Build a multi-agent research assistant powered by context engineering — integrating documents, web search, memory, and research papers to deliver accurate, contextual responses.
This project demonstrates how to apply Context Engineering (CE) principles to build a multi-agent system that gathers and filters information from multiple sources before generating an intelligent response.
Traditional prompt engineering focuses on “magic words.”
Context Engineering, on the other hand, ensures that LLMs receive:
- ✅ The right information
- 🧰 The right tools
- 🧩 In the right format
This allows the model to reason more effectively and generate more accurate, grounded outputs.
The assistant gathers, filters, and synthesizes information using the following context engineering pipeline:
- User submits a query
- Fetch context from:
- 📄 Documents (via Tensorlake)
- 🌐 Web (via Firecrawl)
- 🧠 Memory (via Zep)
- 📚 Research papers (via ArXiv API)
- Aggregate and filter context using a context evaluation agent
- Generate response using a synthesizer agent
- Save the final response back to memory (Zep)
This project is adapted from the “Daily Dose of DS” article on Context Engineering.
Original concept and architecture credit go to Daily Dose of DS.
🔗 Source: Daily Dose of DS – Context Engineering Demo (original thread reference)
In this implementation,
🧩 I have replaced the original LLM integration with the Gemini API for response generation, while maintaining the same multi-agent workflow structure.