How they work
The input-to-output mapping for this particular technique can be broken down into 3 separate components that form a pipeline for input processing.
Enhanced Accuracy: By accessing real-time or domain-specific information, RAG significantly reduces the likelihood of incorrect or outdated responses.
Knowledge Integration: Can incorporate proprietary or external databases, enabling the system to answer niche or organisation-specific queries more effectively.
Reduced Hallucinations: Retrieval mechanisms help ground the model’s output in factual data, minimising fabrications common in purely generative models.
Dependency on Data Quality: The effectiveness of RAG depends on the quality and relevance of the retrieved data. Poorly maintained or outdated databases can compromise output quality.
Latency: The retrieval process can introduce additional latency, especially if accessing large datasets or performing complex searches. Optimization is required for real-time applications.
Complex Integration: Implementing a RAG system requires combining LLMs with retrieval infrastructure, which can be technically complex and resource-intensive. This will require specialised expertise and maintenance. Retrieval systems will vary in complexity depending on the complexity of the downstream task. The more specialised the task, the more complex the system will need to be.
RAG-powered chatbots can provide tenants with highly accurate and contextually relevant responses by retrieving information directly from council policies, legal frameworks, or tenant guidelines. This can significantly reduce the risk of misinformation, which may occur with simple prompted LLM chatbots.
For example, a RAG system could provide tailored responses about rent payment schedules, maintenance updates, or eligibility for assistance programs by dynamically accessing tenant profiles and policy documents. This reduces the burden on staff and ensures consistent, informed communication with tenants. Additionally, RAG will offer the sufficient robustness to use procured advice to assist tenants on simple repairs/fixes which can be done without the need of a contractor.
While costs may be higher due to increased query volumes and infrastructure complexity, the improved accuracy and compliance will likely justify the investment, particularly in applications where misinformation could have serious consequences.
For social housing providers, RAG systems significantly improve querying complex housing datasets. By dynamically retrieving relevant documents or datasets—such as tenant records, contractor performance reports, or compliance guidelines—RAG systems can respond to nuanced queries that require multiple results from multiple data sources, including unstructured types with higher accuracy. For instance, a RAG system can retrieve and summarise relevant sections of building codes, safety regulations, or tenant rights documents in response to queries about construction work or home safety. This reduces the likelihood of errors or omissions in compliance-related decisions.
The prompted LLM system ensures that answers are grounded in specific and up-to-date datasets, offering richer insights than text-to-SQL models. Nevertheless, the integration of retrieval modules may increase infrastructure costs, and maintaining the quality of external databases is essential to ensure reliability. Since it can be seen as an appendage to the system described in AI Technique 1, councils don’t have to onboard the entire risk of RAG immediately, and can instead evaluate its necessity once the base prompted models are implemented and evaluated.
RAG has the potential to add a new layer of support to planning teams by dynamically accessing and querying vast repositories of legislative, historical, planning, and strategic vision documents.
For example, when evaluating a development's compliance with the Local Plan, RAG can extract relevant sections of legislation or past precedents in real-time. Where in prompted LLMs, it allowed for the static summarisation of non-digitised (and digitised) documents, it still required manual indexing of output summarisations. Instead RAG allows the dynamic retrieval of relevant documents according to an input query, tailoring responses according to the exact query it is provided, it can then provide links/references to the documents used, allowing for the simple verification of its responses. This tool can have a transformative impact on the internal functions of the planning team, allowing for the instant retrieval and summarisation of relevant documents for their intended task.
The cost implications are proportional to the complexity of indexing and maintaining historic datasets; however, these expenses are offset by the system’s ability to eliminate redundancies in manual research and improve accuracy. By grounding outputs in retrievable data, RAG minimises the risk of errors that could derail strategic objectives or conflict with established precedents, delivering value through both improved decision-making, reduced operational inefficiencies, and consequences of error.
When assessing complex planning applications, RAG systems excel by retrieving and integrating relevant regulatory frameworks, past decisions, and strategic plans to inform the evaluation process.
This ensures that applications with significant implications for the locality are reviewed comprehensively, with reference to past decisions and existing frameworks. For instance, when reviewing a mixed-use development proposal, RAG can pull zoning regulations, environmental compliance data, and community feedback, enabling planners to generate well-informed evaluations. Planning decisions are critical, therefore a full reliance on RAG systems carries too much risk, and should instead be used as an overview of the application, streamlining the time-to-decision process, and allowing for the provision of relevant documentation for the decision making process.
While RAG systems require higher upfront costs (compared to Prompted LLMs) for database setup, integration, and retrieval design, they can reduce long-term expenditures by minimising manual effort and ensuring higher-quality, defensible decisions that mitigate the risk of costly disputes or appeals.
As discussed in the other impact areas, RAG offers a robust solution to the querying of both structured and unstructured data according to input queries. Prompted LLMs offered solutions to complete the foundational processing of the data, but did not have the capacity to query the resulting data.
This subtle difference has a transformative impact on the internal functions of any team within the council. The ability to automatically retrieve the required information according to a specific need essentially eliminates the manual process of retrieving documents from a vast and dense database of documents that councils possess, which is not only time consuming but also prone to error and inconsistencies.
Although the cost implications are moderate, due to the increased complexity of the system, it is internally facing meaning traffic will be relatively low, and running costs will be minor. There is no doubt that the time saved to access key information, and the impact on service delivery justify the investment.
Residents often contact councils for various reasons, such as reporting issues, requesting information on services, or seeking guidance on application processes.
For instance, through chatbots residents can inquire about road closures, parking availability, or public transport schedules. Through RAG, live data can be retrieved to offer real-time updates, improving resident satisfaction and reducing call-center burden. For applications like parking permits, the chatbot can guide residents through the application process by retrieving the necessary forms, explaining eligibility criteria, and answering FAQs. Since this is an externally facing service meaning higher traffic level, and there is a need to integrate various forms of live data, the costs will be moderate, and further cost-benefit analysis will need to be conducted in order comprehensively justify the investment.
RAG-powered systems can help support care providers by creating detailed, dynamic case files that pull together a patient’s care history, current health data, and standard recommendations for their condition. These files give healthcare providers a clear, up-to-date summary, ensuring continuity of care and building on previous treatments for more personalised support over time. Unlike simple prompted LLMs, RAG ensures the information is accurate and contextually relevant, offering more reliable insights while improving the efficiency and consistency of patient care.
Similarly, a querying system can be built ontop of this data for on-demand access to relevant information during visits, adding additional flexibility and support, with similar costs as seen with other teams.
Chatbots powered by RAG can provide personalised advice at scale, acting as a first point of contact for residents seeking information on health or social care services. By retrieving updated medical guidelines, social care policies, or other trusted sources, these chatbots can deliver accurate and context-specific responses. On the riskier end of this kind of service, the chatbot can be a source of health information, offering recommendations for action, which obviously comes with liability, and since 100% cannot be guaranteed and errors are inevitable, the consequence of error may be too severe to justify implementation. The safer service may point residents towards resources, programmes, etc. that the council offers.
This can be integrated with remote monitoring, RAG-powered chatbots combined with sentiment analysis can guide regular check-ins, retrieving relevant health advice or risk indicators from NHS data sources to tailor their interactions.
© 2025 DG Cities. All rights reserved.