Question and Answering System with Rust and Burn Framework

Word Document Question Answering built in Rust with Burn framework
(developed in < 9 days)

Quick Summary

• Goal: Semantic QA over .docx documents
• Current state: Structured + semantic retrieval (TF-IDF + cosine)
• Language: Rust
• ML framework: Burn
• Status:
  • Phase 1 (Structured QA) → Complete
  • Phase 2 (Semantic QA) → Complete
  • Phase 3 (Full transformer training) → Planned

System Flow (Current)

Document Store (.docx / structured text) ↓ Document Loader ↓ Structured Data Representation ↓ Tokenization Layer ↓ Embedding Generation ↓ Vector Similarity Engine ↓ Semantic Retrieval Engine ↓ QA Inference Engine ↓ Command-Line Interface

Planned Transformer (Phase 3)

  Input Tokens
           ↓

Token Embeddings (256-dim) ↓ Positional Embeddings (256-dim) ↓ Transformer Encoder Layer 1 ↓ Transformer Encoder Layer 2 ↓ Transformer Encoder Layer 3 ↓ Transformer Encoder Layer 4 ↓ Transformer Encoder Layer 5 ↓ Transformer Encoder Layer 6 ↓ Contextual Representations ↓ Output Projection Layer ↓ Answer Span Prediction Head

Example Questions & Answers

• Q: What month and date will the 2026 End of year Graduation Ceremony be held?
  A: December 2026-12-31 – End of Year Graduation Ceremony

• Q: Which events occur in December?
  A: Day of Reconciliation, Christmas Day, End of Year Graduation Ceremony

• Q: What public holidays occur in 2026?
  A: Human Rights Day, Freedom Day, Workers Day, Youth Day, National Women’s Day, Heritage Day, Day of Reconciliation, Christmas Day

Key Design Choices

• Rust + Burn for performance & future ML pipeline
• Phased approach due to < 9-day constraint
• Modular structure (data / retrieval / inference separation)
• TF-IDF embeddings (fast, low-resource) → transformer planned

Challenges

• Extreme time limit (< 9 days) → solved with phased architecture
• Burn dev-dependencies issue → features = ["test"] commented out
• Large scope → modular + staged implementation

What Works Well

• Grounded answers from documents
• Stable semantic retrieval
• Clean, extensible architecture

Limitations (now)

• No neural model yet
• No contextual embeddings
• Basic multi-answer handling
• No temporal/entity reasoning

Next Steps (Phase 3)

• Burn Dataset + training pipeline
• Transformer-based QA model
• End-to-end training & checkpointing
• Neural inference

A functional ML system skeleton built under tight time pressure — ready to grow into full transformer-powered document QA.