A lightweight, pure web search solution for large language models, supporting multi-engine aggregated search, deep reflection and result evaluation. A balanced approach between web search and deep research, providing a framework-free implementation for easy developer integration.
Web search is a common feature for large language models, but traditional solutions have limitations:
- Limited search result quality and reflection effectiveness
- Requires powerful models and paid search engines
- Small models often struggle with tool calling patterns
- Contextual understanding can be unstable across different model sizes
Deep Search Lighting provides:
- Framework-free implementation with no restrictions
- Works with free APIs while maintaining good query quality
- Adjustable depth parameters to balance speed and results
- Reflection mechanism for model self-evaluation
- Supports models of any size, including smaller ones
[Experimental Planning]:
- Simplified design without web parsing or text chunking
- Considering adding RL-trained small recall models
-
Multi-engine aggregated search:
- ✅ Baidu (free)
- ✅ DuckDuckGo (free but requires VPN)
- ✅ Bocha (requires API key)
- ✅ Tavily (requires registration key)
-
Reflection strategies and controllable evaluation
-
Custom pipelines for all LLM models
-
OpenAI-style API compatibility
-
Pure model source code for easy integration
-
Built-in MCP server support
