acceptodds
Under review as a conference paper at ICLR 2027

PCB-RAG: Conditional Retrieval over Multimodal PCB Component Datasheets

Abstract

Electronic component datasheets are essential for PCB design, yet their multi modal content and conditional specifications challenge retrieval-augmented gen eration (RAG).Existing systems may retrieve semantically relevant but condition ally incompatible evidence, producing incorrect specifications . We propose PCB RAG, which formulates datasheet QA as condition-aware evidence acquisition . It constructs a conditional multimodal knowledge graph that separates property concepts from specification statements and compiles effective contexts through scoped inheritance and per-variable overrides . A shared property encoder aligns equivalent properties while distinguishing evidence applicability across condi tions . Guided by a structured query representation (CQIR), an evidence-state pol icy selects retrieval actions and verifies condition coverage before answer gener ation . We also introduce PCB-CQA, a four-task benchmark with strict data splits for evaluating answer correctness and evidence sufficiency. Experiments show that PCB-RAG outperforms strong baselines in accuracy and evidence reliability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Retrieval-Augmented Generation, Conditional Retrieval, Knowledge Graph, PCB Component Datasheets, Multimodal

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3TL9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PCB-RAG, a condition-aware retrieval framework for multimodal PCB component datasheets. The authors argue that most specifications in datasheets are conditional assertions rather than universal facts. Therefore, retrieving relevant evidence alone is insufficient; the evidence must also be applicable to the conditions specified in the query.

The paper identifies three challenges: conditions are distributed across different document structures, property equivalence and evidence applicability require different supervision, and evidence sufficiency depends on the query structure.

To address these challenges, PCB-RAG constructs a conditional multimodal knowledge graph organized into four canonical views: Pin, Characteristics, Geometry, and Function. It also uses condition inheritance and overrides, a structured query representation (CQIR), and an evidence-state retrieval policy to acquire and verify relevant evidence.

The authors evaluate the framework using PCB-CQA, including experiments on unseen condition compositions, device families, and vendor styles.

Critical strengths
  1. Condition-aware retrieval for engineering documents

The paper addresses an interesting problem by considering not only whether retrieved evidence is relevant to a query, but also whether it is applicable under the requested conditions. I think this is a useful approach for engineering domains, where specifications often depend on operating conditions.

  1. Well-defined condition handling

The paper clearly defines how conditions are inherited and overridden across different scopes. In particular, the handling of variable bindings and rebindings helps distinguish different condition branches while preserving inherited conditions.

  1. Generalization experiments

I appreciate that the authors evaluate generalization across unseen condition compositions, device families, and vendor styles. These experiments provide useful evidence about the framework's ability to handle different conditions and datasheet structures.

Critical weaknesses
  1. Limited baseline comparisons and implementation details

Although the paper discusses PCB-QA and MMCircuitEval in the related work, these methods are not included in the experimental comparisons. Comparisons with existing multimodal engineering QA approaches, where applicable, would help provide a more comprehensive evaluation.

The paper also includes GraphRAG and KG Retrieval baselines, but the knowledge graph construction and retrieval settings of the KG Retrieval baseline could be explained more clearly to better understand the differences between the methods.

  1. Limited modality-specific analysis

The paper uses different types of evidence, including text, tables, and images. However, the experimental results do not clearly show how the framework performs on each evidence type separately.

Providing the distribution of these evidence types and modality-specific performance results would help readers better understand the contribution of multimodal evidence to the overall performance.

Questions for authors
  1. Could the authors explain why existing engineering QA approaches, such as PCB-QA or MMCircuitEval, were not included in the experimental comparisons?

  2. Could the authors provide more details about the knowledge graph construction and retrieval settings used in the KG Retrieval baseline, particularly how they differ from the proposed method?

  3. In Table 1, what specific criteria were used to classify questions as "condition-heavy"?

  4. Could the authors report the distribution of text, table, and image evidence in the dataset, along with separate performance results for each evidence type?

  5. Could the authors report the distribution of the verifier statuses (APPLIES, PARTIAL, DISJOINT, and UNKNOWN) and provide an analysis of verification errors for each status?

  6. Have the authors evaluated PCB-RAG using smaller models? If not, could they discuss how model size might affect performance?

Reviewer_1hR2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses question answering over multimodal PCB component datasheets, where most specifications hold only under conditions scattered across table headers, rows, footnotes and figures. Existing approaches apply chunk-based, graph-based or adaptive RAG, retrieving passages, communities or qualified facts by semantic relevance. These methods do not represent whether retrieved evidence actually covers the queried operating conditions and value roles, so they return relevant but inapplicable specifications and declare completion with insufficient evidence. The authors propose PCB-RAG, which compiles datasheets into a conditional knowledge graph with scoped condition inheritance, trains a shared encoder with separate equivalence, topic and retrieval projections, parses questions into a structured representation of conditions, value roles and quantifiers, and runs an evidence-state policy whose stopping is controlled by a deterministic verifier; they also introduce the four-task PCB-CQA benchmark and report gains over lexical, dense, graph and fixed-order baselines in accuracy, evidence recall and wrong-branch and false-complete rates. The distinguishing idea is treating condition applicability as an explicit retrieval state that is verified separately from proposition validity rather than as a property of similarity ranking.

Critical strengths
  1. The paper identifies a real and underexplored failure mode of retrieval, namely evidence that is topically relevant but conditionally inapplicable, and formalizes it through scoped condition compilation, value roles and a verifier that separates condition coverage from proposition validity; this problem framing is general beyond PCB and is the main reason the work merits consideration.

  2. The design of a shared encoder with separate equivalence, topic and retrieval projections, supervised by condition interventions so that the same record can be an equivalence positive and a retrieval hard negative, is a clean representation learning idea, and the ablations in Table 4 support that each of the condition executor, shared representations and verifier gate contributes to the reported gains.

Critical weaknesses
  1. All experiments are conducted on the authors' own benchmark, and the key metrics Evidence Closure, Wrong-Branch and False-Complete are defined in terms of the proposed graph's record fields and verifier states, so the reported gains cannot be separated from benchmark and metric design that mirrors the method, and no human agreement study or external benchmark is provided to validate them.

  2. The strongest baseline Rule/CQIR shares the proposed graph, parser and reader, while the external baselines operate on different infrastructure, so the main table cannot isolate how much improvement comes from the learned policy versus the extraction pipeline, and no independent agentic retrieval baseline using the same reader over raw documents is included.

  3. The learned components are described only abstractly, with no details on the parser and policy architectures, training set sizes for property pairs and state-action labels, or how questions, condition interventions and acceptable action sets are generated, which leaves open whether test questions follow the CQIR schema used for training.

Minor: Section 3.4 contains a broken figure reference and a missing space before "Cross-document" in Section 3.2, and the Table 1 column headers are garbled.

Questions for authors

Clarifications:

  1. How were the questions, condition interventions and acceptable action sets in PCB-CQA generated, and were any of them produced from CQIR templates or by the same extraction model used in the pipeline?

  2. For BM25, Dense RAG and GraphRAG, how are Evidence Closure and False-Complete computed when these systems have no explicit notion of condition records or completion declarations?

  3. What is the agreement between the programmatic scoring and the human review mentioned in Section 4.1, and how many instances were reviewed?

Future work:

  1. Could the authors report results on at least one external datasheet or conditional QA benchmark, or an agentic baseline that uses the same reader over raw PDF pages without the proposed graph, to show that the gains are not specific to the benchmark design?
Reviewer_9AjzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PCB-RAG, a framework for question answering over electronic component datasheets. The core method the authors present organizes specifications and their conditions in a multimodal graph. They also use an evidence state policy to retrieve missing information and design a verifier to check whether the evidence supports the requested answer. The paper also introduces PCB-CQA to evaluate the method on document-disjoint and family-disjoint evaluation setups.

Critical strengths
  1. The problem the paper tackles is practical and useful, and the problem setup and definition are clear. The task is important with downstream applications. The main issue they capture is insightful: a retrieved value could be relevant but still incorrect for requested devices or conditions. The method they propose does improve on such setup.

  2. The experiment is overall clear and shows substantial improvement over the baselines, across several reported tasks. The paper also includes useful abllations of the condition representation, retrieval policy, verifier etc.. The study helps readers understand how the different components contribute to the whole system. The description of the whole system is fairly clear.

Critical weaknesses
  1. The framework already tracks missing evidence and unmet requirements through its evidence state and verifier feedback. However, the main baseline still follows a fixed retrieval order. While the ablations show that using the evidence state helps, it remains unclear how much of the improvement comes from learning the policy itself versus simply adapting retrieval based on missing evidence.

  2. The trained policy higher probability to actions that help ask missing evidence or resolve uncertainty, it indeed be helpful. However, different acceptable actions may have very different costs and contributions to completing the query. The training objective doesn't really distinguish between locally useful actions and those that lead to more efficient evidence acquisition. Moreover, the learned encoder and policy introduce additional computational overhead.

Questions for authors
  1. Could the authors compare the policy they learned with a deterministic policy that selects actions based on the same missing requirements and unresolved evidence dependencies? Let's say keeping the graph, candidate actions, verifier, reader, and budget fixed would help isolate what is gained from learning the action selection.

  2. Also for the failure cases, it might be helpful to break down those into different type of failures and have some analysis? As in particular, I was wondering what happens when CQIR omits a constraint from the original question? Will a comparison using corrected CQIR on a small subset help identify whether the main bottleneck is finding the evidence or correctly representing what needs to be checked?

  3. Since False-Complete is measured among cases that declared complete, could the authors also report the number and fraction of questions for which each method declares completion? Will the system just learn to be conservative and refuse to answer uncertain or challenging questions? Having more of this would help distinguish more reliable answers from a more conservative stopping policy.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.