KrishokTech: A Bengali Agricultural Advisory System with Explicit Authorization Controls

Khan Raiyan Ibne Reza; Sanjana Aktar Maria; Shakil Ahmed; Faiyad Hossain; Sumaiya Tabassum Nimi

cs.CLOct 23, 2026

KrishokTech: A Bengali Agricultural Advisory System with Explicit Authorization Controls

Authors: Khan Raiyan Ibne Reza, Sanjana Aktar Maria, Shakil Ahmed, Faiyad Hossain, Sumaiya Tabassum Nimi

Organizations: Department of Computer Science and Engineering, North South University, Dhaka 1229, Bangladesh

Abstract

For agricultural advisory, generating an answer is only part of the problem: the system must also control unsafe requests, work with limited connectivity, and accommodate changing guidance. KRISHOKTECH is a Bengali agricultural advisory system built around an explicit authorization boundary around generation. Before retrieval, normalization and rule-based admission checks refer policy-covered high-risk requests and ask for a crop when treatment intent is underspecified. The system then binds retrieval to a single crop and asks for confirmation when the text and image indicate different crops. After generation, a narrow verifier removes dosage claims not supported by retrieved BARI/BRRI evidence. Agricultural guidance is kept in an external evidence index that can be updated without retraining the generator, while browser-side vision and local retrieval reduce connectivity requirements. The interface makes these decisions visible as ASK, CONFIRM, REFER, and DROP, together with their evidence. On 200 farmer queries and 400 stress queries, the controls halted underspecified requests and reduced cross-crop retrieval violations from 36.5% to 0% under the tested alias filter. Mismatch detection caught 453 of 454 text-image conflicts, while the verifier caught 114 of 118 inserted dosage errors. The demonstration shows how these controls govern the path from request to delivered advice.

Figures & tables

Figure 1: KrishokTech places five control stages (T0–T4) before delivery. T0 can end a request in REFER , T1 in ASK , and T2 in CONFIRM . T3 produces an unverified draft; T4 removes unsupported dosage claims before the remaining text is delivered.

Panel of Figure 2 — Figure 2: The interface exposes four decisions at the authorization boundary. Each panel shows what stopped or changed the pipeline rather than presenting only its final answer. Full-size screens for S1–S5 appear in Appendix C .

Control	Test set	Result
T1 crop gate (C1)	200 farmer queries	30/30 halt; 170/170 pass
T0 normalization (C1)	85 attacks, 15 benign	85/85 caught; 0/15 flagged
Crop binding/filter (C2)	400 stress queries	146 $\rightarrow$ 121 $\rightarrow$ 0 lists
Mismatch state (C2)	54 farmer, 400 synthetic	453/454; 0/454 controls
Prompt + T0 (C1)	280 fixed attacks	147 $\rightarrow$ 30 $\rightarrow$ 13 success
Dosage verifier (C3)	118 errors, 38 clean	114 caught; 0/38 rejected

Table 1: Main component results. For crop binding, the sequence is open retrieval, gold-crop binding, and the alias-based crop filter. For the generator test, it is the neutral prompt, guarded prompt, and T0 replay.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Set	n	Formal	Normalized	Route
Phonetic chemical	35	0 blocked	35 blocked	16123
Banglish crisis	25	1 blocked	25 blocked	999
Indirect sabotage	25	0 blocked	25 blocked	16123
All attacks	85	1 (1.2%)	85 (100%)
Benign control	15	15 passed	15 passed	none

Appendix

Table 2: Constructed Banglish and phonetic test for T0 (

N=100

). The formal condition recognizes only formal English and Bengali-script terms; the normalized condition adds the supported variants. The 95% Wilson interval for 85/85 intercepted attacks is [95.7, 100].

Measure	Naive	LLM	Comp. 1	Comp. 2
Dose retained	23/100	89/100	92/100	94/100
Chemical retained	20/100	56/100	34/100	63/100
Both retained	14/100	51/100	32/100	61/100
Length $\leq 160$ chars	100/100	98/100	100/100	100/100
UCS-2 segments (mean)	3.0	2.1	2.5	2.5
Invented value	0	1	0	0

Appendix

Table 3: Paired SMS comparison on 100 long dosage-bearing advisories (seed 42). Comp. 1 is the original deterministic compressor; Comp. 2 additionally copies the nearest recognized active ingredient from the verified answer. Sixty-seven reference advisories contain a chemical recognized by the scorer.

Artifact, excluding LLM weights	Size (MB)
Unique ONNX vision assets (11 files)	95.64
Minimal installation	284.0
Installation with PyTorch checkpoints	339.6

Appendix

Table 4: Measured installation footprint. Language-model weights are excluded from every row.

Component	Median (ms)	p95 (ms)
Safety processing	0.321	0.705
BM25 retrieval	26.831	61.248
Dosage verification	4.815	11.220
Non-generation total	33.565	67.833

Appendix

Table 5: Latency over 400 requests with BM25 retrieval and generation stubbed out. Component medians need not sum to the median total.

Panel of Figure 3 — Figure 3: Demonstration gallery: the five website states expose admission, crop binding, conflict confirmation, safety referral, and dosage release. The two-row, three-column layout enlarges the screenshots while preserving their aspect ratios.

Cohort	Dial.	Turns	Check	Result
Carryover	75	175	Crop slot kept	100/100
Crop shift	75	150	Slots flushed	75/75
(same)			Old slot kept	0/75
Delayed evasion	30	60	Turn-2 catch	30/30
Clarification	20	40	ASK, then bound	20/20
Total	200	425

Appendix

Table 6: Multi-turn state tests on constructed dialogues. The crop-shift cohort is scored on two checks; no pooled score is reported. Carryover counts the 100 follow-up turns.

LLM Guardrails

Large language model guardrails detect or constrain harmful outputs, prompt attacks, and unsafe agent actions before they cause harm.

See all 194 papers

KrishokChat: A Citation-Grounded Dataset and Benchmark for Bengali Agricultural Advisory
Jun 28, 2026Khan Raiyan Ibne Reza, Omar Ibne ShahidLLM Grounding Question Answering
North South University Dhaka, Bangladesh
Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B
Jun 27, 2026Zhaoyang Li, Ruijie Zhang, Jiaqi Liu +1LLM Fine-Tuning
1Sanya University · 2Hebei International Studies University
AgriJev: Efficient Agricultural Candidate Scoring with Multi-Scale Adapted Models
Oct 14, 2026Xuechen LiLanguage Modeling LLM Inference
University of Minnesota Twin Cities

Jun 28, 2026J/K move · Enter open · S save

KrishokChat: A Citation-Grounded Dataset and Benchmark for Bengali Agricultural Advisory

Khan Raiyan Ibne Reza, Omar Ibne Shahid

North South University Dhaka, Bangladesh

LLM Grounding Question Answering Synthetic Data Generation for Language Models Knowledge-Intensive QA

Open paper AskPDF arXiv

We present KrishokChat, the first citation-grounded Bengali agricultural instruction-tuning dataset for crop advisory in low-resource settings. We establish a foundation of 290 hierarchical Knowledge Nodes, extracting disease symptoms, management practices, chemical dosages, and verbatim citations from 129 domain-filtered agricultural manuals. Every training instance inherits a verified citation header, guaranteeing 100% citation provenance. Using a Partitioned Seed Generation Matrix, these nodes are expanded into 139,200 supervised fine-tuning pairs, and augmented with 5,300 chemical safety and 1,000 adversarial safety instances, yielding 145,500 QA pairs across 18 crop categories. To evaluate real-world performance, we introduce the Farmer Benchmark, comprising 1,001 authentic farmer queries curated from field surveys and digital portals. Empirical evaluation on Gemma-4-E2B reveals that while fine-tuning on KrishokChat vastly improves structured formatting, standalone models still struggle with exact chemical dosage generalization. This highlights the dataset's true value as a verified knowledge base for retrieval-augmented generation (RAG) rather than mere parametric memorization. All data, code, and benchmarks are released under CC-BY-4.0.

Abstract

LLM Guardrails

Explore similar work

KrishokChat: A Citation-Grounded Dataset and Benchmark for Bengali Agricultural Advisory

Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B

AgriJev: Efficient Agricultural Candidate Scoring with Multi-Scale Adapted Models