Pith. sign in

REVIEW 5 major objections 5 minor 16 references

Vision Meets Language: A RAG-Augmented YOLOv8 Framework for Coffee Disease Diagnosis and Farmer Assistance

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that chaining YOLOv8 leaf-disease detection with retrieval-augmented generation can give farmers reference-grounded diagnoses and remedies, and reports mAP@0.5 of 0.681 on a coffee-leaf validation set.

desk verdict A modest engineering integration of YOLOv8+RAG+LLM for coffee leaf disease that has a real detection table but a load-bearing dataset contradiction; fix the split ambiguity and it could be a useful applied paper. read the letter →

arxiv 2505.21544 v1 pith:GHIZDXL3 submitted 2025-05-24 cs.CV cs.CL

Semanto Mondal This is my paper
classification cs.CVcs.CL
keywords LLMRAGNLPYOLOv8PrecisionAgricultureCoffeeleafdiseasedetectionRetrieval-AugmentedGeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to combine computer vision and language generation into one agricultural assistant: a YOLOv8 model detects coffee leaf diseases from a photo, a retrieval-augmented generation (RAG) pipeline pulls matching passages from a curated knowledge base, and a large language model turns both into a plain-language explanation with remedies. The author argues that chaining these three components addresses what neither part can do alone, because detectors label regions but cannot explain why or what to do, while language models answer fluently but hallucinate without external grounding. On the detection side, the paper reports an overall mAP@0.5 of 0.681 on a validation set of 228 images, with the best class (leaf miner) reaching 0.894. If the approach holds, a farmer could upload a leaf photo and receive a diagnosis plus source-referenced treatment suggestions in one interface.

What carries the argument

The load-bearing mechanism is the retrieval-then-read chain. A YOLOv8 detector first emits predicted disease labels and bounding boxes from the uploaded image; each detected label is then turned into a query against a vector store built by chunking and embedding a hand-written knowledge base of coffee leaf diseases and remedies. The top retrieved passages are appended to the user's prompt, and a large language model generates the final explanation and treatment advice from that augmented context. A short-term conversational memory keeps follow-up questions in context, so the system functions as a dialogue rather than a one-shot classifier.

What would settle it

Re-train YOLOv8n under each of the two BRACOL dataset versions described in Section 5, evaluate both on the same held-out 228-image, 402-instance validation set, and check whether either configuration reproduces the reported overall mAP@0.5 of 0.681; separately, compare remedy answers generated with retrieval enabled versus disabled and count unsupported statements to test the hallucination-reduction claim.

Watch

Extended reading notes

Core claim

The central claim is that a three-part pipeline called the RAG-Augmented YOLOv8 Framework can identify coffee leaf diseases and generate explainable, reference-backed remedies in a single system. The author presents the design as neurosymbolic: YOLOv8 and the large language model are the neural components, while an external knowledge base of disease symptoms, causes, and treatments is the symbolic component that grounds the language output. The paper states that this retrieval-then-read mechanism mitigates the hallucination problem of large language models and enables fine-grained text output and conversational follow-up questions. It reports that the YOLOv8n detector, trained for 100 epochs at 640 by 640 resolution, achieves an overall mAP@0.5 of 0.681 and an overall mAP@0.5:0.95 of 0.454 on the validation set, with the class imbalance appearing to hurt rust detection most.

Load-bearing premise

The evaluation's validity rests on knowing which version of the BRACOL dataset produced Table 1, because Section 5.1 says the original 1,899-annotation version trained best while Section 7.1 reports results after training on the re-annotated 8,226-annotation version, and the paper never resolves this discrepancy.

Editorial extensions

If this is right

  • If the reported metrics are reproducible, YOLOv8n at 640 by 640 resolution is enough for usable coffee leaf disease detection on a consumer GPU, making on-farm deployment plausible.
  • If RAG grounding works as claimed, the remedy text will track the curated knowledge base, so updating that document updates the advice without retraining the detector or the language model.
  • The same three-part chain can be re-targeted to other crops by swapping the detector's training data and the knowledge-base document.
  • The conversational memory component turns a one-shot diagnosis into a follow-up dialogue, which is the intended path toward adaptive treatment plans.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension would be to measure the RAG component's effect directly by running the same user questions with and without retrieved passages and counting unsupported statements; the paper describes the mechanism but reports no such evaluation.
  • Because the knowledge base is a single hand-written document, the quality of the generated advice is bounded by that document's coverage, so a formal comparison against expert-annotated remedy pairs would show whether the hallucination-mitigation claim holds in practice.
  • The class imbalance in the curated dataset, with rust far outnumbering cercospora, suggests the reported per-class mAP would shift under a rebalanced validation set, and severity estimation, a natural next step for farmer assistance, is not addressed.
  • If the detector's false negatives correlate with visual similarity to healthy leaf texture, as the paper says for rust, then a healthy-leaf rejection stage before the LLM query could reduce wasted retrievals and improve trust in the assistant.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript describes a coffee-leaf disease diagnosis system that combines a fine-tuned YOLOv8n detector with a Retrieval-Augmented Generation (RAG) pipeline and an LLM (accessed through the Groq API). The detection module is evaluated on the BRACOL dataset, reporting an overall mAP@0.5 of 0.681 on a validation set. The language module is illustrated through a Streamlit interface with example disease explanations, remedy suggestions, and follow-up question handling. The paper's stated contribution is an end-to-end tool that grounds LLM responses in a domain-specific knowledge base in order to reduce hallucination and provide actionable, reference-backed agricultural advice.

Significance. If the detection numbers are reliable and the RAG component demonstrably improves answer quality, the system would be a useful practical integration for precision agriculture. The authors provide a transparent account of the training hyperparameters and use a public benchmark dataset (BRACOL), which helps reproducibility. The class-wise detection metrics in Table 1 are a reasonable starting point for assessing the YOLOv8n component. However, at present the evidence for the central claims is incomplete: the detection evaluation is ambiguous because of the dataset-version contradiction, the RAG/LLM component is evaluated only through qualitative screenshots, no comparison is made against a non-RAG baseline or against existing integrated systems such as FarmTalk-Nexus, and the knowledge base is entirely author-created. These gaps prevent the paper from currently supporting its stated claims of hallucination mitigation and satisfactory integrated performance.

major comments (5)
  1. [§5.1, §6, §7.1, Table 1] Section 5.1 states that the best performance was obtained when the model was trained using the original version of the BRACOL dataset, whereas Section 7.1 reports all validation results "after training the YOLOv8n model on the curated and re-annotated BRACOL dataset." These are two different annotation regimes (1,899 vs. 8,226 annotations). Section 6 does not define the exact training/validation split or state which dataset version produced the numbers in Table 1. Because every detection metric in Table 1 inherits this ambiguity, the central quantitative result of the paper is currently uninterpretable. The authors must state which dataset version, which split, and which annotation counts were used for the numbers in Table 1, and, if both versions were used, report results separately.
  2. [§7.1 vs. §5.2] The sentence "it is evident that classes with a higher number of representative annotations (Miner and Phoma) yield better results" is contradicted by the authors' own dataset statistics in Section 5.2: Rust has the highest annotation count (6,013) yet the lowest mAP@0.5 (0.415), while Miner has 341 annotations (fewer than Phoma's 1,671 and close to Cercospora's 201) yet the highest mAP@0.5 (0.894). The explanation in terms of annotation count cannot be correct as stated. The authors should either remove this claim or provide a different, evidence-based explanation, for example involving visual similarity to healthy tissue, bounding-box size, or annotation quality.
  3. [§7.2, §4.2–4.3] The RAG/LLM component is not evaluated at all. Section 7.2 presents screenshots and example interactions, but there are no quantitative or structured qualitative measurements of retrieval accuracy, answer faithfulness, hallucination rate, or user satisfaction, and no comparison with the same LLM without RAG. The abstract and Section 1 claim that the system "mitigate[s] the hallucination issue of LLM," but the paper provides no evidence for this claim. At minimum, the authors should include a small evaluation set with expert-validated gold answers, report retrieval hit rates and answer-groundedness scores, and compare RAG vs. no-RAG outputs.
  4. [§4.2] The knowledge base was created by the authors themselves ("We created a detailed document that serves as a knowledge base..."), and no external sources, citations, or expert validation are given for its content. This makes the "grounded" answers self-referential in the sense that the LLM is constrained only by text that the authors wrote. To support the claim of factual grounding, the authors should cite the agricultural sources used to build the knowledge base and/or have the content reviewed by a plant pathology expert.
  5. [§2, §7.1] The paper does not compare the proposed YOLOv8n model with other detectors or with the BRACOL baselines reported in the literature, including the YOLO-based results in Reference [16]. Adding such a comparison would contextualize the reported 0.681 mAP@0.5 and support the claim of satisfactory performance. Without a baseline, the reader cannot assess whether the proposed system improves over readily available alternatives.
minor comments (5)
  1. [Abstract] The Abstract contains wording such as "model AI" that should be "AI models," and the opening sentence about humans as social beings is tangential to the technical content; the abstract would be stronger if it moved directly to the problem and contribution.
  2. [Throughout] The notation "YOLOV8" appears in the keywords and in some headings, while the body uses "YOLOv8"; please use a single consistent spelling.
  3. [§4.2, §7.2] The text writes "F AISS" with a space; this should be "FAISS".
  4. [§1] The phrase "most affection one" should be "most favored one" or "most popular one."
  5. [References] Reference [2] is malformed ("al., L.C.C.:") and should be completed with the author names and report title; Reference [9] is a GitHub link with no version or access date, which makes it difficult to verify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the YOLOv8 detection result is an independent benchmark measurement and the RAG pipeline is a system design rather than a derivation that reduces to its inputs.

full rationale

The paper's central detection claim is the YOLOv8n mAP@0.5 of 0.681 on the BRACOL validation set (Section 7.1, Table 1). That number is a measured performance on a public dataset and is not produced by fitting a parameter to the claimed output; it is an external benchmark result. The RAG-LLM component is described as retrieving passages from a knowledge base that the authors created ('We created a detailed document that serves as a knowledge base describing various coffee leaf diseases, symptoms, causes, and remedies' in Section 4.2) and passing them to the LLM. This is a retrieval pipeline by construction, not a derivation of an independently validated claim, so it is not circular in the sense of a prediction reducing to its inputs. There is no load-bearing self-citation: the cited BRACOL datasets [15,16] are external, and no uniqueness theorem or prior-work-forced ansatz is invoked. Two issues in the manuscript are correctness/reproducibility problems rather than circularity. First, Section 5.1 states 'the best performance was obtained when the model was trained using the original version of the dataset,' while Section 7.1 says the results are 'After training the YOLOv8n model on the curated and re-annotated BRACOL dataset,' and Section 6 gives no explicit split definition; this leaves the exact provenance of Table 1 ambiguous. Second, Section 7.1 says 'classes with a higher number of representative annotations (Miner and Phoma) yield better results,' but Section 5.2 reports rust with 6,013 annotations and the lowest mAP@0.5 (0.415), while miner with 341 annotations has the highest (0.894). These are internal consistency and evaluation-provenance flaws that affect interpretability, but neither is a case where a claimed result is equivalent to its input by construction. Accordingly, the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a post-hoc dataset-version choice, an unreleased and unvalidated knowledge base, and several unstated assumptions about annotation quality, hyperparameter sufficiency, validation representativeness, and LLM grounding. These are the elements the paper does not independently support.

free parameters (1)
  • BRACOL dataset version selection = original version with 1,899 annotations
    Section 5.1 states both dataset versions were trained and the original version was chosen because it gave better performance; this post-hoc selection determines all reported detection metrics and is a choice made after seeing results.
assumptions (5)
  • domain assumption BRACOL dataset annotations are accurate.
    The detection results depend on the ground-truth labels; no error analysis of the annotations is provided, and the paper relies on them without verification.
  • domain assumption The self-authored knowledge base document is factually correct and complete.
    Section 4.2 says a 'detailed document' was created by the authors; no external validation or release is provided, yet the system's remedy outputs rely entirely on it.
  • domain assumption Default Ultralytics hyperparameters are adequate for coffee leaf detection.
    Training used default settings (Section 4.1) without tuning; the paper assumes these are suitable, which may or may not hold for this domain.
  • domain assumption The validation set of 228 images is representative of real field conditions.
    Section 7.1 uses a single validation split; no test set or cross-validation is reported, so class performance could be split-specific.
  • domain assumption The Groq-hosted LLM generates factual text when given retrieved context.
    Section 4.3 assumes retrieval-augmented prompting yields grounded answers, but no factuality evaluation is performed, and hallucination mitigation is claimed without evidence.
invented entities (1)
  • Author-created coffee leaf disease knowledge base
    purpose: Provide grounding passages for the RAG pipeline so the LLM generates disease explanations and remedies
    The document is described in Section 4.2 but not released or externally validated; the system's factuality depends entirely on this self-authored artifact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision Meets Language: A RAG-Augmented YOLOv8 Framework for Coffee Disease Diagnosis and Farmer Assistance." pith.science (2026). https://pith.science/paper/GHIZDXL3

@misc{pith2026250521544,
  author       = {Pith},
  title        = {Pith review of: Vision Meets Language: A RAG-Augmented YOLOv8 Framework for Coffee Disease Diagnosis and Farmer Assistance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GHIZDXL3}},
  note         = {Machine review of arXiv:2505.21544}
}
read the original abstract

As a social being, we have an intimate bond with the environment. A plethora of things in human life, such as lifestyle, health, and food are dependent on the environment and agriculture. It comes under our responsibility to support the environment as well as agriculture. However, traditional farming practices often result in inefficient resource use and environmental challenges. To address these issues, precision agriculture has emerged as a promising approach that leverages advanced technologies to optimise agricultural processes. In this work, a hybrid approach is proposed that combines the three different potential fields of model AI: object detection, large language model (LLM), and Retrieval-Augmented Generation (RAG). In this novel framework, we have tried to combine the vision and language models to work together to identify potential diseases in the tree leaf. This study introduces a novel AI-based precision agriculture system that uses Retrieval Augmented Generation (RAG) to provide context-aware diagnoses and natural language processing (NLP) and YOLOv8 for crop disease detection. The system aims to tackle major issues with large language models (LLMs), especially hallucinations and allows for adaptive treatment plans and real-time disease detection. The system provides an easy-to-use interface to the farmers, which they can use to detect the different diseases related to coffee leaves by just submitting the image of the affected leaf the model will detect the diseases as well as suggest potential remediation methodologies which aim to lower the use of pesticides, preserving livelihoods, and encouraging environmentally friendly methods. With an emphasis on scalability, dependability, and user-friendliness, the project intends to improve RAG-integrated object detection systems for wider agricultural applications in the future.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

  1. [9]

    github (2025)

    Abhishek Chaudhari: A real-time platform for plant disease identification and management. github (2025)

    work page 2025

  2. [16]

    Applied Sciences 15(9) (2025) 14

    Fragoso, J., Silva, C., Paix˜ ao, T., Alvarez, A.B., J´ unior, O.C., Florez, R., Palomino-Quispe, F., Savian, L.G., Trazzi, P.A.: Coffee-leaf diseases and pests detection based on yolo models. Applied Sciences 15(9) (2025) 14

    work page 2025

  3. [1]

    Food and Agriculture Organization of the United Nations (2022)

    F AO: Plant production and protection. Food and Agriculture Organization of the United Nations (2022)

    work page 2022

  4. [2]

    efsa (2023)

    al., L.C.C.: The 2021 european union report on pesticide residues in food. efsa (2023)

    work page 2023

  5. [3]

    EIP (2022)

    Berger, L.T., Doll, D., Schwitzky, E., Lavelle, K., Skalsky, M., Spinelli, F.: Sus- tainable ways to reduce pesticides in pome and stone fruit production. EIP (2022)

    work page 2022

  6. [4]

    Ultralytics (2024)

    Ultralytics: Models supported by ultralytics. Ultralytics (2024)

    work page 2024

  7. [5]

    Heliyon 10 (2024) https://doi.org/10.1016/j

    Abid, M.S., Jahan, B., Mamun, A., Hossen, Mazumder, S.: Bangladeshi crops leaf disease detection using yolov8. Heliyon 10 (2024) https://doi.org/10.1016/j. heliyon.2024.e36694

    doi:10.1016/j 2024

  8. [6]

    Engineering, Technology and Applied Science Research 14, 13584–13591 (2024) https://doi.org/10.48084/etasr.7064

    Zayani, H., Ammar, I., Ghodhbane, R., Maqbool, A., Saidani, T., Ben Slimane, J., Kachoukh, A., Kouki, M., Kallel, M., Alsuwaylimi, A., Alenezi, S.: Deep learning for tomato disease detection with yolov8. Engineering, Technology and Applied Science Research 14, 13584–13591 (2024) https://doi.org/10.48084/etasr.7064

    work page doi:10.48084/etasr.7064 2024

Show all 16 references
  1. [7]

    Frontiers in Plant Science 15 (2025) https: //doi.org/10.3389/fpls.2024.1514832

    Miao, Y., Meng, W., Zhou, X.: Serpensgate-yolov8: an enhanced yolov8 model for accurate plant disease detection. Frontiers in Plant Science 15 (2025) https: //doi.org/10.3389/fpls.2024.1514832

    2025

  2. [8]

    Plant Methods 20 (2024) https://doi.org/10.1186/s13007-024-01275-3

    Cao, Q., Zhao, D., Li, J., Li, J., Li, G., Feng, S., Xu, T.: Pyramid-yolov8: a detection algorithm for precise detection of rice leaf blast. Plant Methods 20 (2024) https://doi.org/10.1186/s13007-024-01275-3

    2024 doi

  3. [10]

    https://arxiv.org/abs/2504.02280

    Yu, Y., Zutty, J.: LLM-Guided Evolution: An Autonomous Model Optimization for Object Detection (2025). https://arxiv.org/abs/2504.02280

    2025 arXiv

  4. [11]

    https://arxiv.org/abs/2504.20419

    Roumeliotis, K.I., Sapkota, R., Karkee, M., Tselikas, N.D., Nasiopoulos, D.K.: Plant Disease Detection through Multimodal Large Language Models and Convolutional Neural Networks (2025). https://arxiv.org/abs/2504.20419

    2025 arXiv

  5. [12]

    https://arxiv.org/abs/2310.05149 13

    Feng, Z., Feng, X., Zhao, D., Yang, M., Qin, B.: Retrieval-Generation Synergy Augmented Large Language Models (2023). https://arxiv.org/abs/2310.05149 13

    2023 arXiv

  6. [13]

    https: //arxiv.org/abs/2309.17050

    Louis, A., Dijck, G., Spanakis, G.: Interpretable Long-Form Legal Question Answering with Retrieval-Augmented Large Language Models (2023). https: //arxiv.org/abs/2309.17050

    2023 arXiv

  7. [14]

    https://arxiv.org/abs/2311.05903

    Dodgson, J., Nanzheng, L., Peh, J., Pattirane, A.R.J., Alhajir, A.D., Dinarto, E.R., Lim, J., Ahmad, S.D.: Establishing Performance Baselines in Fine-Tuning, Retrieval-Augmented Generation and Soft-Prompting for Non-Specialist LLM Users (2024). https://arxiv.org/abs/2311.05903

    2024 arXiv

  8. [15]

    Computers and Electronics in Agriculture 169, 105162 (2020) https://doi.org/10.1016/j.compag.2019.105162

    Esgario, J.G.M., Krohling, R.A., Ventura, J.A.: Deep learning for classification and severity estimation of coffee leaf biotic stress. Computers and Electronics in Agriculture 169, 105162 (2020) https://doi.org/10.1016/j.compag.2019.105162

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.