⚠️ WARNING: NOT EXPERT-REVIEWED (DRAFT DATA)
This repository contains the model assets and configuration for an offline-first mobile RAG assistant.
The indexing contains 13 records which are currently in draft status and have NOT been verified by a sericulture expert or a native Kannada translator.
Repository Contents
db/database.db: SQLite database holding passage embeddings (float32 BLOBs) and metadata.
models/embedding/model_int8.onnx: intfloat/multilingual-e5-small quantized to INT8 for local mobile embeddings.
models/embedding/tokenizer/: Standard XLM-RoBERTa tokenization files.
models/llm/qwen3-0.6b-q4_k_m.gguf: Qwen/Qwen3-0.6B model quantized to Q4_K_M for on-device inference via llama.cpp.
config/app_config.json: Calibration thresholds, prompt templates, and per-language operational configurations.
kotlin/: Sample boilerplate code for integrating the repository, ONNX Runtime, and RAG logic in an Android App.
MANIFEST.json: Contains expected file paths, byte sizes, and SHA-256 hashes to safely coordinate app updates.
Model Details & Conversion Steps
Embedding Model: intfloat/multilingual-e5-small exported to ONNX (FP32) using PyTorch exporter, then dynamically quantized to INT8 with onnxruntime.quantization.
Language Model: Qwen/Qwen3-0.6B converted to high-precision GGUF (FP16) using llama.cpp scripts, and subsequently quantized to Q4_K_M using llama-quantize.
Attribution: Built using resources from the Central Sericultural Research & Training Institute, TNAU Agritech Portal, and Verified ResearchGate Literature.
