← 論文一覧に戻る

PCF-Scout：持続可能な検索・推薦のためのエージェント型製品カーボンフットプリント推定

PCF-Scout: Agentic Product Carbon Footprint Estimation for Sustainable Retrieval and Recommendation (原題)

Giuseppe Spillo, Allegra De Filippo, Cataldo Musto, Michele Patella, Michela Milano, Giovanni Maria Semeraro

Zenodo (CERN European Organization for Nuclear Research)ジャーナル2026-11-01#AI×ESGOrigin: EU経営インパクト: 調達リスク対象セクター: cross_sector
DOI: 10.5281/zenodo.23861944
原典: https://doi.org/10.5281/zenodo.23861944

🤖 gxceed AI 要約

日本語

PCF-Scoutは、LangGraphのReActエージェントとWeb検索を用いて製品のカーボンフットプリント（PCF）を推定するオープンリソースである。電子機器・衣類・家庭用品の3領域で各194製品についてEPDベースの正解PCFを整備し、Gemini 2.5 FlashとOpenAI o3-miniによるゼロショット推定とエージェント呼び出しのログを公開している。各推定には値・出典・説明が付与され、持続可能な検索・推薦への応用が想定される。

English

PCF-Scout is an open resource for estimating product carbon footprints (PCF) using a LangGraph ReAct agent with web search. It provides EPD-based ground-truth PCF for 194 products each across Electronics, Clothing, and Home & Kitchen, plus zero-shot and agent estimates from Gemini 2.5 Flash and OpenAI o3-mini with full execution logs. Each estimate includes value, source, and explanation, targeting sustainable retrieval and recommendation.

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

📝 gxceed 編集解説 — Why this matters

日本のGX文脈において

Scope 3のカテゴリ1（購入製品・サービス）算定は日本企業のSSBJ・有報開示で最大の負担であり、LLMエージェントによる製品別PCF推定の自動化は実務的な示唆が大きい。サプライヤー一次データが乏しい製品群での推計手法として注目される。

In the global GX context

Automated product-level PCF estimation directly addresses Scope 3 Category 1, the largest and hardest-to-measure emissions category under GHG Protocol, ISSB/CSRD and SEC climate disclosure. Agentic LLM approaches with source attribution offer a scalable complement to supplier-specific EPD data.

👥 読者別の含意

🔬研究者:LLMエージェントによるPCF推定のベンチマークと正解データセットを再利用できる。

🏢実務担当者:サプライヤー一次データが無い製品のScope 3推計にエージェント型手法を試験導入できる。

🏛政策担当者:製品別炭素情報の自動推計・開示インフラ設計の参考になる。

📄 Abstract（原文）

PCF-Scout: agent source code, ground truth and large files This record contains the source code of the PCF-Scout agent, the ground truth, and the large files of the PCF-Scout resource, which are not stored in the Git repository, but accessible with DVC or this record: https://github.com/giuspillo/react_pcf_retrieval The full code, the use-case splits and the PCF annotations of the catalog are available in the repository. Files File Content Size (zipped / unzipped) MD5 pcf_scout_source.zip Source code of the agent: pcf_scout/ (LangGraph ReAct agent, script to run it on the ground truth or on a catalog, zero-shot baseline), its configuration files (config/dataset.yaml, config/emissions_enrichment.yaml, config/zero_shot.yaml), pyproject.toml and .env.example 24 KB / 77 KB cd1513fed74d7a8b612141ff3a40b678 ground_truth_data.zip ground_truth/data/: the ground truth, one JSONL file per domain (Electronics, Clothing, Home and Kitchen; 194 products each). Each line contains the product ID and title, the EPD-based PCF (pcf_kg), the emission class, the zero-shot estimates and the PCF-Scout calls (value, source, explanation) with Gemini 2.5 Flash and OpenAI o3-mini 222 KB / 1.3 MB 708be1043a8f1fcad68cda735884871f ground_truth_logs.zip ground_truth/logs/: execution logs of PCF-Scout on the ground truth (4,446 logs, one per call, with Gemini 2.5 Flash and OpenAI o3-mini) and index.tsv, which maps each log to its item and answer 24 MB / 85 MB 6509bdb60f68fe99bd301ea239503eb0 use_case_logs.zip use_case/logs/: execution logs of PCF-Scout (Gemini 2.5 Flash) on the products of the use-case catalog 86 MB / 283 MB df810d998aa16f437cf06c0a6742d7ad use_case_raw_data.zip use_case/data/electronics/raw/: Amazon Reviews'23 Electronics user reviews and item metadata of the catalog 90 MB / 311 MB 3c92dd9b7ab14ca06e8da9f50ce58c3a Each log reports the input product, the web-search queries issued by the agent with their results, and the final answer (PCF in kg CO2e, source, explanation). Usage Agent. Unzip pcf_scout_source.zip, install the project (pip install -e .), set the API keys listed in .env.example, and run python -m pcf_scout.run (agent) or python -m pcf_scout.zero_shot (zero-shot baseline) from the unzipped folder. To run them on the ground truth, unzip ground_truth_data.zip in the same folder. Large files. Clone the repository and unzip the files in its root: the paths inside the archives are relative to it, so the folders are restored in place. git clone https://github.com/giuspillo/react_pcf_retrieval cd react_pcf_retrieval unzip /path/to/ground_truth_data.zip # already in the repository; optional unzip /path/to/ground_truth_logs.zip unzip /path/to/use_case_logs.zip unzip /path/to/use_case_raw_data.zip The same files can also be downloaded with DVC (dvc pull), as described in the repository README. Notes The Amazon Reviews'23 data in use_case_raw_data.zip is redistributed from the original dataset (Hou et al., 2024) and remains subject to its terms of use.

🔗 Provenance — このレコードを発見したソース

🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。

gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。