Latest / 1 More Paper / Development of an Agentic Retrieval-Augmented Generation System with Hybrid Retrieval for Budget Planning and Validation — Anggara et al. (2026)
Transcript
- You’re listening to “Development of an Agentic Retrieval-Augmented Generation System with Hybrid Retrieval for Budget Planning and Validation,” by Devi Willieam Anggara and colleagues. Published in 2026. Abstract. Budget planning and validation require users to examine cost references, compare proposed budget items with relevant guidelines and historical records, and provide traceable evidence for each decision. This process becomes challenging when the required information is scattered across formal cost standards, budget plans, project realization records, and internal narrative documents. This study develops a budget validation system based on Agentic Retrieval-Augmented Generation (RAG) with hybrid retrieval. Small and medium enterprises are used as the application domain, while the main contribution lies in combining structured retrieval and semantic retrieval to support evidence-grounded budget validation. Structured information, including cost guideline items, budget components, and historical realization data, is stored in a relational database. Meanwhile, narrative documents are converted into Markdown chunks and indexed in a vector database. Three workflows are implemented and compared: Single-Agent RAG, Routing-Based Agentic RAG, and Orchestrator-Workers Agentic RAG. The evaluation is conducted using 30 test cases and four RAG-based metrics: context precision, context recall, faithfulness, and answer relevancy. Latency is also measured as an additional indicator of response time. The results show that Routing-Based Agentic RAG achieves the highest context recall score of 0.873 and the highest faithfulness score of 0.878, while sharing the highest context precision score of 0.731 with Orchestrator-Workers Agentic RAG. Single-Agent RAG obtains the highest answer relevancy score of 0.775 and the lowest latency of 26.763 seconds. Based on these findings, Routing-Based Agentic RAG is selected as the most suitable workflow because it provides stronger retrieval quality and factual grounding, which are critical for evidence-based budget validation. Introduction. Budget planning is an essential activity in project, operational, and organizational management. A budget does not only represent an estimation of costs, but also serves as a basis for decision making, cost control, and accountability. For this reason, each budget component should be supported by a clear calculation basis, comparable to relevant references, and traceable to supporting evidence. One of the main issues in budgeting is cost unreasonableness. A proposed budget value may exceed market or reference values, the proposed volume may not correspond to the actual activity, or a budget item may be submitted without a clear justification. In budgeting practice, this condition may lead to budget mark up, which refers to an unreasonable increase in cost or budget value beyond its actual basis. Such practices can result in inefficiency, inaccurate financial reporting, reduced trust, and potential legal risks. Corruption-prevention documents also indicate that planning and budgeting processes are vulnerable to inflated standards, fictitious budget components, and markup risks, making budget review against valid cost references necessary. Budget validation can be described as the process of assessing the reasonableness of budget components using relevant comparison sources. This process does not only focus on the total budget value, but also examines item names, categories, units, quantities, unit prices, planned costs, and conformity with applicable cost guidelines or historical records. In this study, the reference sources include formal cost guidelines such as PMK/SBM, INKINDO, ITB cost standards, and historical project budget and realization records,. Manual budget validation has several limitations. Users need to search through multiple documents, inspect cost tables, compare proposed values with historical records, and document the sources used as evidence. The required information is also stored in heterogeneous formats. Some sources are structured, such as cost tables, budget items, planned values, realization values, and variance records. Other sources are narrative, such as guideline explanations, internal notes, and supporting policy context. This condition creates a need for a system that can retrieve structured tabular data accurately while also understanding narrative documents through semantic retrieval. Retrieval-Augmented Generation (RAG) offers a potential solution by retrieving external knowledge before generating a response. However, a vector-only RAG approach is not sufficient for budget validation because numerical information, such as cost values, units, categories, and realization amounts, requires a more precise retrieval mechanism than semantic similarity alone. Therefore, this study develops an Agentic RAG system with hybrid retrieval. Structured cost and budget data are stored in a relational database, while narrative documents are indexed in a vector database. The system is evaluated by comparing three workflows: Single-Agent RAG, Routing-Based Agentic RAG, and Orchestrator-Workers Agentic RAG. The contributions of this study are threefold. First, this study designs a hybrid retrieval approach for budget validation by separating structured financial data from narrative knowledge. Second, it implements three Agentic RAG workflows to support cost guideline question answering, budget validation, and budget recommendation. Third, it evaluates the implemented workflows using RAGAS-based metrics and analyses which workflow is the most suitable for evidence-grounded budget validation. 2 Literature Study 2.1 Budget Planning and Cost Guideline Knowledge Budget reports support cost monitoring, comparison between planned and actual spending, and managerial decision making in project activities. In budget validation, cost guidelines are used as knowledge sources that define standard values, units, conditions, and constraints for assessing budget reasonableness. This study uses PMK/SBM as a reference for government standard input costs, INKINDO as a minimum standard for consulting services, and ITB cost standards as institutional references for academic activities,. These guidelines contain both tabular and narrative information. Cost tables include precise values such as rates, units, categories, positions, locations, and cost limits. Narrative sections provide supporting context, including scope, exceptions, and usage conditions. As a result, cost guidelines need to be represented in a form that preserves both numerical accuracy and contextual meaning. This requirement motivates the use of hybrid retrieval, where structured cost data and narrative guideline documents are handled through different but complementary retrieval mechanism. 2.2 Retrieval-Augmented Generation and Hybrid Retrieval Retrieval-Augmented Generation (RAG) combines information retrieval and text generation by first retrieving relevant external context and then using that context to guide the language model in producing an answer. This approach is useful for reducing unsupported or hallucinated responses because the generated answer can be grounded in retrieved evidence. However, budget validation is not only a semantic question answering task. It also requires the system to handle numerical values, units, categories, and relationships between budget components, guideline references, and project records. In this study, hybrid retrieval is applied by separating retrieval strategies according to the form of the data and the required level of precision. Structured information, such as cost guideline items, budget items, planned values, and realization values, is retrieved from a relational database. Meanwhile, narrative documents are retrieved through vector search. This design follows the idea that RAG systems working with heterogeneous documents need to preserve tabular structures instead of simply converting all tables into plain text. Financial question answering also requires retrieval strategies that can handle both numerical precision and contextual interpretation. In addition, enterprise RAG systems show the importance of combining structured and unstructured data sources before generating the final response. 2.3 Agentic RAG Workflows Agentic RAG extends conventional RAG by introducing decision-making capabilities into the retrieval and generation process. These capabilities may include tool selection, task planning, query routing, and collaboration between agents. Agentic RAG describe as an approach that improves system adaptability by allowing retrieval and reasoning steps to be adjusted based on the characteristics of the task. This capability is relevant to budget validation because cost guideline questions, budget validation questions, and budget recommendation questions often require different sources, retrieval paths, and reasoning patterns. This study compares three Agentic RAG workflow patterns. The first workflow is Single-Agent RAG, where a single agent is responsible for selecting sources, retrieving evidence, and generating the final response. The second workflow is Routing-Based Agentic RAG, where a router first classifies the user request into a specific use case before the system performs retrieval and answer generation. This pattern is suitable when different categories of input are more effectively handled by specialized processes. The third workflow is Orchestrator-Workers Agentic RAG, where an orchestrator plans the task, assigns subtasks to worker agents, and synthesizes their outputs into a final response. This pattern can be useful for tasks that require dynamic delegation across multiple evidence sources, although it may also introduce additional coordination complexity. 2.4 Evaluation of RAG Systems RAG systems need to be evaluated beyond the fluency of their generated responses. The evaluation should also consider the quality of retrieved contexts and the degree to which the final answer is grounded in the retrieved evidence. RAGAS provides several metrics for this purpose, including context precision, context recall, faithfulness, and answer relevancy. Context precision at K measures whether the top retrieved contexts are relevant to the question, as expressed in this formula. ∑ K k=1 (Precision@k x vk) Context Precision@K = Total relevant items in top K Context recall measures the proportion of reference answer claims that are supported by the retrieved contexts, as expressed in this formula. Context Recall = Supported reference claims Total reference claims Faithfulness measures the proportion of claims in the generated response that are supported by the retrieved contexts, as expressed in this formula. Total response claims When answer components can be compared against a reference answer, answer correctness can also be represented using the F1 score, as expressed in this formula. Answer Correctness (F 1) = 2 x Precision x Recall Precision + Recall Since RAG systems generate natural language responses, their evaluation is also related to LLM-as-a-judge methods. show that LLM-based judges can support scalable evaluation of conversational responses, while review various LLM-based evaluation methods along with their limitations. However, LLM-based evaluation may be influenced by prompt sensitivity, position bias, verbosity bias, and inconsistency. To reduce these risks, this study uses predefined RAGAS metrics, applies the same ground-truth dataset across all workflows, and complements the metric-based evaluation with manual review of the generated answers. 3 Methodology This study follows the Cross-Industry Standard Process for Data Mining (CRISP-DM), which consists of business understanding, data understanding, data preparation, modeling, evaluation, and deployment. The method is adapted to the development of an RAG-based budget validation system. Figure 1 explains the research stages used in this study. The process begins by defining the budgeting problem, continues with data understanding and preparation, and then moves to workflow modeling and evaluation. 3.1 Business Understanding The business understanding stage defines the problem context and the objectives that the system is expected to support. In this study, the main problem is the difficulty of preparing, checking, and reviewing budget plans when cost references, historical records, and supporting documents are distributed across different sources. Users need to search for relevant cost guidelines, compare proposed budget items with available references, assess the reasonableness of the proposed values, and provide traceable evidence for each validation result. This study uses small and medium enterprises as the application domain because organizations at this scale still require budget control but often have limited resources for manual review. However, the system is not designed to measure the business impact on SMEs. Instead, the focus is on developing a retrieval-based system that can support budget validation using available evidence. Based on the identified problem, the system is designed to support three main use cases. The first use case is cost guideline question answering, where the system provides source-grounded answers about cost standards and usage conditions. The second use case is budget validation, where the system compares proposed budget items with guideline and historical evidence. The third use case is budget recommendation, where the system provides initial recommendations based on guideline, historical, and contextual evidence. These use cases are summarized in Table I shows the three user needs supported by the system. UC1 handles cost guideline questions, UC2 validates submitted budget items, and UC3 provides budget preparation recommendations. 3.2 Data Understanding The data understanding stage is conducted to identify the data sources, their characteristics, and the quality issues that may affect system implementation. This stage is important because budget validation does not only require general document search, but also precise retrieval of cost values, units, planned amounts, realization amounts, and variances. Therefore, the system cannot treat all sources as ordinary text documents. The data sources used in this study consist of cost guideline documents, internal company documents, budget plan data, project realization data, and source metadata. Cost guideline documents include PMK/SBM, INKINDO, PER-033/2023 ITB, and PP-34/2024 ITB. These documents are used as references for answering guideline questions, assessing budget reasonableness, and supporting budget recommendations. Budget plan and realization data are used as validation objects and historical comparison sources. Internal documents provide organizational context, activity descriptions, and other supporting information relevant to budget planning.The analysis shows that the data sources are heterogeneous. Some sources contain structured information, such as cost tables, budget items, planned values, actual values, units, categories, and variance records. Other sources contain narrative information, such as guideline explanations, usage conditions, exceptions, internal notes, and policy context. Some documents are also semi-structured or require OCR assistance, which may introduce extraction errors or ambiguity in the relationship between item names, units, values, and their usage context. This stage identifies two main retrieval requirements. First, structured retrieval is needed to obtain numerical values and tabular attributes accurately, especially for cost guideline items, budget items, planned amounts, realization values, and variances. Second, semantic retrieval is needed to find contextual explanations from narrative documents, such as usage conditions, exceptions, and supporting internal information. This finding becomes the basis for using hybrid retrieval in the proposed system. Data quality also affects the validation behaviour of the system. Ambiguous numerical data should not be used as quantitative evidence before being validated. If the system cannot find relevant comparison evidence, it should state the limitation of the analysis instead of generating unsupported cost benchmarks. Therefore, source metadata, validation status, and traceability information must be preserved so that each generated answer can be linked back to its original document, page, table, row, or chunk. 3.3 Data Preparation The data preparation stage separates tabular information from narrative information because budget validation requires both precise numerical retrieval and contextual semantic retrieval. Tabular data are prepared for cost values, units, categories, planned amounts, realization amounts, and variances. Narrative documents are prepared to capture guideline explanations, usage conditions, exceptions, and internal contextual information. The data preparation and storage strategy is summarized in Table II. Based on Table II explains how each data source is prepared and stored. Structured tables are kept in the relational database for exact comparison, while narrative documents are prepared for semantic retrieval in the vector database. Structured data are stored in a relational database using schemas for guideline documents, guideline cost items, aliases, projects, budget categories, budget items, and project files. This storage supports exact retrieval of numerical values, units, categories, source documents, and historical comparisons. Narrative documents are converted into Markdown before being indexed in the vector database. Each chunk stores metadata such as source name, source type, page information when available, chunk identifier, and cleaned text. The loader workflow reads the Markdown chunks, generates embeddings, and stores the vectors and metadata in the vector database. This design allows the system to use relational retrieval for precise budget evidence and vector retrieval for contextual explanations. As a result, the system can avoid relying only on semantic similarity when answering questions that require exact numerical comparison. 3.4 System Overview The proposed system is designed as an Agentic RAG-based budget validation system with hybrid retrieval. The system follows a processing pipeline that starts from data acquisition and data preparation, then continues to online system processing and answer generation. Figure 2 illustrates the overall processing pipeline of the proposed system. The process begins with data acquisition, where budget guidelines, RAB files, internal documents, and historical data are collected as the main data sources. The data preparation stage applies table-text component selection to separate tabular information from narrative text. Tabular components are processed by extracting item names, units, prices, quantities, and categories. The extracted data are then normalized by standardizing formats and budget values before being stored in a relational database. Narrative components, such as guideline explanations and internal documents, are processed through document chunking. The text chunks are converted into vector representations and stored in a vector database to support semantic retrieval. After data preparation, three Agentic RAG workflows are modeled and evaluated in this study: Single-Agent RAG, Routing-Based RAG, and Orchestrator-Workers RAG. Each workflow uses the prepared knowledge sources to retrieve structured and narrative evidence before generating a response. The final output of the system consists of three types of responses: budget guideline answers, budget validation results, and budget recommendations. In the implementation, Supabase is used as the relational database for structured cost and budget data, while Qdrant is used as the vector database for embedded narrative chunks. The n8n workflow coordinates the retrieval and generation process, the HuggingFace BAAI/bge-m3 model is used for embedding generation, and OpenRouter provides access to the LLM. 3.5 Workflow Modeling and Implementation Three workflows are implemented using the same data sources and retrieval architecture so that the comparison focuses on workflow design rather than differences in available evidence. The workflows are designed to support the same three use cases: cost guideline question answering, budget validation, and budget recommendation. The main difference between the workflows lies in how each workflow selects sources, organizes retrieval, and generates the final answer. The workflow characteristics are summarized in Table III compares the three workflow designs using the same data sources. This keeps the evaluation focused on workflow behaviour instead of differences in available evidence. Table III. Implemented Workflow Characteristics 3.5.1 Single-Agent RAG Architecture Single-Agent RAG is used as the baseline workflow. As illustrated in Fig. 3, the user query is handled by a single agent that performs source selection, evidence retrieval, and answer generation in one main process. The agent has access to the available knowledge sources, including structured budget data and narrative documents. This workflow is simpler and easier to control, but its performance depends heavily on the ability of one agent to select the most relevant evidence for different types of questions. Figure 3 shows the baseline workflow where one agent handles source selection, retrieval, and answer generation. This design is simpler and faster, but it has less explicit control over use-case-specific retrieval. 3.5.2 Routing-Based Agentic RAG Architecture Routing-Based Agentic RAG introduces a router before retrieval and answer generation. As shown in Fig. 4, the router first classifies the user query into one of the predefined use cases: UC1 for cost guideline question answering, UC2 for budget validation, or UC3 for budget recommendation. After the query is routed, the corresponding path performs retrieval and response generation using the sources and prompt structure that are more suitable for that use case. This workflow is designed to test whether use-case-based routing can improve source selection, evidence coverage, and answer relevance. Figure 4 shows how the router selects a path for UC1, UC2, or UC3 before retrieval and answer generation. This helps the system use a more suitable retrieval strategy for each budgeting task. 3.5.3 Orchestrator-Workers RAG Architecture Orchestrator-Workers Agentic RAG applies a delegation-based workflow. As illustrated in Fig. 5, an orchestrator analyses the user query, determines the types of evidence required, and assigns subtasks to specialized workers. The workers may retrieve guideline evidence, historical realization evidence, or narrative context depending on the query needs. Their partial outputs are then combined by a synthesizer to produce the final answer. This workflow is intended to support questions that require analysis across multiple evidence types, although it may introduce additional coordination complexity and processing time. Figure 5 shows the delegation-based workflow. The orchestrator assigns evidence retrieval and analysis to specialized workers, then the final response is synthesized from their outputs. 3.6 Evaluation Design The evaluation uses one general dataset consisting of 30 test cases covering UC1, UC2, and UC3. Each test case is executed on the three implemented workflows, resulting in 90 workflow executions. All executions are completed without errors, indicating that the workflows can run stably on the evaluation dataset. The workflows are evaluated using four RAGAS-based metrics: context precision, context recall, faithfulness, and answer relevancy. Context precision is calculated using K = 10, which evaluates the top ten evidence contexts submitted to the evaluation script. This value is different from the retrieval top-K used inside the workflows. The knowledge retrieval uses top-K = 8, the internal retrieval in Routing-Based Agentic RAG and Orchestrator-Workers Agentic RAG uses top-K = 4, and the internal retriever in Single-Agent RAG uses top-K = 10.The same ground-truth dataset is used across all workflows to ensure a fair comparison. The main comparison is based on the aggregate result across 30 test cases, while the per-use-case result is used as supporting diagnostic analysis. 4 Results and Discussion 4.1 Comparative Evaluation The comparative evaluation result is presented in Table IV. The highlighted cells indicate the best score for each metric. When two workflows obtain the same best value, both cells are highlighted. Overall, the results show that each workflow has different strengths, and no workflow dominates all evaluation metrics. Table IV. Comparative Evaluation Result on 30 Test Cases Based on Table IV summarizes the average evaluation results across 30 test cases. The highlighted cells mark the best score for each metric, showing that Routing-Based Agentic RAG provides the strongest retrieval coverage and faithfulness, while Single-Agent RAG is fastest. Routing-Based Agentic RAG achieves the strongest overall RAG quality among the evaluated workflows. It obtains the highest context recall score of 0.873 and the highest faithfulness score of 0.878. It also shares the highest context precision score of 0.731 with Orchestrator-Workers Agentic RAG. These results indicate that the routing-based workflow is able to retrieve relevant evidence, cover more supporting information required by the reference answer, and generate responses that remain strongly grounded in the retrieved sources. Orchestrator-Workers Agentic RAG also shows strong retrieval performance. It obtains a context precision score of 0.731, equal to Routing-Based Agentic RAG, and a context recall score of 0.866. This indicates that delegating retrieval and analysis to specialized workers can help structure the evidence collection process. However, this workflow does not achieve the highest recall or faithfulness, and it produces the highest latency at 86.659 seconds. The result suggests that additional coordination and synthesis steps may improve task decomposition, but they also increase response time and do not automatically improve the final answer quality. Single-Agent RAG achieves the highest answer relevancy score of 0.775 and the lowest latency of 26.763 seconds. This result indicates that a simpler workflow can produce responses that are more direct and faster. However, its context precision score of 0.640 and context recall score of 0.823 are lower than those of the agentic workflows. This suggests that Single-Agent RAG is efficient and responsive, but it has weaker retrieval control and evidence coverage compared with workflows that explicitly separate routing or worker-based retrieval. Based on the aggregate results, Routing-Based Agentic RAG is considered the most suitable workflow among the evaluated alternatives. Although Single-Agent RAG is superior in answer relevancy and latency, Routing-Based Agentic RAG provides stronger evidence retrieval and factual grounding. These aspects are more important for the budget validation task because the system must not only answer user questions fluently but also support each claim with relevant and traceable evidence. 4.2 Discussion The evaluation shows that no workflow dominates all metrics. Each workflow provides different strengths depending on the type of task and the complexity of evidence required. In this study, faithfulness is treated as a minimum requirement because budget validation responses should not generate unsupported cost claims. A response may appear relevant to the user’s question, but it still needs to be grounded in retrieved evidence, especially when discussing cost standards, budget reasonableness, or possible mark up. After faithfulness is sufficiently met, context recall and answer relevancy become important for determining the most suitable workflow. Budget validation does not only require a correct final statement, but also complete supporting evidence from cost guidelines, budget records, historical realization data, and narrative documents. Therefore, a workflow with better evidence coverage and better alignment with the user’s intent is more suitable for this system. Single-Agent RAG remains useful as a fast and responsive baseline. Its simpler structure allows the system to generate direct answers with lower workflow complexity. However, because one agent handles source selection, retrieval, and answer generation, the workflow has weaker control when the request requires different evidence types. This limitation becomes more visible in budget validation and recommendation tasks, where the system may need to combine guideline values, historical records, uploaded RAB items, and narrative explanations. Routing-Based Agentic RAG is selected as the main workflow because it provides the best balance between evidence coverage, response relevance, and workflow controllability. The router separates cost guideline questions, budget validation requests, and budget recommendation queries into different paths. This allows the system to apply a more suitable retrieval strategy and answer structure for each use case. The evaluation results support this design because Routing-Based Agentic RAG achieves the highest context recall and faithfulness. Higher context recall indicates that the workflow retrieves more complete supporting evidence, while higher faithfulness indicates that the generated response remains grounded in that evidence. Although Single-Agent RAG attains higher answer relevancy, relevance alone is not sufficient for budget validation, where every claim must be traceable to a retrieved source. Orchestrator-Workers Agentic RAG demonstrates that multi-agent decomposition can support more structured evidence processing. The workflow allows different workers to focus on different evidence sources, such as cost guidelines, historical realization data, and narrative documents. However, the additional coordination between the orchestrator, workers, and synthesizer increases workflow complexity and latency. The result suggests that increasing the number of agents does not automatically improve overall performance. The workflow may still be useful for more complex scenarios, but it requires better worker activation, evidence filtering, and synthesis mechanisms. Overall, the results indicate that Agentic RAG with hybrid retrieval is a feasible approach for supporting budget validation. Structured retrieval helps preserve numerical accuracy for cost guideline items, budget components, and historical realization values, while semantic retrieval provides contextual explanations from narrative documents. This combination allows the system to generate evidence-grounded responses and identify budget items that may require further review when relevant comparison sources are available. 4.3 Routing-Based RAG Case Examples The selected workflow is further demonstrated through three case examples: cost guideline question answering (UC1), budget validation (UC2), and budget recommendation (UC3). These examples show how the same user interface can support different budget-related tasks while the internal processing path is adapted by the routing mechanism. This design is important because each use case requires different evidence priorities. A cost guideline question requires precise guideline evidence, a budget validation request requires comparison between uploaded budget items and reference data, while a recommendation request requires synthesis of relevant budget components. In general, the Routing-Based Agentic RAG workflow follows five main stages. First, the user submits a question or uploads a budget file through the interface. Second, the Router Agent identifies the user intent and assigns the request to the appropriate route. Third, the system retrieves evidence from the relational database and vector database based on the selected route. Fourth, the retrieved evidence is filtered and arranged into an evidence packet. Finally, the Final Analyst Agent generates the response based on the selected evidence and returns the answer with supporting sources. Figure 6 shows the UC1 case, where the user asks about cost guideline information. In this use case, the main objective is to retrieve the most relevant cost reference from official guideline data, such as PMK/SBM, INKINDO, or ITB cost standards. Therefore, the system prioritizes structured guideline data stored in the relational database. Narrative evidence from the vector database is only used as supporting context when needed. After the question is submitted, the Router Agent classifies the request as a cost guideline question because the user asks for a specific cost reference and does not request budget validation or recommendation. The workflow then retrieves relevant rows from the guideline data by considering the item name, source, unit, amount, and source metadata. The final response presents the requested guideline information in natural language and includes supporting sources so that the answer can be traced back to the retrieved evidence. Figure 7 shows the UC2 case, where the user uploads a budget file and asks the system to validate the proposed budget items. In this use case, the uploaded budget file becomes the target object of validation. The system compares the proposed items with available cost guidelines, historical realization data, or other relevant references. The workflow extracts budget rows from the uploaded file and stores the related information in structured tables, such as item description, quantity, unit, unit price, total amount, and category. The Router Agent classifies the request as a budget validation task. The system then retrieves the uploaded budget items, searches for matching guideline references, and retrieves historical or realization data when relevant. The validation output is generated at the item level. The system compares the uploaded budget value with the available reference and classifies the result based on the evidence. Items can be marked as aligned with the available reference, potentially exceeding the reference, or not quantitatively judgeable due to insufficient evidence. This distinction is important because not all budget items have a direct official guideline. For items such as venue rental, goodie bag, documentation, VIP dinner, or contingency, the system should not force an unrelated comparison. Instead, it should state that the item cannot be judged quantitatively from the available references and recommend additional supporting evidence, such as vendor quotations, scope details, or internal justification. This case shows that the uploaded RAB is treated as the object being evaluated, while guideline and historical data are used as comparison sources. Figure 8 shows the UC3 case, where the user asks for a budget recommendation for a planned activity, such as a small seminar. Unlike UC2, this use case does not validate an existing budget file. Instead, the system helps the user prepare an initial budget structure based on available cost references and supporting context. After receiving the request, the Router Agent classifies it as a budget recommendation task. The workflow then retrieves relevant budget components from official guidelines, historical budget data, and narrative context. For a seminar case, the retrieved components may include speaker honorarium, moderator honorarium, master of ceremony, consumption, snacks, transportation, documentation, or other event-related items. The recommendation output includes suggested budget components, source-based rates when available, units, calculation assumptions, and limitations. For example, if a guideline provides an honorarium rate per person-hour, the system can recommend the component while still stating that the final amount depends on the number of speakers and session duration. For components that do not have a direct guideline, the system can recommend obtaining vendor quotations or adding internal justification. This case demonstrates that the system supports recommendation tasks as an initial planning aid, not as a final budget decision. Overall, these case examples show that the routing mechanism helps the system adapt the retrieval and response process based on the user’s intent. Without routing, every request would be processed using the same retrieval strategy, which may retrieve irrelevant evidence or apply an unsuitable answer structure. By separating guideline questions, validation requests, and recommendation requests into different paths, the system can provide responses that are more aligned with the user’s need and more traceable to the supporting evidence. 5 Conclusion and Future Work This study developed and evaluated an Agentic Retrieval-Augmented Generation system with hybrid retrieval for budget planning and validation. The system was designed to support source-grounded answers for cost guideline questions, budget validation, and budget recommendation. Structured data such as cost guideline items, budget plans, and historical realization records are stored in a relational database, while narrative documents and contextual knowledge are stored in a vector database. This separation allows the system to retrieve numerical and tabular information more precisely while still using semantic retrieval to obtain relevant narrative evidence. The evaluation shows that each workflow has different strengths. Single-Agent RAG provides a simple and responsive baseline. Routing-Based Agentic RAG provides better control by directing different user requests to more suitable retrieval and reasoning paths. Orchestrator-Workers RAG supports a more flexible multi-agent process, but its additional coordination does not always improve the final answer quality. Overall, Routing-Based Agentic RAG is selected as the most suitable workflow for the proposed system because it provides a better balance between evidence coverage, response relevance, and workflow controllability. 5.1 Single-Agent RAG Single-Agent RAG is the simplest workflow in this study. In this workflow, one agent handles source selection, evidence retrieval, and answer generation. This makes the workflow easier to implement and more responsive compared with more complex agentic workflows. The evaluation results show that Single-Agent RAG can produce focused and direct answers, making it useful as a lightweight baseline for budget-related question answering. However, Single-Agent RAG has limitations when the user request requires different types of evidence. Budget validation may require guideline values, historical budget records, uploaded RAB items, and narrative explanations at the same time. Since all retrieval and reasoning decisions are handled by one agent, the workflow has weaker control over which evidence source should be prioritized. As a result, Single-Agent RAG is more suitable for simpler lookup questions than complex validation or recommendation tasks. 5.2 Routing-Based Agentic RAG Routing-Based Agentic RAG is selected as the most suitable workflow for the proposed system. This workflow uses a router to classify the user request before retrieval and answer generation. The request can be directed into different paths, such as cost guideline question answering, budget validation, or budget recommendation. By separating these paths, the system can use a more appropriate retrieval strategy and answer structure for each use case. This design is important because budget validation does not always require the same type of evidence. A cost guideline question may prioritize formal cost standards, while a budget validation request may require comparison between proposed budget items, guideline values, and historical realization data. A recommendation request may require broader supporting context. The evaluation results support this design because Routing-Based Agentic RAG achieves stronger evidence coverage and answer relevance compared with the other workflows. Therefore, the routing-based workflow provides the best balance between agentic capability and controllability for the proposed system. 5.3 Orchestrator-Workers RAG Orchestrator-Workers RAG provides the most complex workflow among the three evaluated approaches. In this workflow, an orchestrator analyzes the user request, assigns subtasks to worker agents, and combines the worker outputs into a final response. This structure is conceptually useful for multi-source analysis because different workers can focus on different types of evidence, such as guideline data, historical realization data, or narrative documents. However, the evaluation shows that adding more agents does not automatically improve the final answer quality. The coordination between the orchestrator, workers, and synthesizer increases workflow complexity and latency. In some cases, the final response may become less focused because the system needs to combine several worker outputs. Although this workflow remains promising for more complex budget analysis, it requires better worker activation control, evidence filtering, and synthesis mechanisms before it can outperform the routing-based workflow. 5.4 Limitations and Future Work This study has several limitations. First, the evaluation uses a limited number of test cases, so the results may not fully represent all possible budgeting scenarios. Future work should expand the evaluation dataset by including more diverse budget documents, different RAB formats, and more complex validation cases. Evaluation with finance experts or domain users should also be conducted to assess whether the system output is useful in real decision-making contexts. Second, the system depends on the quality and completeness of the available data. If the cost guidelines, RAB data, historical records, or narrative documents are incomplete, the generated answer may also be limited. Future work should improve document parsing, data validation, metadata management, and source prioritization so that the system can retrieve more accurate and relevant evidence. Third, the current system is designed as a decision-support tool, not as a final decision maker. Indications of unreasonable costs or potential mark up still require user review and domain judgment. Future development should provide clearer validation explanations, such as structured comparison tables between proposed budget values, guideline values, historical values, differences, validation status, and supporting sources. Finally, the workflow and system performance can still be improved. Retrieval can be enhanced using better ranking, filtering, and source selection strategies. Latency can be reduced by limiting unnecessary LLM calls, applying caching, and separating deterministic data retrieval from response generation. For production use, the system should also include authentication, access control, audit logs, and stronger protection for sensitive budget and internal documents. Competing interests The authors have no relevant financial or non-financial interests to disclose. All authors certify that they have no affiliations with or involvement in any organization or entity with any financial or non-financial interest in the subject matter or materials discussed in this manuscript. Ethics approval This study did not involve human participants, their identifiable data, or biological material, nor did it involve animals. The data used in this study consist of institutional budget plans, formal cost-standard references (PMK/SBM, INKINDO, and ITB cost standards), and project realization records. Therefore, approval from a research ethics committee was not required. Consent to participate Not applicable. This study did not involve human participants; therefore, informed consent to participate was not required. Consent to publish Not applicable. This manuscript does not contain any individual person’s identifiable data, images, or personal details; therefore, consent to publish was not required. Funding This research was funded by Hibah P2MI ITB (Project ID: STEI.PPMI-1-47-2026). D.W.A., M.R.H., M.S.M.R., S.H.S., L.P.Y., and A.A.K. contributed to the study conception and design. D.W.A., M.R.H., and A.A.K. carried out the system design, implementation, data preparation, and experiments. M.S.M.R., S.H.S., and L.P.Y. provided methodology guidance, supervision, and critical review of the manuscript. D.W.A. wrote the first draft of the manuscript, and all authors commented on previous versions. All authors read and approved the final manuscript. Acknowledgement The authors would like to thank Institut Teknologi Bandung for the support provided through the P2MI ITB research grant (Project ID: STEI.PPMI-1-47-2026). The authors also gratefully acknowledge the AI & Cyber Author Contribution Security for Smart X (AISX) Research Collaboration Center, ITB, for the facilities and research environment that supported this work Data Availability. The budget plan and project realization datasets that support the findings of this study are available from Institut Teknologi Bandung, but restrictions apply to the availability of these data, which were used under permission for the current study and so are not publicly available. The data are, however, available upon reasonable request and with the permission of Institut Teknologi Bandung. The formal cost-standard references used in this study (PMK/SBM, INKINDO, and the ITB cost standards) are publicly available from their respective issuing institutions and are cited in the manuscript. Footnotes 1. School of Electrical Engineering and Informatics, Institut Teknologi Bandung, Bandung, Indonesia. 2. Artificial Intelligence & Cyber Security for Smart X (AISX) Research Collaboration Center, Institut Teknologi. Bandung, Bandung, Indonesia 3. School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia, Johor, Malaysia. 4. UTM-IRDA Digital Media Centre, Media and Game Innovation of Excellence (MaGICX), Institute of Human. Centered Engineering (iHumEn), Universiti Teknologi Malaysia, Johor, Malaysia