trivial but essential if you're shipping agents in high-stakes domains.

Limitations and When Not to Use This

τ-Knowledge assumes that success can be defined via verifiable state changes and policy compliance checking, which works for structured domains like banking but may not generalize to open-ended domains (creative writing, strategic advice) where correctness is subjective. The paper relies on a finite set of pre-defined policies and tools in τ-Banking; it does not address scenarios where agents must compose novel solutions or handle ambiguous policies that require interpretation, which is common in real customer support. The benchmark likely assumes that unstructured knowledge is well-organized enough to retrieve (via embeddings or search); it does not measure performance on extremely chaotic, contradictory, or outdated knowledge bases that exist in many legacy enterprises. Finally, the paper does not empirically demonstrate that agents trained or selected via τ-Knowledge evaluation actually perform better in production—the link between benchmark performance and real-world outcomes remains unvalidated.

Research Context

This work builds directly on τ-Bench (prior work on long-horizon agent evaluation) and extends it from closed-world reasoning into knowledge-intensive settings, addressing a long-standing critique of RAG and tool-use benchmarks: they isolate problems that agents must solve jointly in production. It contrasts with prior work like HotpotQA (multi-hop QA over structured data) and ToolBench (tool use in isolation) by forcing both to occur together over unstructured corpora in stateful environments. τ-Knowledge opens a new research direction in agent evaluation—moving from synthetic, clean benchmarks toward realistic, regulatory-aware evaluation frameworks that capture the operational complexity of deployed conversational AI. This is particularly timely as large language models are increasingly integrated into customer-facing systems in regulated industries, making the gap between 'can the model retrieve facts?' and 'can the model correctly change system state while respecting policy?' a critical research frontier.

---

:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::

---

Back to Research Lab → · Subscribe to AI Letters →