How shubham0204/android-document-qa works
Project type and purpose
• This is a full-stack Android application (mobile client-only) implementing a retrieval-augmented generation (RAG) pipeline for document question-answering.
• Purpose: let users import PDF/DOCX, split text into chunks, generate on-device embeddings, store/retrieve them in a local vector DB, send top-K retrieved context + user query to a remote LLM (Gemini) for answer generation, and display results in a chat UI.
File structure
• Top-level “app” module (Gradle/Kotlin DSL) with:
• src/main/java/com/ml/.../docqa
– data (ObjectBox schemas, API key)
– domain (readers, splitters, embeddings provider, LLM API)
– di (AppModule)
– ui (MVVM screens + components)
• src/main/assets (ONNX embedding model, tokenizer.json)
• src/main/res (UI resources, prompt template in strings.xml)
• libs (sentence_embeddings.aar), objectbox-models (DB schema JSON)
• External config: proguard, backup rules, data extraction rules.
• Build: Gradle wrapper, versions TOML, GitHub Actions for release.
Extract architecture insights from README
• Clean-architecture layers: UI (Compose + ViewModels), Domain (use-cases: parse → split → embed → store → retrieve → prompt → call LLM), Data (ObjectBox vector store + tokenization + API).
• Technologies: Apache POI & iTextPDF (parsers), ONNX Runtime (embeddings), ObjectBox, Gemini Android SDK, Kotlin.
Elements
• Clear labels for each component/block (e.g., PDFReader, WhiteSpaceSplitter)
• Blue rectangles for UI components
• Green for domain services