[Epic] Phase 2: VAZHI Full (Offline Mode)

[CryptoYogiLLC]

Overview

Add offline AI inference capability to complement the hybrid retrieval architecture.

Status

Hybrid retrieval architecture is COMPLETE (ADR-009). The app provides full value via SQLite-backed deterministic lookups for 10 knowledge categories. This phase now focuses on adding optional on-device AI inference for conversational capabilities.

Completed

• Hybrid retrieval architecture (Epic Epic: Hybrid Retrieval Architecture for VAZHI App #14, sub-issues Create SQLite database structure and migrations #15-Integration testing for hybrid architecture #21)
• Query router with 10 category matchers and hybrid/deterministic/AI classification
• 7 retrieval services (Thirukkural, Schemes, Emergency, Healthcare, GenericData x6 categories)
• Model download service with pause/resume, network detection, cellular warning
• Download dialog UI with progress, speed, ETA
• Model status indicator widget
• SHA256 model integrity verification
• SSL certificate pinning for download URLs
• Hybrid chat provider integrating knowledge retrieval + AI responses
• Bilingual UI (English + Tamil) across result cards and chat bubbles
• Expandable content with Show more/Show less
• Smart routing: conversational queries route as hybrid (SQLite + AI), not forced deterministic
• 232 tests passing

Remaining (blocked on model training)

• Successful model training (v3.4 Qwen3-0.6B-Base pending — see models/TRAINING_LOG.md)
• Convert trained model to GGUF format (Issue Convert VAZHI model to GGUF format for mobile (<1GB target) #13)
• Integrate llama.cpp / llamadart for on-device inference
• Test inference quality on 4GB+ RAM devices (Android 10+ / iOS 15+)
• Wire AI response path in HybridChatProvider to actual model inference
• Validate Tamil output quality of quantized model
• End-to-end test: hybrid query returns SQLite data + AI explanation

Technical Scope

• Inference: llamadart (GGUF format) for on-device processing
• Model target: Qwen3-0.6B-Base, Q4_K_M quantization (<1GB)
• Min device: 4GB RAM, Android 10+ / iOS 15+
• Fallback: App already works fully without model via hybrid retrieval

Dependencies

• Model training success (v3.4 or fallback approach)
• Issue Convert VAZHI model to GGUF format for mobile (<1GB target) #13 (GGUF conversion)

Related ADRs

• ADR-009: Hybrid Retrieval Architecture (implemented)
• ADR-005: Incremental Pack Downloads (superseded)