An autonomous agent designed for automated research. It uses a dynamic tool selection mechanism and a vector-based memory to ingest and summarize information efficiently, while optimizing for cost through cloud infrastructure choices.
-
Automated Research: Capable of ingesting and summarizing information from multiple sources.
-
Dynamic Tool Selection: Uses prompt engineering to allow the LLM to choose the right tool for a task (e.g., Google Search, Arxiv API, Browse).
-
Vector-Based Memory: Employs a FAISS vector database to retrieve relevant information, providing the agent with long-term context.
-
Cost-Efficient Infrastructure: Leverages GCP Preemptible VMs and Docker to significantly reduce operational costs.
-
Feedback Loop: A rule-based system provides basic feedback to the agent to improve tool-use over time.
- Agent Core: Python, Asyncio, LangChain
- Model Backend: fine-tuned Llama 2 7B
- Memory: FAISS VectorDB
- Tooling: Google Search API, Arxiv API, Browse, Code Execution
- Infrastructure: GCP Preemptible VMs, FastAPI, Docker
-
Systems-Level Thinking: The ability to integrate multiple technologies (LLMs, vector databases, APIs, cloud infrastructure) into a single, functional system.
-
Advanced Concepts: A practical understanding of vector embeddings, dynamic tool usage, and prompt engineering.
-
Problem-Solving: The deliberate choice to use cost-saving infrastructure and a feedback loop shows a focus on practical, real-world constraints.
This project is licensed under the MIT License. Your agent should be yours to own and control.
We're open to open-source collaborations and R&D partnerships.