TRACE: Trustworthy Retrieval-Augmented Conversational Engine

TRACE: Trustworthy Retrieval-Augmented Conversational Engine proposes a retrieval-based, constraint-aware framework for public service conversational systems. The paper investigates the impact of retrieval quality on constraint-aware recommendation in public service conversational systems built over noisy and heterogeneous service directories. The authors propose TRACE, which parses input user queries into structural and semantic constraints for downstream retrieval, with the help of a dual data representation schema. The framework extracts structural constraints (e.g., city/county/ZIP) that are enforced exactly via knowledge graphs, and semantic constraints (e.g., eligibility requirements) that are checked using text fields. If no structural constraints are present, the system asks a clarification question. The retrieval process first generates a top-k candidate set via graph querying based on structural constraints, then applies evidence-based semantic filtering; if no candidates satisfy semantic constraints, it fetches the next top-k candidates.

The evaluation uses a curated statewide food pantry directory of approximately 800 pantries from Kansas, with structured fields (name, address, city, county, ZIP code, phone, hours) and unstructured fields (eligibility). The authors constructed a synthetic benchmark of 1,000 user queries across five families: location-only requests (500 queries), eligibility questions (200), open-hours queries (200), recall-style queries mentioning a specific pantry name (10), and combined queries (90). They compared multiple knowledge representation variants: KG-0 (no-KG baseline with text-only indexing), KG-1 (Location KG), KG-2 (Hours KG), and KG-3 (Location + Hours KG). They evaluated 15 open-source instruction-tuned LLMs across small (≤3B), medium (4–8B), and large (>8B) parameter groups, plus one proprietary model, GPT-5.5.

Key results show that KG-based representations substantially improve retrieval compared to the no-KG baseline. At k=3, KG-3 achieves the strongest retrieval with P@3=0.87, R@3=0.91, and F1@3=0.86, while KG-0 remains the weakest. For end-to-end reliability, averaged across all models, constraint satisfaction increases from 64.15% (KG-0) to 88.48% (KG-3), and semantic similarity improves from 0.67 to 0.77. Hallucination rate drops from 10.54% under KG-0 to 2.12% under KG-3 on average. The paper also finds that model-to-model variance is the largest under KG-0 and shrinks under stronger KG variants, meaning that stronger retrieval reduces sensitivity to model choice and enables more robust performance even with smaller models. The authors conclude that retrieval quality is the primary lever for dependable public service conversational access and that improving retrieval consistently improves end-to-end recommendation reliability, and that stronger retrieval stabilizes performance across diverse LLMs.

Improvements to AI systems:

1. Constraint-Aware Retrieval Pipeline: Implement a dual-layer retrieval mechanism that first extracts and enforces hard structural constraints (e.g., geo-location, IDs) via graph-based exact matching, then applies soft semantic constraints (e.g., eligibility, hours) via embedding-based filtering on text fields. This reduces hallucination by grounding answers in verified candidates.

2. Adaptive Clarification Trigger: Add a proactive clarification module that detects when no structural constraints are present in a user query and asks a single, targeted follow-up question (e.g., Which city or county?) before retrieval. This improves precision for under-specified requests.

3. Iterative Candidate Expansion: When the initial top-k graph query yields no candidates passing semantic filtering, automatically fetch the next top-k candidates and re-run semantic checks. This prevents premature failure and improves recall for noisy or heterogeneous directories.

4. Model-Robustness via Retrieval Strength: Use stronger knowledge graph representations (e.g., combining location and hours) to reduce sensitivity to LLM choice. This allows deploying smaller, faster models (≤3B parameters) without sacrificing end-to-end reliability, cutting inference cost and latency.

5. Hallucination Suppression via Evidence Filtering: Enforce that every generated response must cite at least one retrieved candidate that satisfies all extracted constraints. If none exist, the system returns a no match response instead of generating an ungrounded answer, reducing hallucination rates from ~10% to ~2%.

6. Structured-Semantic Hybrid Indexing: Build a dual representation schema where each service entry has a structured graph node (for exact filters) and an unstructured text embedding (for semantic similarity). This enables both SQL-like filtering and natural language matching in a single pipeline.

What the improved AI system can do:

• Answer location-based queries (e.g., food pantries in Sedgwick County) with 91% recall at k=3, even over noisy directories.

• Handle eligibility questions (e.g., Do I qualify if I'm a veteran?) by semantically filtering candidates and returning only those whose text fields match.

• Respond to combined queries (e.g., pantries open after 6pm in Wichita for families with children) by jointly enforcing structural and semantic constraints.

• Ask clarifying questions only when necessary, reducing user friction.

• Maintain high reliability (88% constraint satisfaction) across a wide range of LLMs, including small open-source models, making deployment feasible on edge devices or with limited compute.

• Avoid fabricating services; if no candidate satisfies all constraints, it explicitly states no match, improving trust in public service contexts.

Public service chatbots are expected to deliver recommendations from an underlying public service directory, while also making sure that the recommendations respect explicit user constraints. In practice, public service directories are noisy and inconsistent, and general-purpose large language model (LLM) or AI-based chatbots frequently generate unreliable recommendations, citing unverified sources from the web. We investigate the impact of retrieval quality on constraint-aware recommendation in public service conversational systems built over noisy and heterogeneous service directories. We propose TRACE (Trustworthy Retrieval-Augmented Conversational Engine), a retrieval-based, constraint-aware framework that parses input user queries into structural and semantic constraints for downstream retrieval, with the help of a dual data representation schema. Using a curated statewide pantry directory and a synthetic query benchmark, we evaluate multiple knowledge-representation variants with and without knowledge graphs (KGs). We experiment with several open-source LLMs and a proprietary model, showing that strengthening retrieval substantially improves user constraint satisfaction while reducing hallucinated recommendations. Performance differences across LLMs narrowed in our experiments as retrieval quality improved, making results less sensitive to model size. These findings suggest that the quality of retrieval is key for robust public service conversational systems.