KRISHOKTECH
A generative AI system for Bengali-language agricultural advisory scenarios. It controls the safety of generated content through explicit authorization boundaries, including admission checks before retrieval, crop binding during retrieval, and dosage verification after generation.
Timeline (last 90 days)
The core design idea of the Bengali agricultural advisory system KRISHOKTECH is to set explicit authorization boundaries before and after the generation process, rather than relying solely on the capability of the generative model.
KRISHOKTECH performs normalization and rule-based admission checks before retrieval, refers high-risk requests covered by policy (REFER), and proactively asks (ASK) for the specific crop when the treatment intent description is insufficient.
KRISHOKTECH binds retrieval to a single crop and requires the user to confirm (CONFIRM) when the text and image indicate different crops.
After generation, KRISHOKTECH uses a narrow-scope verifier to remove dosage claims that are not supported by the retrieved BARI/BRRI evidence.
KRISHOKTECH keeps agricultural guidance in an external evidence index, which can be updated without retraining the generator.
The study tested KRISHOKTECH's control measures on 200 farmer queries and 400 stress queries.
Under the alias filter tested, KRISHOKTECH's cross-crop retrieval violations dropped from 36.5% to 0%.
KRISHOKTECH's mismatch detection caught 453 of 454 text-image conflicts.
KRISHOKTECH's dosage verifier caught 114 of 118 inserted dosage errors.
The KRISHOKTECH interface presents the decision process as four states—ASK, CONFIRM, REFER, and DROP—along with the corresponding evidence.
4 more timeline events
All Facts (14)
The core design idea of the Bengali agricultural advisory system KRISHOKTECH is to set explicit authorization boundaries before and after the generation process, rather than relying solely on the capability of the generative model.
50%UnverifiedKRISHOKTECH performs normalization and rule-based admission checks before retrieval, refers high-risk requests covered by policy (REFER), and proactively asks (ASK) for the specific crop when the treatment intent description is insufficient.
50%UnverifiedKRISHOKTECH binds retrieval to a single crop and requires the user to confirm (CONFIRM) when the text and image indicate different crops.
50%UnverifiedAfter generation, KRISHOKTECH uses a narrow-scope verifier to remove dosage claims that are not supported by the retrieved BARI/BRRI evidence.
50%UnverifiedKRISHOKTECH keeps agricultural guidance in an external evidence index, which can be updated without retraining the generator.
50%UnverifiedThe study tested KRISHOKTECH's control measures on 200 farmer queries and 400 stress queries.
50%UnverifiedUnder the alias filter tested, KRISHOKTECH's cross-crop retrieval violations dropped from 36.5% to 0%.
50%UnverifiedKRISHOKTECH's mismatch detection caught 453 of 454 text-image conflicts.
50%UnverifiedKRISHOKTECH's dosage verifier caught 114 of 118 inserted dosage errors.
50%UnverifiedThe KRISHOKTECH interface presents the decision process as four states—ASK, CONFIRM, REFER, and DROP—along with the corresponding evidence.
50%UnverifiedThe four decision states ASK, CONFIRM, REFER, and DROP correspond respectively to: insufficient request information, multimodal conflict or ambiguous matching, high-risk situations covered by policy, and violation of safety rules that cannot be remedied.
50%UnverifiedKRISHOKTECH's dosage verifier adopts a narrow-scope design that targets only dosage claims and does not cover other types of errors such as crop disease identification errors or deviations in application timing advice.
50%UnverifiedThe KRISHOKTECH system is built for Bengali and the BARI/BRRI evidence system, and migrating it to other languages and agricultural systems requires redesigning the evidence index and rules.
50%UnverifiedThis study is a demonstration work with limited testing scale.
50%