What is Kheish
Kheish is an open-source platform for multi-agent orchestration based on large language models (LLMs). It uses multiple specialized roles (agents) and flexible workflows to coordinate the various steps of complex tasks, such as proposal generation, review, validation, and formatting, producing high-quality results. The platform can seamlessly integrate external modules, such as file system access, shell commands, and vector storage, enriching context and handling large codebases or documents. With features such as modularity, extensibility, feedback loops, and chat-style prompting, Kheish supports customization and extension to new tasks or domains.
Kheish's main features
- Multi-agent workflow: Define multiple roles (proposer, reviewer, validator, formatter) and execute them in sequence according to a YAML-based workflow configuration to handle complex tasks.
- Module integration: Integrate multiple modules, such as file system access (fs), shell commands (sh), vector storage (vector_store), etc., to extend functionality and processing capabilities.
- Chat-style prompting: Interaction with the LLM adopts a conversational structure, including system, user, and assistant, maintaining context and clarifying instructions.
- Feedback and revision loop: The workflow supports iterative feedback. If the reviewer or validator finds problems, it requests revisions until the solution meets the criteria.
- RAG and embeddings: Integrate vector storage to handle large documents, allowing the model to retrieve relevant snippets without overloading the prompt.
- Modularity and extensibility: Supports adding new modules or agents without breaking the existing structure, encouraging customization and extension to new tasks or domains.
Kheish's technical principles
- Task manager: Design a task manager responsible for loading task configurations (YAML), processing context, running workflows, and integrating modules as needed.
- Context processing: The task manager reads the initial file or text, prepares the TaskContext, and provides contextual information for the subsequent agent workflow.
- Workflow execution: Agents execute tasks in the defined workflow order, including generating solutions, checking proposal quality, ensuring final correctness, and outputting validation results.
- Module request invocation: Agents can issue MODULE_REQUEST calls to obtain other data, read files, or perform semantic searches, etc.
- Result caching: To avoid repeated operations, Kheish caches the results of agent requests to modules.
- Integration with RAG: Integrates with retrieval-augmented generation (RAG) models to easily handle large documents, retrieving relevant snippets through vector storage.
Kheish's project address
- GitHub repository：https://github.com/graniet/kheish
Kheish's application scenarios
- Code audit: Used in automated code review to identify security vulnerabilities and quality issues in code through agent workflows, and propose improvement suggestions.
- Legal document analysis: In the legal field, process and analyze large volumes of legal documents, extract key information, and assist legal research and case preparation.
- Customer service automation: As a chatbot, handle customer inquiries and support requests, providing immediate automated responses.
- Content creation assistant: Assists content creators in generating and editing articles, blogs, and other marketing materials, improving the quality and appeal of the content.
- Personalized education platform: Provides personalized learning materials and practice questions based on students' learning progress, assisting online education and remote learning.