ParseNote is an AI-powered learning tool that transforms large-scale, unstructured technical content—such as lecture transcripts or textbook excerpts—into semantically structured, high-quality documentation. The system uses large language models (LLMs), embedding-based similarity, and retrieval-augmented generation (RAG) to produce a distilled version of the content alongside a logically ordered outline. This distillation preserves core ideas while enhancing structure and clarity, forming the foundation for generating organized documentation and enabling intelligent retrieval through embeddings.
- OpenAI GPT-4: For outline generation and structured HTML note synthesis
- OpenAI text-embedding-3-small: For embedding and semantic similarity scoring
- Flask + Flask-SocketIO: For local web-based document processing and interaction
- Pandas, NumPy: For data handling, transformation, and CSV/PKL serialization
- Prompt Engineering: For controlling GPT output format and semantic consistency
The system begins with raw, unordered files containing academic or technical material. These may include lecture dumps, notes, or instructional documents in their original, unedited form.
The raw input is sent to a language model that produces a high-level semantic outline:
- Topics are ordered from foundational to advanced.
- The outline acts as both a distilled version of the original content and a semantic guide for organizing the document.
- Output is stored as a serialized DataFrame for downstream use.
To semantically align the unstructured input with the generated outline, the system applies a sliding window embedding strategy combined with cosine similarity scoring. This functions as a custom topic modeling system:
- Chunks of the input are embedded using OpenAI’s
text-embedding-3-small. - Each chunk is compared against precomputed embeddings of outline sections.
- Chunks are assigned to the section with the highest semantic similarity score.
This creates a meaning-based mapping of unordered content into semantically coherent sections without relying on its original sequence or formatting.
Each outline section is embedded and stored to support semantic search and question answering. These embeddings enable high-accuracy chunk retrieval during inference or user interaction.
Organized content sections are sent back to the LLM for structured note generation:
- Each section is rewritten as coherent, technical documentation.
- Output is formatted as valid HTML, ready for integration into Jinja templates or rendered directly via the web interface.
- All content maintains semantic fidelity to the original material while improving clarity and structure.
A built-in chat interface enables users to query the document:
- The query is embedded and used to retrieve the most relevant section(s) based on similarity to the stored embeddings.
- The system includes the distilled document as part of the context.
- GPT-4 generates an answer using both the retrieved section and distilled document to ensure completeness and specificity.
- Clone the repository:
git clone [https://github.com/jleto6/parse-note.git](https://github.com/jleto6/parse-note.git)
- Navigate to the directory:
cd parse-note - Install dependencies:
pip install -r requirements.txt
- Set up API keys:
export OPENAI_API_KEY=your_key_here - Run the application:
python app/main.py
A Flask server will start and open in your browser. Upload your raw content—this can include text files, lecture transcripts, PDFs, or images—and the system will automatically generate a semantic outline, organize the material, and return structured documentation ready for retrieval, analysis, or integration.
- Full-featured web interface with user logins, saved sessions, and workspace management
- PDF and Markdown export of generated documentation
- Human-in-the-loop feedback system for iterative refinement of generated documentation