shubham0204 / android-document-qa

A custom RAG pipeline for multi-document QA from PDF/DOCX documents, in Android

Shubham Panchal

Reach me at

• Send a message on [email protected]
• LinkedIn
• Reddit

Profiles

• Website
• LinkedIn
• Medium
• StackExchange
• Linktree (Link Aggregator)

OSS Projects

Rust

• text-predictor-android: Next word prediction and word auto-complete for Android - with the power of Rust
• tfidf-summarizer.rs : Simple, efficient and cross-platform TFIDF-based text summarizer in Rust
• postagger.rs : NLTK inspired Parts-of-Speech Tagger (Perceptron Tagger) in Rust
• pq.rs: Pure Rust implementation of Product Quantization for Approximate Nearest Neighbor Search
• diff-tool.rs: A simple terminal-based diff tool written in Rust

On-Device ML

• Segment Anything Inference On Android With ONNX
• On-Device Face Recognition with FaceNet ( GitHub )
• Sentence-Embeddings in Android - all-MiniLM-L2-V6 ( GitHub )
• On-Device RAG in Android for Document Q&A
• Depth-Anything v1/v2 Inference On Android With ONNX
• Deploying Scikit-Learn Models In Android Apps With ONNX ( GitHub )
• Deploying TF models on Heroku for Android ( GitHub )
• Realtime Depth Estimation In Android ( GitHub )
• MLKit Selfie Segmentation in Android ( GitHub )
• Using FaceNet For On-Device Face Recognition With Android ( GitHub )
• Text2Summary Android ( GitHub )
• Q-Learning With The Frozen Lake Environment In Android ( GitHub )
• Hyperparameter Optimization With Genetic Algorithms in Kotlin ( GitHub )
• Exploring Random Forests In Light Of Kotlin ( GitHub )
• Coding Feed-Forward Neural Networks in Kotlin (or Android) ( GitHub )
• Gaussian Naive Bayes ( for Iris Classification ) in Android ( GitHub )
• How I made Skinly for Melanoma Detection in Android ( GitHub )
• Designing Decision Trees From Scratch on Android ( GitHub )
• Text Classification in Android with TensorFlow ( GitHub )
• Sarcasm Detection using Word Embeddings in Android ( GitHub )
• Hands-on With Multiple Linear Regression on Android
• Bayes Text Classification in Kotlin for Android without TensorFlow
• Introducing TensorFlow Lite Android Support Library

Web

• shubham0204.github.io: My portfolio built with HTML 5, CSS and Bootstrap 5.3
• WASM-JS-Algorithms: Implementation of various algorithms in WebAssembly and vanilla Javascript

Backend

• Recruiting/Shortlisting CVs with OpenAI's API (FastAPI)
• blogpost-backend (FastAPI)
• INC-DBMS-Project (Ktor)

Data Structures / Algorithms

• text-predictor-android (Trie for text-completion)
• glove.c (Hashtable with open hashing / closed addressing for storing embeddings)
• full-text-search (Inverted Index)
• diff-tool.rs (Longest common subsequence using backtracking)

Kotlin/Android

• google-search-kotlin: Get Google Search results in Kotlin
• Document Scanner
• DistilBert

Blogs

• Building On-Device Face Recognition In Android
• From Python To Android: HF Sentence Transformers (Embeddings)
• WebAssembly: Everything You Need To Know
• On-Device Machine Learning In Android: Frameworks and Ecosystem
• Using C/C++ in Android: A Comprehensive Guide For Beginners
• Building A Cross-Platform TFIDF Text Summarizer In Rust
• Managing Deep Learning Models Easily With TOML Configurations
• glove-android: Using GloVe Word Embeddings for NLP In Android
• Chaquopy: Using Python In Android Apps
• Using Websockets To Run Terminal Commands From Android Apps
• Creating Desktop Apps For TensorFlow Models With Qt
• Building a Simple Document Scanner in Android
• Deploying Scikit-Learn Models In Android Apps With ONNX
• Empirical Distribution — Everything You Need To Know
• Human-In-The-Loop Systems — All You Need To Know
• Deploying TF Models on Heroku for Android Apps
• Accessing App Usage History In Android
• No, Kernels & Filters Are Not The Same
• Deploying Pretrained TF Object Detection Models on Android
• Realtime Selfie Segmentation In Android With MLKit
• Grad-CAM: A Camera For Your Model’s Decision
• MLP Mixer Is All You Need?
• Age + Gender Estimation in Android with TensorFlow
• Using FaceNet For On-Device Face Recognition With Android
• Coding Feed-Forward Neural Networks in Kotlin (or Android)
• Q-Learning With The Frozen Lake Environment In Android
• Hyperparameter Optimization With Genetic Algorithms In Kotlin
• Image Classification With TensorFlow 2.0 ( Without Keras )
• Get Started With TensorFlow 2.0 and Linear Regression
• Cityscape Image Segmentation With TensorFlow 2.0
• 4 Awesome Ways Of Loading ML Data In Google Colab
• Neural Implanting With AutoEncoders and TensorFlow
• Colorizing B/W Images With GANs in TensorFlow
• Exploring Random Forests In Light Of Kotlin
• Gaussian Naive Bayes ( for Iris Classification ) in Android
• Sarcasm Detection using Word Embeddings in Android
• Hands-on With Multiple Linear Regression on Android
• Bayes Text Classification in Kotlin for Android without TensorFlow
• How I made Skinly for Melanoma Detection in Android
• Text Classification in Android with TensorFlow
• Designing Decision Trees From Scratch on Android
• Introducing TensorFlow Lite Android Support Library
• Introducing Text2Summary: Text Summarization on Android

Demystifying Mathematics - Shubham Panchal

• Kernels: Everything You Need to Know
• Principal Component Analysis: Everything You Need To Know
• Demystifying Monte Carlo Integration
• Demystifying Linear Independence
• Demystifying Probability Distributions ( 1 / 3 )
• Demystifying Probability Distributions ( 2 / 3 )
• Demystifying Probability Distributions ( 3 / 3 )
• Demystifying The Moore-Penrose Generalized Inverse
• Taylor Series And The Power Of Approximation
• Entropy In The World Of Computer Science

Google Colab Notebooks - Tutorials

• Age + Gender Estimation in Android with TensorFlow
• Hyperparameter Optimization Using Keras Tuner API
• Image Colorization With GANs
• Chatbot using Seq2Seq LSTM models.
• Neural Machine Translation (NMT) - Translating English sentences to French sentences.
• Neural Machine Translation (NMT) - Translating English sentences to Marathi sentences.
• Image Classification From Scratch With TensorFlow 2.0 ( Without Keras! )
• Cityscape Image Segmentation with TensorFlow 2.0 ( Without Keras! )
• Face Landmark Detection With TensorFlow
• Neural Implanting With TensorFlow
• Tweet Classification with MLP-Mixers ( TF-Keras ) ( Story On Medium )
• From Paper To Keras: MobileNets With TensorFlow
• From Paper To Keras: DenseNets With TensorFlow

Spotlight

• Age + Gender Estimation in Android with TensorFlow used in Realtime Age and Gender estimation using MLKit and Tensor Flow Android Studio | Visitor Analysis YouTube video by The Mobile Dev
• Notebook Tweet Classification With MLP-Mixers ( TF-Keras referred by the paper Distinguish risk communication and crisis communication based on deep learning model

Collaborations / Pull Requests

• handtracking Android example to victordibia/handtracking
• depth-estimation Android sample to SunitRoy2703/Tensorflow-lite-kotlin-samples
• air_drawing demo (a sample for hand landmark detection solution) to googlesamples/mediapipe
• script for building Debian package to ataractic/libcfg
• article How to Integrate Firebase Authentication with the Ktor Auth Feature to mjovanc/awesome-ktor
• multiple Kotlin libraries like kotlindl, dotenv-kotlin, multik etc. to mcxiaoke/awesome-kotlin
• Android demo for LiheYoung/Depth-Anything
• Rust ML frameworks to josephmisiti/awesome-machine-learning
• Corrected Android sample code in README.md of codeandtheory/YCharts
• Added Mediapipe and ExecuTorch frameworks to merveenoyan/awesome-osml-for-devs
• Added gnp/minbpe-rs to karpathy/minbpe as a community extension
• Added rust-bert, lancedb and faer crates in rust-unofficial/awesome-rust

android-document-qa's Issues

local llm, with pdf rag feature

Android 14 Issue

Seems like TextEmbedder causes the application to crash on Android 14

Question about the distance function

I'm seeing you are using the default distance function (Euclidean), as it is not specified:
@HnswIndex(dimensions = 100) var chunkEmbedding: FloatArray = floatArrayOf()

I'm not sure which embedding model you are using. In RAG, often Cosine is preferable. Or, if you know that vectors are normalized, you could use the faster DotProduct.