Quote of the day

A person who never made a mistake never tried anything new.

- Albert Einstein

Contact

Voice-First Agentic RAG Advisory in Moroccan Darija for Smallholder Farmers

Open Agriculture

Voice-First Agentic RAG Advisory in Moroccan Darija for Smallholder Farmers

Voice-First Agentic RAG Advisory in Moroccan Darija for Smallholder Farmers

Description

Develop a voice-first agricultural advisory agent in Moroccan Darija, grounded through retrieval over authoritative agronomic sources and designed to abstain when documentary evidence is insufficient.

Details

Context and Problem Statement

Agricultural extension services remain understaffed relative to the number of farms requiring technical support. Large language models provide a natural interface, but hallucinations on localized agronomic questions such as dosage, crop calendars, and pesticide authorization are unacceptable. Darija is also underrepresented in training corpora, while many farmers prefer oral interaction.

Research Question

Which agentic architecture can ensure factual grounding and controlled abstention for a voice assistant operating in Darija while respecting the latency and cost constraints of entry-level smartphones?

Proposed Approach

The pipeline combines Darija-adapted speech recognition, multilingual retrieval, geo-aware RAG over authoritative sources, tool use for weather and dose calculations, speech synthesis, and a calibration module that decides when to answer, ask for clarification, abstain, or escalate to a human advisor.

Expected Contribution

A dedicated evaluation protocol for factual faithfulness and abstention in a low-resource spoken language, supported by an agronomy-focused Darija question-answer benchmark validated by experts.

Expected Prototype

A lightweight mobile and voice-messaging assistant with source logging and an audit interface for agricultural advisors.

Datasets

Darija-SFT-Mixture, DarijaMMLU, FAO and national agronomic documents, and field-collected speech recordings.

Challenges

Dialectal variability, Darija-French code-switching, limited gold-standard annotations, legal responsibility for incorrect advice, and protection of farmer data.

Research Question

How can the documentary faithfulness and abstention behavior of a voice-based agricultural assistant operating in a low-resource language such as Moroccan Darija be measured and improved?

Innovation

The project shifts the objective from conversational fluency toward verifiable grounding and explicit refusal when evidence is insufficient, within a low-resource spoken language and geo-grounded retrieval setting.

Expected Deliverable

A functional voice prototype, a Darija benchmark for faithfulness and abstention, and an experimental report quantifying the coverage-error trade-off.

Technologies

Large Language Models (Atlas-Chat, Qwen, Llama) Retrieval-Augmented Generation Automatic Speech Recognition (Whisper) Autonomous Agents and Tool Use Faithfulness Evaluation and Calibration

Required Skills

  • Natural Language Processing
  • RAG Engineering and Vector Databases
  • Speech Processing and ASR Fine-Tuning
  • Experimental Evaluation Design

Datasets

  • Darija-SFT-Mixture and DarijaMMLU (MBZUAI-Paris)
  • FAO and national agronomic technical documents
  • Field-collected speech recordings

Morocco & Africa Relevance

The project addresses limited access to localized agricultural advice in Morocco and the Maghreb by reducing both literacy and language barriers to digital agriculture.

References

  1. Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect
  2. Social acceptance of large language models in agricultural extension and advisory: a rapid review
  3. Leveraging large language models for GeoAI-enabled digital agro-advisory

Project Information

Duration: 4 to 6 months
Academic Level: Engineering, Master's or PhD
Research Area: Agriculture

Keywords

Darija Retrieval-Augmented Generation Agricultural Advisory Low-Resource NLP Speech Recognition

Evaluation

Scientific novelty
Impact
Technical feasibility
Data availability
Prototype potential
Industrial relevance

Apply for this topic

To apply for this topic, please send your CV and portfolio to:

contact@yanecode.com
Back to Research