KrishokBondhu: A Retrieval-Augmented Voice-Based Agricultural Advisory Call Center for Bengali Farmers

KrishokBondhu is a voice-enabled, call-center-integrated advisory system for Bengali farmers in Bangladesh that leverages a Retrieval-Augmented Generation (RAG) framework to deliver high-quality, real-time agricultural guidance, achieving a 44.7% performance improvement over existing benchmarks.

The Gist

Imagine a farmer in a remote village in Bangladesh. He looks at his rice field and sees leaves turning yellow. He knows something is wrong, but he doesn't have a smartphone with the internet, he might not be able to read complex manuals, and the nearest agricultural expert is hours away.

This is the problem KrishokBondhu ("Farmer's Friend") was built to solve.

Here is a simple breakdown of how this system works, using everyday analogies.

1. The Problem: The "Library in a Language You Don't Speak"

For years, all the best advice on farming (how to stop pests, when to water, which seeds to buy) has been locked away in thick, heavy books written by experts.

• The Barrier: These books are often in technical Bengali or English. Even if a farmer could read them, they are too heavy to carry and too slow to find answers in an emergency.
• The Old Way: Farmers used to call a "Kisan Call Centre," but that often meant waiting on hold or talking to a human agent who might not have the specific answer right at their fingertips.

2. The Solution: A "Super-Helper" on the Phone

The researchers built KrishokBondhu, which acts like a super-smart, instant librarian that lives inside a phone call. You don't need a computer or the internet; you just dial a number and speak.

Here is how the "magic" happens, step-by-step:

Step A: The "Digital Brain" (The Knowledge Base)

First, the team took hundreds of official government farming manuals, textbooks, and guides (about 2,500 pages worth) and turned them into a digital format.

• The Analogy: Imagine taking a massive, dusty library of encyclopedias and scanning every single page into a computer. But instead of just storing them as pictures, the computer "reads" them and understands the meaning of every word. This is the OCR (Optical Character Recognition) pipeline.

Step B: The "Smart Index" (Retrieval-Augmented Generation)

This is the most important part. If you ask a regular AI (like a basic chatbot) a question, it might "hallucinate"—make up a fake answer because it's trying to be creative.

• The Analogy: KrishokBondhu is different. It's like a student taking an exam with an open textbook.
  1. Retrieval: When you ask, "Why are my tomatoes rotting?", the system doesn't guess. It instantly flips through its digital library, finds the exact page about tomato rot, and pulls out the relevant facts.
  2. Generation: It then takes those facts and uses a smart AI (called Gemma) to write a clear, friendly answer based only on what it found in the book. This ensures the advice is real and safe.

Step C: The "Voice Translator" (Speech-to-Speech)

The farmer doesn't need to type. He speaks in his local Bengali dialect.

• The Analogy: Think of the system as a simultaneous interpreter.
  1. You speak: "My rice plants have brown spots."
  2. The system hears you and turns your voice into text.
  3. It finds the answer in the "Digital Brain."
  4. It turns the answer back into natural-sounding Bengali speech and speaks it back to you through the phone.

3. How Good Is It? (The Results)

The researchers tested this system against the old standard (the Kisan Call Centre logs).

• The Old Way: The answers were short, like a text message: "Use 5kg of fertilizer." (Good, but not helpful if you don't know why or how).
• KrishokBondhu: The answers were like a conversation with a wise uncle. It explained why the plants were sick, how to fix it, how to prevent it next time, and even told the farmer when to call a real expert if it was too serious.
• The Score: On a scale of 1 to 5, the old system got a 3.1. KrishokBondhu got a 4.5. It was nearly 45% better, especially in giving full, complete answers.

4. Why This Matters

• No Reading Required: If you can't read well, you can still get expert advice.
• No Internet Required: It works on a simple phone call, which is available even in the poorest villages.
• Trustworthy: Because it pulls answers from official government books, the farmer knows the advice is safe and accurate, not made up by a robot.

In a Nutshell

KrishokBondhu is like taking the entire national library of farming knowledge, shrinking it down into a phone number, and giving it a voice so that every farmer, regardless of their education or location, can have a conversation with an expert instantly. It turns a complex, silent library into a friendly, talking guide.

Technical Summary

Here is a detailed technical summary of the paper "KrishokBondhu: A Retrieval-Augmented Voice-Based Agricultural Advisory Call Center for Bengali Farmers."

1. Problem Statement

Despite agriculture being a cornerstone of Bangladesh's economy (employing ~35–45% of the workforce), farmers face significant barriers in accessing timely, expert-level guidance.

• Accessibility Gaps: Traditional extension services are overstretched and cannot provide real-time support to remote communities.
• Literacy and Language Barriers: Much agricultural knowledge is locked in English or technical Bengali, inaccessible to farmers with low literacy or those unfamiliar with specialized terminology.
• Limitations of General AI: While Large Language Models (LLMs) exist, using them without domain grounding leads to "hallucinations," vague advice, or a lack of local context (e.g., specific crop varieties, regional pest cycles, and soil conditions).
• Existing Solutions: Previous systems (like KisanQRS) often rely on static datasets, lack voice interaction, or fail to ground responses in authoritative, up-to-date documentation.

2. Methodology: The KrishokBondhu System

The authors propose KrishokBondhu ("Farmer's Friend"), a voice-enabled, Retrieval-Augmented Generation (RAG) system designed specifically for Bengali speakers. The system architecture consists of four main components:

A. Data Collection and Curation

• Sources: The knowledge base aggregates ~2,500 pages of authoritative documents from the Government of Bangladesh and NGOs, including the Krishi Projukti Haat-boi, BARC handbooks, DAE manuals, agricultural textbooks, and WorldFish repositories.
• Coverage: Topics include major crops (rice, wheat, jute, etc.), livestock, fisheries, and integrated farming.

B. Document Processing Pipeline (OCR & Cleaning)

To handle heterogeneous inputs (scanned images, low-quality PDFs), a robust pipeline was developed:

1. Pre-processing: Image enhancement, skew correction, and noise reduction.
2. OCR: Bengali-capable Optical Character Recognition to extract text.
3. Post-processing: Rule-based and contextual validation to correct common Bengali character errors and normalize Unicode encoding.
4. Segmentation: Cleaned text is split into semantically coherent chunks (150–300 tokens) enriched with metadata (source, topic, position) and stored in Markdown format.

C. Retrieval and Generation Engine

• Vector Database: Text segments are embedded using a sentence-transformer model and indexed in LanceDB.
• Retrieval: User queries are embedded and matched via cosine similarity. A lightweight re-ranking layer combines semantic similarity with metadata filtering to retrieve the top-$k$ relevant context.
• Generation: The retrieved context is fed into Gemma 3-4B (a 4-billion parameter LLM) via LM Studio. The model is prompted to generate factual, grounded, and practical advice while acknowledging uncertainty.
• Voice Interface: Integrated via VAPI, the system handles:
  • Speech-to-Text (STT): Converts Bengali spoken queries to text.
  • Text-to-Speech (TTS): Synthesizes the LLM's response into natural Bengali speech.
  • Dialogue Management: Supports multi-turn interactions and fuzzy matching to handle recognition errors.

3. Key Contributions

1. Authoritative Knowledge Base: Creation of a Bangladesh-specific, curated agricultural corpus processed through a custom OCR-cleaning-segmentation pipeline.
2. End-to-End Voice RAG System: Implementation of a full-stack system integrating semantic search (LanceDB), a localized LLM (Gemma 3-4B), and Bengali STT/TTS modules for phone-based interaction.
3. Novel Evaluation Framework: Development of a practical assessment method for low-resource settings where standardized Bengali agricultural QA benchmarks do not exist. This involves expert rating and comparative analysis against the KisanQRS benchmark.

4. Results and Evaluation

The system was evaluated on a test set of diverse agricultural queries (crop diseases, pests, fertilizers, etc.) and compared against the KisanQRS benchmark.

• Performance Accuracy: KrishokBondhu produced high-quality responses for 72.7% of queries. Moderate quality (9.1%) occurred when context was partial, and poor quality (18.2%) was linked to under-represented topics.
• Comparative Metrics (vs. KisanQRS):
  • Overall Score: KrishokBondhu achieved 4.53/5.00, a 44.7% improvement over KisanQRS (3.13).
  • Contextual Richness: Massive gain of +367% (4.67 vs 1.00).
  • Completeness: +100.4% improvement (4.67 vs 2.33).
  • Response Length: Average response length increased from 87 characters (KisanQRS) to 692 characters, providing step-by-step guidance.
  • Relevance & Specificity: Maintained parity with KisanQRS (5.00 and 4.00 respectively).
• Information Coverage: KrishokBondhu provided cause explanations, prevention measures, and expert referrals in 100% of tested cases, whereas KisanQRS largely omitted these (0% for causes and prevention).
• Correlation Analysis: A strong positive correlation was found between semantic similarity scores and response quality, confirming that effective retrieval is a primary predictor of answer accuracy.

5. Significance and Future Work

• Significance: KrishokBondhu demonstrates the feasibility of deploying AI-driven agricultural advisory in low-resource, low-literacy environments. By combining voice interfaces with RAG, it bridges the gap between expert knowledge and remote farmers, ensuring advice is not only accessible but also contextually accurate and actionable.
• Limitations: The system relies on expert-based evaluation due to the lack of standardized benchmarks. The knowledge base requires continuous updates to reflect new crop varieties and emerging diseases.
• Future Directions: Plans include expanding document coverage, integrating structured knowledge graphs, adding real-time data (weather, market prices), and exploring domain-specific fine-tuning of the LLM.

In conclusion, KrishokBondhu represents a significant step toward democratizing agricultural knowledge in Bangladesh, moving from static, text-heavy resources to dynamic, voice-based, AI-assisted guidance.