RECIPER: A Dual-View Retrieval Pipeline for Procedure-Oriented Materials Question Answering

1. Problem Statement

Retrieving procedure-oriented evidence from materials science literature is a significant challenge for Retrieval-Augmented Generation (RAG) systems. The core difficulties include:

• Scattered Information: Key synthesis details (steps, materials, conditions) are often dispersed across long, context-heavy paragraphs rather than being isolated in structured sections.
• Limitations of Dense Retrieval: Standard dense retrieval (paragraph-level) often fails to capture fine-grained procedural logic, returning unstructured text chunks that make specific steps hard to identify.
• Fragmentation: Chunk-level retrieval breaks the logical connections between interdependent procedural sections.
• LLM Hallucinations: Without precise retrieval, Large Language Models (LLMs) may hallucinate scientific procedures due to static training data limitations.

Existing approaches using expert priors or structured representations have not fully established whether compact procedural abstractions can serve as an effective auxiliary retrieval view alongside standard text.

2. Methodology: The RECIPER Framework

RECIPER proposes a dual-view retrieval pipeline that treats procedural knowledge as a complementary signal to standard contextual evidence. The framework consists of four main stages:

A. Procedure-Centric Knowledge Extraction

• Goal: To transform dispersed synthesis workflows into a retrieval-friendly format.
• Process: An instruction-following LLM (DeepSeek-R1-Distill-Qwen-32B) processes the full text of scientific papers to generate compact procedural summaries (referred to as "Recipes").
• Format: These summaries are step-oriented descriptions covering materials, operations, and conditions, compressing long-form methodology while preserving key cues.
• Representation: The system maintains two views for each paper:
  1. Paragraph View ($P$): Standard text chunks preserving broad context.
  2. Recipe View ($R$): Compact, LLM-extracted procedural summaries.

B. Dual-View Candidate Retrieval

• Independent Retrieval: Given a user query $q$, the system retrieves candidates independently from both the Paragraph view ($P_q$) and the Recipe view ($R_q$) using dense embeddings.
• Scoring: Candidates are scored based on embedding-space distance ($d$), where $s_i = \frac{1}{1 + d(q, e_i)}$.
• Complementarity: The Paragraph view provides narrative context, while the Recipe view emphasizes condensed synthesis logic.

C. Candidate Merging and Stream-Aware Deduplication

• Merging: The two candidate sets are merged into a unified pool ($C_q^{(0)} = P_q \cup R_q$).
• Deduplication: To prevent redundancy (e.g., multiple chunks from the same paper or stream), a stream-aware deduplication step is applied. For every unique (paper_id, stream) pair, only the highest-ranked candidate is retained. This preserves cross-view complementarity while eliminating within-stream noise.

D. Query-Aware Lexical Reranking

• Motivation: Dense retrieval handles semantic matching well but may miss exact term alignment.
• Mechanism: A lightweight reranking step calculates a query-coverage score ($o_i$) based on the intersection of query tokens and candidate tokens (title + body).
• Final Score: The final score combines the semantic score ($s_i$) and the lexical score ($o_i$):
  $$\hat{s}_i = s_i + \lambda o_i$$
  (where $\lambda = 0.1$). This promotes candidates that explicitly cover query terms without disrupting the primary semantic ordering.

3. Key Contributions

1. Dual-View Pipeline: Introduction of a novel framework that jointly indexes paragraph-level context and LLM-extracted procedural summaries ("Recipes") for materials science QA.
2. Empirical Validation of Complementary Signals: Demonstration that while procedural summaries are weak as standalone retrieval units, they provide significant complementary value when fused with paragraph retrieval.
3. Backbone-Agnostic Robustness: The method shows consistent improvements across four different dense retrieval backbones (BGE, MiniLM, Contriever, E5), proving the approach is not tied to a specific embedding model.
4. Improved Downstream QA: Evidence that better retrieval of procedural steps leads to higher quality answers, particularly for smaller LLMs with limited parametric knowledge.

4. Experimental Results

The system was evaluated on a benchmark of 1,024 query-document pairs derived from 300+ materials science articles.

Retrieval Performance (BGE-large-en-v1.5 Backbone)

RECIPER significantly outperforms both paragraph-only baselines and naive hybrid fusion methods:

• Recall@1: Improved from 84.08% (Paragraph-only) to 86.82% (+2.74).
• Recall@5: Improved to 97.07%.
• Recall@10: Improved to 97.85%.
• nDCG@10: Achieved 0.9283 (vs. 0.9060 for baseline).
• MRR: Achieved 0.9116 (vs. 0.8875 for baseline).

Cross-Backbone Analysis

Across all four tested backbones, RECIPER achieved average gains of:

• +3.73 in Recall@1.
• +2.85 in nDCG@10.
• +3.13 in MRR.

Downstream QA Performance

When integrated into RAG pipelines with LLMs ranging from 0.5B to 40B parameters:

• RECIPER consistently outperformed "No-RAG" and "Paragraph-Dense RAG" baselines on metrics like BERTScore-F1, ROUGE-L, and BLEURT.
• Impact on Small Models: The improvement was most pronounced for smaller models (e.g., Qwen-2.5-7B), suggesting that high-quality procedural retrieval can compensate for limited internal knowledge.

5. Significance

• Scientific QA Advancement: RECIPER addresses a critical gap in scientific literature retrieval by specifically targeting "how-to" synthesis questions, which are often poorly served by standard semantic search.
• Efficiency vs. Accuracy: By using lightweight LLM extraction and lexical reranking, the system achieves high precision without the heavy computational cost of complex generative retrieval models.
• Scalability: The framework offers a scalable foundation for extracting knowledge from complex, unstructured scientific literature, potentially aiding researchers in accelerating material discovery and synthesis planning.
• Generalizability: The success across multiple embedding backbones suggests this dual-view strategy could be adapted to other domains requiring procedural or step-by-step reasoning.