AI-Driven Climate Equity: Mapping Policy Practices with RAG-LLM Analysis

AI-Driven Climate Equity: Mapping Policy Practices with RAG-LLM Analysis

Table of Contents

• Introduction
• Why This Matters
• AI in Planning: Planners, Jobs, and RAG
• Semantic Review of Climate Equity Plans
• Policy Recommendation Engine: How It Works
• Patterns, Gaps, and Geographic Clustering
• Key Takeaways
• Sources & Further Reading

Introduction

Climate equity is not just about cutting carbon; it’s about making sure the benefits and burdens of climate action are shared fairly. That idea sits at the center of a new line of work exploring how AI—specifically retrieval-augmented generation (RAG) with large language models (LLMs)—can aid policymakers and planners. This blog post dives into a fresh study that maps, compares, and learns from climate equity policy practices across U.S. cities using an RAG LLM-based semantic analysis and a recommendation system. The work, described in the paper “Mapping and Comparing Climate Equity Policy Practices Using RAG LLM-Based Semantic Analysis and Recommendation Systems” by Seung Jun Choi and colleagues, shows how AI can help digest vast policy texts, surface patterns, and spotlight policy gaps across jurisdictions. You can read the original research here: https://arxiv.org/abs/2601.06703.

What makes this study stand out is its end-to-end use of AI to go from raw planning documents and job postings to actionable cross-city insights. The researchers first examine planning-related job postings to assess how planning roles are evolving in the AI era. Then they apply a LangChain-native retrieval-augmented generation (RAG) pipeline to climate equity plans—pulling out policies, strategies, and actions, specifically in transportation and energy domains. The goal isn’t to prove which AI model is best; it’s to demonstrate how an AI-assisted framework can help planners compare, diagnose, and learn from across-city policy practice, and even suggest missing actions that peer cities have pursued.

If you’re curious about how this actually works in practice, imagine a librarian with a highly skilled search assistant: the AI reads thousands of PDF plans, identifies relevant policy pieces, and then feeds them into a system that highlights what cities are doing similarly or differently. The result is a content-rich, searchable map of climate equity work that can be browsed by city to surface comparable peers and potential policy options. That’s the vision this study offers—a practical, AI-augmented approach to cross-city learning for climate action.

Why This Matters

• Significance in the current moment: AI-enabled policy analysis is moving from curiosity to utility. With climate action accelerating and cities experimenting with climate equity and justice, having scalable, transparent tools to compare plans across dozens or hundreds of jurisdictions is incredibly valuable. This study shows that LLMs, when paired with retrieval techniques, can systematically extract policy content, assess its presence, and surface similarities and gaps. The method directly addresses a long-standing planning bottleneck: how to compare policies across cities in a way that’s both rigorous and usable for practitioners.

• Real-world scenario you could use today: Imagine a regional planning agency or a city’s climate office that wants to benchmark its climate equity plan against peer cities. They could deploy a RAG-based system, fed by publicly available plans, to automatically identify which cities have adopted similar actions (e.g., expanding public transit, deploying energy efficiency programs, or ensuring equitable access to charging infrastructure) and which actions are underrepresented in their own city. Suddenly, policymakers have a digestible, evidence-backed menu of options with clear references to source materials and pages, enabling faster, more informed decision-making and dialogue with stakeholders.

• How this builds on prior AI research: The study ties into a growing body of work showing that LLMs can assist in normative planning (where judgment and ethics matter) without supplanting human experts. It extends prior AI-driven policy analysis by integrating a robust RAG pipeline with a structured, content-based recommendation mechanism. In contrast to earlier studies that relied on limited samples or non-scalable prompts, this work demonstrates large-scale retrieval, chunking, embedding, and evaluation across thousands of pages and dozens of plans. It also foregrounds the importance of governance and transparency, using prompt constraints and explicit sourcing to reduce hallucinations and improve credibility.

In short, the research is timely because it aligns with how cities actually work today: large, diverse policy documents, a need for cross-city learning, and a push toward AI-assisted planning that preserves human judgment and equity commitments.

AI in Planning: Planners, Jobs, and RAG

The evolving role of planners in the age of AI

One of the study’s early insights is about the job market around planning in an AI-enabled era. The authors scraped planning-related job postings (69,116 in total) and found that after deduplication and filtering, 83 postings remained for detailed analysis. Most postings still emphasize traditional planner duties—transportation, environmental planning, land use, housing, and related domains—alongside a strong emphasis on communicative responsibilities. In other words, despite the hype around GenAI, the day-to-day expectations of planners have not shifted into “AI replacement” mode. Instead, the data suggest a hybrid professional model: planners need to blend coordination and policy analysis with technical capacities like GIS, data analytics, and regulatory knowledge.

For readers, the takeaway is practical: AI in planning is not about replacing frontline planners; it’s about augmenting their ability to analyze, synthesize, and communicate complex policy landscapes at scale. The study’s sentiment is consistent with the broader planning literature that frames AI as a co-creative tool that supports, rather than overrides, human judgment (Sager, 2017; Forester, 1982). The paper even notes that while GenAI can reshape certain workflows, trust and governance remain essential.

How the AI setup works in this study

The authors use a LangChain-native retrieval-augmented generation (RAG) pipeline. Text from PDFs is chunked into 1,000-word segments with a 200-word overlap to maintain semantic continuity. Each chunk gets embedded with OpenAI’s text-embedding-3-small and indexed in a FAISS vector store. Retrieval uses a maximal marginal relevance strategy (k = 5, fetch_k = 20, λ = 0.7) to balance relevance and diversity—plus random sampling to diversify contexts and reduce bias. For generation, they employed a deterministic GPT-4o-mini setup (temperature fixed at 0) to ensure reproducibility. The system prompts ChatGPT to extract a hierarchical policy structure—policies, strategies, and actions—and explicitly cites page numbers to ground the outputs in source documents.

The practical upshot is an auditable pipeline: a user can press a city and get back a compact, source-backed set of policies, strategies, and actions drawn from the universe of climate equity plans the study analyzed.

If you want a nerdy-but-usable mental model, think of the RAG system as a librarian that not only fetches relevant books but also pulls out precise quotes, cites exact pages, and then uses a careful, constrained prompt to summarize what each policy concept means in context.

Semantic Review of Climate Equity Plans

What the document corpus looked like

The study examined climate equity plans across U.S. cities with populations over 100,000. From a pool of 346 cities, 192 had some form of regional climate equity plan. The plan documents—collected as PDFs—were then subjected to content analysis. The researchers also dug into the table of contents to identify which topics repeatedly show up as core elements, especially in the transportation and energy arenas.

In terms of scale, the authors analyzed 21,489 pages totaling 5,834,348 words. The documents span roughly 2009–2025, with a heavy concentration published between 2020 and 2025 (129 documents, about 69%). The geographic distribution is shown to be more dense in the contiguous U.S., with a handful of outliers in Alaska and Hawaii.

The textual analytics at work

The analysis starts with a pretty classic NLP stack: you extract terms from the table of contents (after removing generic stopwords) and apply term frequency–inverse document frequency (TF-IDF) weighting. Then, to surface latent themes across documents, they apply Latent Semantic Analysis (LSA) via truncated singular value decomposition (SVD). The outcome is topic scores for each document, which can be visualized as heatmaps to compare thematic emphasis across city plans.

The study also contrasts a table-of-contents-driven view with the RAG-generated outputs to show how machine-augmented synthesis aligns with human-placed emphasis. The themes that repeatedly surface across transportation plans include multimodal transport, transit accessibility, vehicle emissions, and complete streets. In energy policy, renewable energy, energy efficiency, grid modernization, and electrification appear as recurring pillars. In short, policy narratives tend to cluster around decarbonization through transportation and energy system transformation.

Positive language vs. caveats: a semantic sentiment angle

Beyond topic extraction, the authors used a BERT-based classifier to gauge semantic polarity in the policy-text outputs generated by the RAG process. They labeled outputs as positive (affirmative, forward-looking) or negative (acknowledging limits, costs, or uncertainties). Most plan sections—policies, strategies, and actions—tended toward affirmative language. This aligns with the practical aim of climate plans: to articulate a vision and commit to concrete steps.

But there are clear caveats. Some plans included explicit constraints or acknowledged gaps. The study provides representative samples illustrating both ends of the spectrum: positive cases from Albuquerque, Sacramento, and Cleveland show clear dedicated sections for policies, strategies, and actions with grounded references; negative samples from Brownsville, South Bend, and Tampa illustrate sections that are present but lack substantive detail or acknowledge cost or feasibility constraints.

This polarity analysis matters because it surfaces not just what is being promised in plans, but how strongly jurisdictions articulate their intentions. In policy terms, language matters: it shapes stakeholder expectations, accountability, and the perceived seriousness of commitments.

Themes in transport and energy policies

Tables in the study enumerate the most common themes:

• Transportation: themes include land-use integration with transportation, transit-oriented development, active and multimodal transport, EV charging infrastructure, transportation demand management, and complete streets. Action items often emphasize installing charging infrastructure and promoting shared mobility, among others. Notably, there are gaps in some areas like transit accessibility enhancements and targeted needs assessments.

• Energy: common themes center on renewable energy adoption, energy efficiency, decarbonization, electrification, grid modernization, and energy storage. Actions frequently focus on efficiency measures and community energy programs, with fewer documented efforts around formal energy impact assessments or advanced grid modernization in some plans.

This semantic portrait highlights that while the climate action narrative is broad, the practical implementation repertoire across cities shows both overlap and variation. The study’s semantic lens helps to reveal where plans converge and where they diverge in meaningful ways.

Policy Recommendation Engine: How It Works

From documents to recommendations

A standout feature of this study is a content-based recommendation system that matches policy content across cities. Unlike approaches that depend on diffusion patterns or user interaction histories, this system relies on the actual composition of policies, strategies, and actions. The researchers built paired city-policy matrices, computed cosine similarities, and identified the top five most similar city pairs for comparison. The goal is diagnostic: identify common practices and, crucially, gaps in target cities by looking at peers with similar policy footprints.

For every policy element, the researchers measured adoption rates among peer cities by averaging binary indicators. This yields a score-based sense of how widely a given policy or action appears across comparable jurisdictions. Then, in a practical twist, they created an interactive search-based interface (implemented with ipywidgets) that lets users input a city and see its most similar peers and which policy elements those peers have adopted that the target city has not. The interface is designed to be an analytical aid—transparent, reproducible, and user-driven—rather than a top-down decision tool.

The pipeline in plain terms

Think of it like this: you feed the system a long bookshelf of climate equity plans, and it splits each book into smaller passages. It then indexes them so a search can quickly pull out the most contextually relevant passages. When a user asks about a city, the system retrieves relevant chunks, compiles a context-rich answer with exact page citations, and returns a list of peer cities along with specific policy elements those peers have implemented. It’s a “policy cookbook” that doesn’t pretend to know what’s best for your city, but it points you to examples that have worked elsewhere under similar conditions.

A few technical notes that matter in practice:
- The chunks are 1,000 words each with a 200-word overlap, ensuring that ideas stay coherent across segments.
- Embeddings come from text-embedding-3-small, and the index is a FAISS store to support fast retrieval.
- Retrieval uses Maximal Marginal Relevance (MMR) with k = 5 retrieved chunks and fetch_k = 20 candidates, plus a λ of 0.7 to balance relevance and diversity.
- GPT-4o-mini runs deterministically (temperature = 0) to keep outputs stable and auditable.
- The system constrains responses with prompts that require explicit citations and a fallback “I don’t know” if evidence is insufficient.

All of this is designed to deliver replicable, source-grounded policy insights that practitioners can actually use to inform strategy and narrative in plan writing.

Real-world implications of the recommendation engine

For city planners and regional authorities, this approach lowers the cost of cross-city learning. It surfaces “missing but actionable” ideas from peer cities, turning passive benchmarking into proactive policy exploration. The study’s examples—Las Vegas’ transportation actions and Fort Lauderdale’s energy actions—illustrate how the tool can surface clusters of cities that share similar practice profiles. It’s not about copying someone else’s plan; it’s about identifying relevant patterns and tailoring them to local governance, budget constraints, and community needs.

This has real-world salience: a city with budget gaps or political constraints can still glean from peers with comparable scales and contexts. It also invites a more iterative policy design process—policy narratives can be refined against a palette of peer experiences and then tested in local forums before formal adoption.

Patterns, Gaps, and Geographic Clustering

Findings on similarity and gaps

The study’s comparative exercise found that no city achieved a perfect 20/20 on transportation or energy theme scores across policy, strategy, and actions. On average, transportation themes scored about 7 for policy, 6 for strategy, and 8 for action. Energy themes fared slightly better, averaging 9 for policy, 8 for strategy, and 6 for action.

A notable pattern: policy practice exhibits geographic clustering. For transportation actions, Las Vegas aligns with several California cities (Chico, Richmond, Berkeley, Long Beach) on shared mobility, community workshops, and awareness campaigns, but tends to miss explicit action items around transit accessibility or vanpool support. For energy actions, Fort Lauderdale’s profile clusters with Eastern U.S. peers (Saint Paul, Des Moines, Columbia, Sterling Heights, Buffalo) around energy transition and stakeholder engagement, but it also reveals gaps in items like energy education workshops, formal energy impact assessments, or aggressive grid modernization.

This geographic clustering is not about causation, but about the mobility of policy conversations. It supports Temenos and McCann’s framing of policy diffusion as a geography of “policy mobilities”—ideas circulating across places with regional flavors and constraints. The takeaway for practitioners is clear: you can learn from nearby peers who face similar regulatory environments and infrastructure realities, then adapt the lessons to local governance, implementation capacity, and political feasibility.

The practical value of seeing patterns and gaps

With a content-based recommendation, a city can quickly identify which peers have tackled similar policy challenges and what actions those peers implemented (or avoided). The system helps planners to:
- Benchmark policy narratives against peers with similar contexts.
- Surface underrepresented or overlooked actions that could be feasible in their jurisdiction.
- Build evidence-supported policy narratives that are easier to defend in stakeholder meetings and council deliberations.

The paper emphasizes that the goal is not a prescriptive recipe but an evidence-backed starting point for reflective planning and iterative policy refinement. It’s a tool for learning, not a replacement for local knowledge or community engagement.

Key Takeaways

• AI can scale cross-city policy learning: RAG LLM pipelines turn vast corpora of climate equity plans into searchable, context-grounded policy insights with explicit page citations.
• Planning work remains human-centered: The study finds planners’ roles continue to hinge on communication and judgment, with AI augmenting capabilities rather than replacing core professional duties.
• Transportation and energy dominate climate equity discourse: Semantic analysis confirms that local plans consistently emphasize decarbonization via transport and energy transformations, with transport infrastructure and EV charging being recurrent themes.
• Actionable gaps are identifiable: While many plans specify policies and strategies, certain concrete actions (like broad transit accessibility improvements, needs assessments, or formal energy impact assessments) are less consistently documented.
• A content-based recommender reveals geographic clustering: Cities with similar policy footprints tend to cluster regionally, enabling peer learning that is sensitive to local context and governance constraints.
• RAG-enhanced policy extraction improves accuracy and reduces hallucinations: By embedding documents and retrieving chunks with constrained prompts, the method narrows factual drift and anchors outputs to specific source material.
• The approach is practical, not prescriptive: The interactive interface helps policymakers explore peers and missing actions, facilitating informed drafting and iterative policy design.
• This is a stepping-stone for institutional adoption: While promising, careful validation, governance, and risk assessment are necessary before deploying such systems in public agencies.

If you’re curious to dive deeper into the specifics, the original paper provides the full methodology, tables of themes, and concrete examples from cities across the United States. It’s a rich, data-driven look at how climate equity work is evolving and how AI can help cities learn from each other in meaningful ways. For a direct read, check out the original research here: https://arxiv.org/abs/2601.06703.

Sources & Further Reading

• Original Research Paper: Mapping and Comparing Climate Equity Policy Practices Using RAG LLM-Based Semantic Analysis and Recommendation Systems
• Authors: Seung Jun Choi

This post pulled together core ideas, numbers, and takeaways from the study to present a readable, practitioner-friendly synthesis. If you want to explore the full depth—down to the specific thematic tables for transportation and energy, the sentiment samples, and the 21,489-page corpus—I highly recommend reading the paper itself. The work sits at an exciting intersection of climate policy, urban planning, and AI-enabled learning, and it offers a concrete path toward more informed, equitable, and responsive city climate action.