Retrieval Augmented Generation of Literature-derived Polymer Knowledge: The Example of a Biodegradable Polymer Expert System

Technical Deep Dive: Retrieval Augmented Generation of Literature-derived Polymer Knowledge

Overview

This work introduces the "Polymer Literature Scholar", a retrieval-augmented system designed to support literature-grounded reasoning for polymer science. It develops and benchmarks two complementary retrieval pipelines - a dense semantic vector-based approach (VectorRAG) and a structured graph-based approach (GraphRAG) - using a curated corpus of over 1,000 polyhydroxyalkanoate (PHA) papers.

Problem & Context

• Polymer science literature is vast, heterogeneous, and largely unstructured, making it difficult to retrieve, integrate, and use for scientific reasoning.
• Existing tools often extract narrow, study-specific facts without providing the integrated context needed to answer cross-cutting scientific questions.
• Retrieval-augmented generation (RAG) offers a promising approach, but its effectiveness depends on how domain knowledge is represented.

Methodology

Corpus Construction

• The literature corpus comprises ~3 million materials science documents, from which a PHA-specific subset of 1,028 papers was extracted and parsed.
• Paragraphs were extracted from the full-text articles for downstream processing.

VectorRAG Pipeline

• Uses dense semantic embeddings and similarity search to retrieve relevant paragraphs.
• Leverages a domain-aware text chunking strategy to preserve scientific context.

GraphRAG Pipeline

• Constructs a canonicalized knowledge graph from the corpus, with entities, relations, and citations.
• Retrieves relevant subgraphs through a hybrid of string-based and embedding-based matching, followed by context-aware re-ranking.

Evaluation

• Benchmark using Recall@K, Recall PID@K, and Accuracy evaluated by domain experts.
• Compare against state-of-the-art systems like GPT and Gemini.

Results

• GraphRAG achieves higher precision and interpretability, while VectorRAG provides broader recall.
• Both pipelines outperform general-purpose systems on domain-specific questions.
• Expert validation confirms the systems produce well-grounded, citation-reliable responses with high relevance.

Interpretation

• Tailoring retrieval to the structure and reporting conventions of polymer literature is crucial for reliable, trustworthy knowledge synthesis.
• Complementary vector-based and graph-based approaches leverage different strengths to address a range of scientific queries.
• Grounding model outputs in explicit source evidence enhances transparency and enables researchers to assess credibility.

Limitations & Uncertainties

• The work focuses on PHAs as a representative case study; further evaluation is needed to generalize the approach to other polymer domains.
• While the benchmark covers a range of query types, real-world scientific inquiries may introduce additional complexities.
• Expert evaluation, while valuable, relies on subjective judgments and may not capture all nuances of scientific reasoning.

What Comes Next

• Explore further enhancements to the retrieval pipelines, such as integrating interactive clarification of user queries.
• Expand the corpus to cover a broader range of polymer systems and scientific disciplines.
• Investigate methods for automatically identifying gaps or inconsistencies in the literature based on the system's outputs.

Sources:

• Retrieval Augmented Generation of Literature-derived Polymer Knowledge: The Example of a Biodegradable Polymer Expert System