Contexpurt

Contexpurt is a powerful, versatile tool designed to optimize interactions with Large Language Models (LLMs) by providing a dynamic and interactive context-building environment. It enables users to craft precise, relevant prompts for diverse applications, including querying codebases, managing personal knowledge bases, and handling work-related repositories.

Features

• Folder and Tag Management

  • Upload and organize files into virtual folders.
  • Enable or disable specific folders for inclusion in context searches.
  • Apply user-defined or AI-generated tags to documents for efficient filtering and retrieval.

• AI-Generated Metadata

  • Generate concise summaries for uploaded documents to provide an overview of their content.
  • Automatically extract and assign keywords to documents for improved searchability.

• Iterative Search and Expansion

  • Use embeddings to retrieve relevant chunks of data based on a user-entered query.
  • Dynamically expand retrieved results by including neighboring chunks before or after the selected text.
  • Allow users to uncheck irrelevant chunks and reassemble the context.

• Interactive Context Builder

  • Assemble context dynamically by selecting, reordering, and refining chunks of text.
  • Display the evolving prompt in a live preview window for validation.
  • Exclude irrelevant chunks using checkboxes or drag-and-drop removal.

• Persistent Context Storage

  • Save assembled contexts for reuse in future queries.
  • Maintain a version history of saved contexts, allowing users to revisit and refine past prompts.

• LLM Integration

  • Send the final assembled context along with the user’s query to the LLM.
  • Integrate user feedback into refining future searches and chunking strategies.

Technical Architecture

• Document Storage and Indexing

  • Store documents as Azure Storage Blobs for scalable and secure storage.
  • Index documents in the MSSQL database using vector embeddings, tags, folder structures, and metadata.
  • Utilize the new Vector data type in Azure MSSQL and SQL Server 2025 for embedding storage and semantic similarity search.
  • Leverage SQL functions such as VECTOR_DISTANCE, VECTOR_NORM, and VECTOR_NORMALIZE for advanced semantic similarity calculations and vector operations.

• API

  • Expose endpoints for document upload and tagging, embedding generation and storage, semantic search, and neighboring chunk retrieval.

• Frontend

  • Folder and tag managers.
  • Interactive context builder with drag-and-drop, checkboxes, and real-time preview.
  • Semantic search bar with expandable results.
  • Metadata display for retrieved chunks (e.g., source, section, relevance score).

• Integration with LLMs

  • Use OpenAI’s text-embedding-ada-002 or similar for chunk-level embeddings.
  • Pass assembled context and queries to the LLM (e.g., OpenAI GPT models).

Future Enhancements

• Advanced Search Features: Combine keyword-based filters with semantic search for hybrid retrieval. Implement AI-driven query suggestions.
• Context Summarization: Automatically summarize assembled contexts to save tokens.
• Collaboration Features: Enable real-time, multi-user context building sessions.
• Multi-Model Support: Integrate different LLMs for varied use cases (e.g., code-specific models like Codex).

Getting Started

Prerequisites

• .NET SDK
• Node.js
• Azure Storage Account
• MSSQL Database

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/contexpurt.git
   cd contexpurt

2. Install the necessary dependencies:

   cd src/UI
   npm install

3. Build the .NET solution:

   dotnet build src/Contexpurt.sln

Usage

1. Start the backend API:

   dotnet run --project src/Api

2. Start the frontend:

   cd src/UI
   npm run dev

3. Open your browser and navigate to http://localhost:3000 to start using Contexpurt.

Contributing

Contributions are welcome! Please read the contributing guidelines for more information.

License

This project is licensed under the MIT License. See the LICENSE file for details.