SpectraQuery: A Hybrid Retrieval-Augmented Conversational Assistant for Battery Science

TL;DR

The paper addresses the challenge of reasoning across structured experimental data and unstructured literature in battery science, specifically operando Raman data. It proposes SpectraQuery, a hybrid retrieval-augmented conversational system with a SUQL-inspired planner that issues coordinated SQL over peak-parameter data and literature queries to a vector store, then synthesizes grounded, cited answers. Key contributions include a SUQL-like planner, an integrated evaluation suite spanning SQL correctness, grounding, retrieval diversity, and expert ratings, and demonstration of strong performance (e.g., ~80% SQL correctness and grounding above 90% with 10–15 retrieved passages). The work demonstrates the feasibility and value of hybrid data architectures for accelerating interpretable scientific insight, and outlines concrete paths for extending the approach to additional modalities, external knowledge bases, and interactive debugging to further improve reliability and usefulness.

Abstract

Scientific reasoning increasingly requires linking structured experimental data with the unstructured literature that explains it, yet most large language model (LLM) assistants cannot reason jointly across these modalities. We introduce SpectraQuery, a hybrid natural-language query framework that integrates a relational Raman spectroscopy database with a vector-indexed scientific literature corpus using a Structured and Unstructured Query Language (SUQL)-inspired design. By combining semantic parsing with retrieval-augmented generation, SpectraQuery translates open-ended questions into coordinated SQL and literature retrieval operations, producing cited answers that unify numerical evidence with mechanistic explanation. Across SQL correctness, answer groundedness, retrieval effectiveness, and expert evaluation, SpectraQuery demonstrates strong performance: approximately 80 percent of generated SQL queries are fully correct, synthesized answers reach 93-97 percent groundedness with 10-15 retrieved passages, and battery scientists rate responses highly across accuracy, relevance, grounding, and clarity (4.1-4.6/5). These results show that hybrid retrieval architectures can meaningfully support scientific workflows by bridging data and discourse for high-volume experimental datasets.

Figures (18)

• Figure 1: Each spatial pixel in the Raman maps (right panels) represents the intensity distribution of characteristic vibrational modes identified by automated peak detection. Representative spectra from individual pixels (left panels) show the corresponding local features—(top) A1g charged transition-metal–oxygen vibration and (bottom) combined unknown carbon peaks
• Figure 2: Data and literature pipelines. (Left) Raman preprocessing → peak fitting → relational tables. (Right) Literature ingestion → metadata embeddings → chunking → vector search.
• Figure 3: The planner parses a natural-language question, emits coordinated SQL over the Raman database (left) and a literature query over the vector index (right), and returns harmonized intermediate tables/snippets for generation.
• Figure 4: Structured outputs (i.e., D/G table; left) and literature snippets (right) are fused by the LLM to produce a grounded, cited natural-language answer (bottom).
• Figure 5: LLM-as-a-judge SQL correctness scores (0, 0.5, 1.0) for three independent runs across the 30 benchmark questions. Darker cells indicate higher correctness.