DocuAction is an AI-powered document analysis system designed to help users extract actionable insights from unstructured documents.
The project focuses on building a practical Retrieval-Augmented Generation (RAG) workflow that improves response grounding, traceability, and document understanding using modern LLM tooling.
Large language models often struggle with hallucination when answering questions about long documents.
DocuAction addresses this by combining:
- Document ingestion and chunking
- Vector-based semantic search
- Retrieval-Augmented Generation (RAG)
- Multi-step reasoning workflows
This allows the system to generate responses that are grounded in the original documents rather than relying solely on model knowledge.
- Upload and process unstructured documents
- Automatically split documents into semantic chunks
- Store embeddings for efficient retrieval
- Retrieve relevant document segments based on user queries
- Use similarity search to locate supporting evidence
- Retrieve relevant context from the vector store
- Provide grounded responses using LLM reasoning
- Improve response traceability by linking answers to source documents
- Enables question-answering over large document collections
- Reduces hallucination by grounding outputs in retrieved context
User Query
│
▼
Semantic Retrieval (Vector Search)
│
▼
Relevant Document Chunks
│
▼
LLM Reasoning (RAG Prompt)
│
▼
Grounded Response + Source Context
The system separates retrieval, reasoning, and response generation, allowing more reliable document understanding.
- Python
- FastAPI
- LangChain
- Vector Embeddings / Semantic Search
- LLM-based reasoning
- Retrieval-Augmented Generation (RAG)
Working with LLM systems highlighted a common challenge:
models can generate confident but incorrect answers when working with large documents.
DocuAction was built to explore practical approaches to:
- improving grounding in LLM outputs
- building reliable document QA pipelines
- structuring multi-step reasoning workflows for LLM applications
The project focuses on practical system design rather than just prompt experimentation, emphasizing reproducible pipelines and clear retrieval logic.
Planned improvements include:
- Document highlighting and citation tracing
- Evaluation pipelines for hallucination detection
- Multi-agent orchestration for document analysis
- UI for interactive document exploration
GitHub:
https://github.com/minjikim19/DocuAction
Minji Kim
AI / Data Systems Developer
Portfolio: https://minjikim19.github.io
LinkedIn: https://linkedin.com/in/minji-kim19