Source-linked AI summary
AgriIR: A Scalable Framework for Domain-Specific Knowledge Retrieval
Shuvam Banerji Seal, Aheli Poddar, Alok Mishra, Dwaipayan Roy
TL;DR
Agricultural information access requires grounded, domain-specific answers without the infrastructure burden and reliability problems associated with large general-purpose models. AgriIR addresses this gap with a configurable, modular RAG pipeline using lightweight models, adaptive retrieval, domain agents, and deterministic citations. On 191 agricultural policy queries, its 27B variant achieved statistical parity with ChatGPT-4o and significantly outperformed GPT-OSS-120B while providing verifiable citations.
Problem
Agricultural RAG deployment is constrained by large-model resource requirements, domain drift, and unreliable outputs without trustworthy grounding.
Method
AgriIR decomposes information access into configurable stages and combines lightweight models with adaptive multi-source retrieval, domain specialization, and deterministic citation.
Results
AgriIR_Gemma3:27B achieved statistical parity with ChatGPT-4o on composite scores (0.820 ± 0.208 vs 0.840 ± 0.233, p = 0.493) and outperformed GPT-OSS-120B (∆= 0.115, p < 0.001, d = 0.45).
Takeaways & Limitations
The results support domain specialization and multi-agent retrieval as an alternative to brute-force parameter scaling for trustworthy agricultural information access.
Takeaways & Limitations
The domain enhancement mechanism requires continuous maintenance because outdated keyword lists or agent configurations can degrade retrieval performance and coverage.
Abstract
from arXiv · showhide
This paper introduces AgriIR, a configurable retrieval augmented generation (RAG) framework designed to deliver grounded, domain-specific answers while maintaining flexibility and low computational cost. Instead of relying on large, monolithic models, AgriIR decomposes the information access process into declarative modular stages -- query refinement, sub-query planning, retrieval, synthesis, and evaluation. This design allows practitioners to adapt the framework to new knowledge verticals without modifying the architecture. Our reference implementation targets Indian agricultural information access, integrating 1B-parameter language models with adaptive retrievers and domain-aware agent catalogues. The system enforces deterministic citation, integrates telemetry for transparency, and includes automated deployment assets to ensure auditable, reproducible operation. By emphasizing architectural design and modular control, AgriIR demonstrates that well-engineered pipelines can achieve domain-accurate, trustworthy retrieval even under constrained resources. We argue that this approach exemplifies ``AI for Agriculture'' by promoting accessibility, sustainability, and accountability in retrieval-augmented generation systems.
1 Introduction
AgriIR addresses unreliable and resource-intensive agricultural information access with a configurable RAG architecture. Its modular, domain-adaptive design uses smaller models while emphasizing groundedness, transparency, and reuse.
- Motivation: Agricultural information access remains difficult despite the sector’s importance, leaving farmers, extension officers, and policymakers without timely, contextually relevant knowledge.Agriculture employed approximately 916 million people worldwide in 2023, while Indian agriculture supports 58% of the rural population and contributes roughly 18% to national GDP.
- Motivation: General-purpose LLMs are constrained by high resource requirements, agricultural domain drift, and unreliable outputs without explicit grounding.These limitations are especially consequential for agricultural decision-making involving crop management, irrigation, and climate resilience.
- Motivation: Existing RAG systems provide grounded responses but commonly depend on large databases, extensive datasets, and GPU-heavy infrastructure unsuitable for many low-resource deployments.The paper identifies agriculture, public health, and climate adaptation as domains where such deployment constraints remain significant.
- Framework: AgriIR decomposes information access into declarative stages for query refinement, decomposition, retrieval, synthesis, and evaluation, enabling component substitution without runtime re-engineering.Temperature controls, pluggable retrieval, and configurable domain artifacts support behavioral tuning and rapid vertical retargeting without code changes or fine-tuning.
- Framework: AgriIR uses architectural intelligence rather than parameter scale to support reliable, evidence-grounded responses with 1B-parameter models and transparent configurable interfaces.The framework targets Indian agriculture while presenting reconfiguration as a path toward domains such as health and climate, alongside reproducibility, accountability, and reusability.
- Reproducibility: The implementation is publicly available, supporting reproducibility and further research.A video demonstration is also provided as a system walkthrough.
2 Related Work
Related work shows that agricultural RAG systems improve domain grounding but often remain dependent on large models, static resources, or manual curation. AgriIR extends efficiency-oriented architectural design into configurable, multi-source agricultural retrieval.
- Domain-Specific IR and RAG Systems: General-purpose RAG systems typically combine dense retrieval with 7B–70B-parameter models, creating deployment challenges in resource-constrained agricultural settings.Agricultural applications such as crop disease diagnosis and fertilizer recommendation demonstrate RAG’s potential but commonly rely on large models and static knowledge bases.
- Domain-Specific IR and RAG Systems: Specialized systems such as ShizishanGPT integrate search, knowledge graphs, retrieval, and tools, illustrating the value of modular agricultural architectures.The described components combine general query handling, real-time updates, structured facts, and RAG-based supplementation.
- Citation Reliability: AgriIR addresses citation reliability through post-hoc sentence-similarity matching that operates independently of LLM generation.The related work reports incorrect citations in 40–60% of GPT-3.5/GPT-4 cases, motivating independent citation insertion.
- Model Efficiency and Architectural Design: Architectural techniques such as chain-of-thought prompting, temperature control, and specialized fine-tuning are presented as ways to improve smaller-model performance beyond model size alone.This motivates AgriIR’s emphasis on task decomposition and temperature stratification.
- Model Efficiency and Architectural Design: Prior work reports that DVFS can reduce LLM energy consumption by up to 30% while preserving model quality.The cited studies analyze energy-performance trade-offs across models and task input characteristics.
- Agricultural Knowledge Systems: Traditional agricultural systems provide valuable content but often lack natural-language flexibility, real-time integration, or scalable semantic curation.AgriIR is positioned as combining structured agricultural databases with real-time web information through a resource-efficient configurable architecture.
3 AgriIR: An IR System for Agricultural Knowledge Access
AgriIR implements agricultural knowledge access as a configurable six-stage pipeline combining lightweight models, adaptive multi-source retrieval, domain agents, and deterministic citation. Its architecture supports modular substitution, parallel retrieval, agentic curation, and auditable synthesis.
- Architecture: AgriIR structures retrieval and synthesis as a declarative pipeline whose stages are controlled through configuration rather than code.The architecture is designed for efficient use in resource-limited settings and can be extended without altering the overall runtime architecture.
- Pipeline Workflow: The pipeline refines a raw query, decomposes it into sub-queries, retrieves evidence in parallel, enhances sub-queries with domain agents, synthesizes an answer, and inserts citations.The algorithm defines six sequential stages with independently configurable components and stage-specific temperature regimes.
- Pipeline Workflow: Query refinement uses a 1B-parameter model at temperature 0.1, while sub-query decomposition uses the same model at temperature 0.5 to introduce controlled diversity.Decomposition separates multifaceted agricultural questions into perspectives such as soil management, pest control, irrigation, and policy support.
- Retrieval: Each sub-query retrieves from structured databases and domain-constrained web sources using adaptive embeddings, ranked candidate selection, and parsed articles.Database retrieval returns top-k passages with metadata, while web retrieval selects and processes the top five articles.
- Retrieval: Parallel retrieval reduced latency from approximately 180 seconds sequentially to approximately 50 seconds for a four-subquery workload in testing.Retrieved database and web passages are unified across sub-queries before synthesis.
- Domain-Agent Enhancement: Domain agents are selected by keyword-overlap scores and append contextual domain keywords to each sub-query’s retrieval scope.The registry includes roles such as crop specialist, soil expert, pest manager, and sustainability advisor.
- Answer Synthesis: Synthesis selects between a 1B model and a higher-parameter model according to query type, using temperature 0.2 to balance factual precision and readability.The generator integrates evidence from retrieved documents into the final response.
- Citation Enforcement: Deterministic citation insertion embeds answer sentences, compares them with retrieved chunks using cosine similarity, and appends source identifiers when similarity exceeds 0.75.Multiple citations are added when a sentence synthesizes several sources, producing a verifiable answer with source metadata.
4 Results
AgriIR is evaluated on 191 agricultural queries annotated by 30 human evaluators using answer quality, citation quality, and a composite score. The evaluation examines model configurations, retrieval augmentation, and statistical differences under constrained resources.
- Evaluation setup: 191 agricultural queries were manually annotated by 30 evaluators with relevant domain backgrounds.Each question-answer pair was assigned to three annotators, who graded answer and citation satisfaction.
- Evaluation setup: The evaluation compares AgriIR configurations using Llama3.2:3B and Gemma3 models, with and without database integration, against open and commercial baselines.The comparison assesses answer quality, citation quality, and system efficiency.
- Metrics: The composite performance score combines answer and citation quality through a linear combination with λ = 0.7.The weighting was selected after preliminary experimentation on a heldout subset to balance factual correctness and citation grounding.
- Performance results: AgriIR_Gemma-3-27B achieves a composite score of 0.820 ± 0.208, statistically equivalent to ChatGPT-4o at 0.840 ± 0.233 (p = 0.493).The same model significantly outperforms Gemini 1.5 Flash at 0.779 ± 0.250 and GPT-OSS-120B at 0.705 ± 0.246 (p < 0.001).
- Performance results: AgriIR models achieve 73-84% perfect citations, whereas baseline models lack citation tracking.Table 1 reports citation quality separately because baseline models do not support citation tracking.
5 Conclusion and Future Work
AgriIR combines deterministic citation, intelligent multi-phase retrieval, and autonomous domain knowledge acquisition for agricultural information access. Its evaluation reports parity with ChatGPT-4o and verifiable citations, while future work targets multimodal data, maintenance, causal reasoning, privacy, and personalization.
- Conclusion: AgriIR uses sentence-level semantic similarity for deterministic citation tracking, achieving 59-73% perfect citation accuracy.The mechanism operates independently of language-model generation and measures semantic overlap between generated sentences and retrieved sources.
- Conclusion: Its multi-phase web retrieval combines multistrategy candidate gathering, LLM-based article selection, and comprehensive content extraction.The stated goal is to reduce retrieval noise while prioritizing authoritative agricultural sources.
- Conclusion: Specialized agents with persistent duplicate tracking collected 15,247 agricultural entries without manual curation.This addresses knowledge staleness in static RAG systems through autonomous knowledge acquisition.
- Conclusion: On 191 agricultural policy queries, AgriIR_Gemma3:27B reaches statistical parity with ChatGPT-4o and outperforms GPT-OSS-120B despite using 4.4 fewer parameters.Composite scores are 0.820 ± 0.208 versus 0.840 ± 0.233 for ChatGPT-4o, with p = 0.493; the GPT-OSS comparison reports ∆ = 0.115 and p < 0.001.
- Future work: Future extensions include multimodal retrieval, causal reasoning, federated learning, citation-graph analytics, and personalization by region, crop type, or farm size.The proposed modalities include satellite imagery, soil moisture, IoT sensor data, and visual question answering.
- Future work: Domain-agent enhancement requires continuous maintenance because outdated keyword lists or agent configurations can introduce inconsistencies and degrade retrieval performance.The stated maintenance boundary concerns domain-specific keyword lists and agent configurations.
- Conclusion: The framework’s principles of deterministic citation, domain grounding, and autonomous data acquisition are presented as applicable beyond agriculture to healthcare, law, and education.The broader scope is tied to settings where trust and verifiability are essential.