Krishi Sathi: AI Agricultural Advisor

• Krishi Sathi is an AI-powered agricultural advisory chatbot that uses multi-turn dialogue and retrieval-augmented generation to provide personalized crop-specific guidance.
• The system specializes in grapes and onions by employing intent-aware slot filling combined with a curated agricultural knowledge base for precise recommendations.
• It features a bilingual text and speech interface that enhances accessibility for low-literacy farmers while delivering strong performance metrics and context-aware responses.

Krishi Sathi is an AI-powered agricultural advisory chatbot for Indian farmers that combines multi-turn dialogue, intent-aware context acquisition, and retrieval-augmented generation to deliver personalized answers through text and speech in English and Hindi. In its current reported form, the system is specialized to two crops—grapes and onions—and is designed to ask follow-up questions before answering underspecified agricultural queries, rather than relying on one-shot responses. Its architecture couples an instruction-fine-tuned domain model with a curated agricultural knowledge base and a bilingual speech interface, with the stated aim of improving accessibility for low-literacy and digitally inexperienced users (Vijayvargia et al., 28 Jul 2025).

1. Definition and problem setting

Krishi Sathi is framed as a response to a knowledge-access problem in Indian agriculture: advisory queries are often short, colloquial, and context-dependent, while existing helplines and portals can be constrained by personalization, language mismatch, and limited scalability. The system therefore treats agricultural question answering not as generic open-domain chat, but as a task-oriented advisory workflow in which crop type, state, season, seed variety, soil condition, and related agronomic attributes must often be elicited before a response can be generated (Vijayvargia et al., 28 Jul 2025).

A defining feature of the system is its rejection of immediate answer generation for incomplete inputs. A query such as a fertilizer or irrigation question is treated as potentially unsafe or uninformative unless the relevant slots for the identified intent have been filled. This places Krishi Sathi in the class of dialogue systems closer to task-oriented agricultural assistants than to static FAQ bots. The paper also positions accessibility as a core requirement rather than an auxiliary feature: the system supports speech input and output, and bilingual interaction in English and Hindi, because many intended users may prefer oral interaction or may not be comfortable with text-heavy interfaces (Vijayvargia et al., 28 Jul 2025).

The current scope is intentionally narrow. The reported system is specialized to grapes and onions, with crop-specific intent inventories and slot schemas, rather than a generalized agriculture-wide ontology. This bounded design suggests a prototype-oriented architecture in which domain depth within selected crops is prioritized over immediate breadth across all Indian farming systems (Vijayvargia et al., 28 Jul 2025).

2. System architecture and inference pipeline

Krishi Sathi is organized as a modular pipeline that transforms a farmer query into a grounded agricultural response. The pipeline accepts either text or speech. Speech input must be a 16 kHz .wav file. Speech is transcribed, language is detected, Hindi queries are translated into English, the normalized query is routed into one of three classes—Domain Specific, General Knowledge, or Casual Queries—and, if domain-specific, it is passed through crop classification, intent recognition, slot extraction, clarification dialogue, dense retrieval, and grounded generation before the answer is returned as text or speech (Vijayvargia et al., 28 Jul 2025).

Component	Stated implementation	Function
ASR	In-house 120M-parameter model	Speech-to-text
Language handling	Rule-based character-frequency method; Hindi translated with Google Cloud Translate	Query normalization
Retriever	all-mpnet-base-v2, 768-dimensional embeddings, Qdrant, top-1 search	Passage retrieval
Generator	Fine-tuned Param-1-2.9B model	Grounded answer generation
TTS	Hindi F5TTS-small, 150M parameters	Spoken response

The retrieval stage is formulated as nearest-neighbor search in an embedding space. Documents from the agricultural corpus are embedded offline using all-mpnet-base-v2, stored as 768-dimensional vectors in Qdrant, and queried with top-1 retrieval, i.e., $k = 1$. The generation stage then conditions on three inputs: the enriched query, the retrieved agricultural passage, and a few-shot prompt template containing 2–3 curated exemplars. The paper characterizes this as retrieval-augmented generation intended to reduce hallucination by grounding responses in expert-reviewed agricultural text (Vijayvargia et al., 28 Jul 2025).

The routing and classification layer contains a manuscript-level inconsistency. One part of the paper assigns routing and classification to Mistral-Nemo-Instruct-2407, whereas another assigns router, crop classifier, intent classifier, and question generator roles to Param-1-2.9B-instruct. The paper does not reconcile this discrepancy. The safest characterization is therefore that the system includes an LLM-based routing and classification layer, but the exact deployed model assignment is not specified consistently in the manuscript (Vijayvargia et al., 28 Jul 2025).

3. Knowledge curation, intent schema, and model training

The knowledge substrate behind Krishi Sathi is derived from curated Indian agricultural sources: ICAR, Vikaspedia, the National Research Centre for Grapes, and the Directorate of Onion and Garlic Research. The initial scraped corpus comprised approximately 20.4 million tokens; after filtering and curation, this was reduced to approximately 12 million tokens, about 59% of the original. The curation process ran for two weeks and combined automated preprocessing, custom scripts, and human review by three domain experts (Vijayvargia et al., 28 Jul 2025).

From this curated corpus, the system constructed a “Refined-Passage-Dataset” of 150,000 domain-refined, expert-reviewed passages. These passages serve as the retrieval corpus for the RAG subsystem. The same passage set was then converted into an instruction-based dataset containing 15 conversation-based instruction tasks, again reviewed by three domain experts, to support supervised fine-tuning of the answer-generation model (Vijayvargia et al., 28 Jul 2025).

Separate from passage curation, the dialogue layer required an explicit intent-slot schema. The reported system defines 25 intents for grapes and 22 intents for onions, with 2–5 associated slots per intent. Approximately 18,000 annotated examples were created across both domains by three trained annotators. Each intent corresponds to a high-level user goal, while slots encode the agronomic information required to answer that goal appropriately. This schema is the backbone of the clarification dialogue, since the system checks slot completeness before triggering retrieval and generation (Vijayvargia et al., 28 Jul 2025).

The fine-tuning setup for the domain generator is unusually specific. The paper reports supervised fine-tuning of Param-1-2.9B using NVIDIA NeMo Megatron GPT SFT on a single-node cluster with 8 NVIDIA H200 GPUs, bf16 precision, 3 epochs, global batch size 1024, micro-batch size 4 per GPU, gradient accumulation 32, maximum sequence length 2048, Distributed Fused Adam with $\beta_1 = 0.9$, $\beta_2 = 0.98$, weight decay $= 0.01$, a linear learning-rate schedule, 1200 warmup steps, 12000 total steps, peak learning rate $5 \times 10^{-6}$, no gradient clipping, and checkpointing every 1000 steps with the top 20 models retained by validation loss as .nemo files (Vijayvargia et al., 28 Jul 2025).

4. Dialogue management, retrieval logic, and multilingual interaction

Krishi Sathi’s distinguishing feature is its intent-aware multi-turn dialogue. The system does not treat the initial query as sufficient by default. Instead, it predicts an intent, checks the required slot schema for that intent, and asks follow-up questions if any slots are missing or ambiguous. The paper states that this iterative slot-filling process typically takes 2–3 dialogue turns. Context is accumulated through slot filling rather than via a separately formalized dialogue-state vector (Vijayvargia et al., 28 Jul 2025).

The crop-specific intent design is concrete. For grapes, the paper lists examples such as “Vineyard Variety Selection,” with slots including Grape Variety, Climate, Expected Yield Potential, and Soil Type; “Irrigation Management,” with slots including State, Season, Seed Variety, and Soil Testing; and “Fertilization and Nutrient Management,” with slots such as Fertilizer Type, Fertilization Schedule, and Micronutrient Deficiency in Soil. For onions, examples include “Time of Transplanting,” “Integrated Pest Management Protocols,” and “Land Clearing and Tilling,” each with its own slot schema (Vijayvargia et al., 28 Jul 2025).

Retrieval is explicitly conditioned on intent plus slot values. Rather than embedding the raw question alone, the system augments the query using extracted contextual details and then performs dense retrieval over the passage corpus. This makes retrieval dependent on structured agricultural context such as crop, state, season, variety, or soil attributes. The paper does not give a formal query-construction equation, but it repeatedly describes retrieval as occurring only after sufficient contextual enrichment (Vijayvargia et al., 28 Jul 2025).

The multilingual and speech interface is asymmetric but operationally central. Speech input is handled by an in-house ASR model with 120M parameters deployed on one H100 GPU. Speech output is handled by a Hindi TTS model based on F5TTS-small with 150M parameters, also deployed on one H100 GPU. Language identification is rule-based over custom English and Hindi dictionaries, and Hindi queries are translated into English using Google Cloud Translate so that the core reasoning pipeline operates in English. This implies that bilingual interaction is implemented via a translate-to-pivot-language design rather than a natively multilingual reasoning stack (Vijayvargia et al., 28 Jul 2025).

5. Reported performance, evaluation protocol, and limitations

The paper reports strong system-level metrics. Query Response Accuracy is reported as 97.53%, Personalization & Contextual Relevance as 91.35%, Follow-up Question Relevance as 82.85%, Query Completion Rate as 97.53%, and average response time as approximately 5.96 seconds. The results section also notes that 53.66% of responses took more than 3 seconds. Multimodal text-output delivery rate is reported as 98%, uptime rate as approximately 100% during testing, and error rate as negligible. Training loss is reported to decrease from 2.3049 to 0.2239, with evaluation loss reaching 0.3343 (Vijayvargia et al., 28 Jul 2025).

Metric	Reported value	Note
Query Response Accuracy	97.53%	System-level
Personalization & Contextual Relevance	91.35%	System-level
Follow-up Question Relevance	82.85%	System-level
Query Completion Rate	97.53%	System-level
Average Response Time	$\sim 5.96$ s	More than 3 s for 53.66% of responses

These figures should be read with caution because the evaluation methodology is only partially specified. The paper does not clearly report the size of the evaluation set, the number of users, train/dev/test split details, annotation procedures for the evaluation labels, formulas for most system-level metrics, confidence intervals, statistical significance tests, ablation studies, or baseline comparisons against other agricultural chatbots or simpler single-turn/RAG-free variants. The reported values therefore indicate promising internal performance, but they do not support a strong comparative claim against alternative architectures (Vijayvargia et al., 28 Jul 2025).

The system’s limitations are substantial and explicitly acknowledged. Current crop coverage is limited to grapes and onions. Current language support is limited to English and Hindi, although the paper proposes eventual scaling to 22 official Indic languages. The evaluation is methodologically sparse. Latency remains nontrivial, given that a majority of responses exceed 3 seconds. The paper also contains an unresolved model-allocation inconsistency in the routing/classification layer. These limitations make the reported system best understood as a specialized prototype or early production candidate, rather than a fully generalized agricultural assistant for India (Vijayvargia et al., 28 Jul 2025).

6. Relation to the broader agricultural AI ecosystem

Krishi Sathi, as described in the intent-aware multi-turn QA paper, is primarily an advisory dialogue system. Related work in the same period points toward a broader modular stack that could surround such a core. Multi-crop leaf-disease classification has been shown on a unified dataset of 93,136 images spanning 17 crops, 34 diseases, and 51 classes, using a lightweight ResNet9-based residual CNN with reported overall accuracy of about 99.03%; however, the paper also reports sharply weaker performance for some classes such as wheat-healthy, indicating that broad-coverage disease modules still require careful class-wise validation before deployment in a farmer-facing assistant (Yadav et al., 3 Jul 2025).

A separate line of work proposes a hybrid crop-selection engine that combines agronomic suitability and market forecasting through a Random Forest classifier and an LSTM, exposed through a Kannada voice-first interface. That system reports 98.5% suitability accuracy and one-month-ahead price forecasting with average RMSE of ₹3.50 per kg and MAPE of approximately 5.8%, but its actual decision rule is price maximization among top agronomically suitable crops rather than full net-profit optimization. This suggests that an agriculture assistant can couple agronomy, economics, and voice accessibility, but also that “profitability” claims need careful interpretation when cost-of-cultivation is not yet modeled explicitly (Sindhur et al., 6 Jul 2025).

Retrieval-based farmer-support systems built on Kisan Call Center logs provide another adjacent module. One such system reports baseline semantic-retrieval accuracy of about 56%, rising to 86% after synonym normalization and crop-name entity extraction; another WhatsApp-oriented bot built with RASA and TF-IDF-based phrase matching reports about 96% response accuracy in a proof-of-concept setting. These systems indicate that domain-specific normalization, metadata preservation, and escalation to human agents remain central design patterns for agricultural assistants, especially where repeated advisory queries dominate (Jain et al., 25 Sep 2025, Darapaneni et al., 2022).

Two other research directions are especially relevant for expansion beyond QA. First, farmer-query streams can function as a pest-surveillance layer: a national-scale KCC analysis over 10,981,793 queries from 2015–2020 extracted 867,337 pest-related queries and showed that query-derived hotspots aligned with reported outbreaks, sometimes more than a month before major news reporting. This suggests that an assistant such as Krishi Sathi could eventually serve not only as a response system but also as a sensor for spatio-temporal pest intelligence (Adhikari et al., 2021). Second, multilingual structured corpora such as AgriGov—covering 50 farmer welfare schemes in English, Hindi, and Marathi, with about 2,100–2,200 source segments and about 8,000 Hindi–Marathi aligned sentence pairs—suggest a route toward scheme discovery, eligibility checking, and retrieval-grounded public-service guidance within the same assistant stack (Bilal et al., 6 Jun 2026).

The paper on Krishi Sathi itself already points toward such expansion. Its stated future work includes image integration through the i-SARATHI mobile app, soil-testing integration via SAMBHAV, possible interoperability with mKisan and KISAAN 2.0, multimodal agro-environmental parameter integration, and expansion to 22 Indic languages (Vijayvargia et al., 28 Jul 2025). This suggests that “Krishi Sathi” is best understood not merely as a chatbot, but as an emerging architectural pattern for agricultural assistance in which dialogue management, retrieval grounding, crop-specific schemas, speech accessibility, and eventually sensing and external services converge into a unified advisory platform.