This agent system, based on LangGraph, generates a survey on a requested topic using the most relevant papers on arXiv.
A FastAPI implementation is also provided, which serves:
/agent: an endpoint to the summary agent/query: an endpoint for semantic search
Both terminal use and Docker deployment require a Qdrant instance serving the following points:
| Field | Description |
|---|---|
id |
Random uuid |
vector |
Embedding for the paper |
title |
Title for the paper |
authors |
Authors of the paper |
summary |
Abstract for the paper |
categories |
arXiv categories for the paper |
published |
Date of publication on arXiv (format YYYY-MM-DD) |
arxiv-id |
arXiv ID for the paper |
All Qdrant and arXiv interactions are implemented in qdrant_tool.py and arxiv_tool.py.
Embeddings are implemented with SentenceTransformer, with the all-mpnet-base-v2 model by default.
Most configuration is done through environment variables:
| Variable | Description |
|---|---|
QDRANT_HOST |
Url for the Qdrant instance |
QDRANT_PORT |
Port for the Qdrant instance |
QDRANT_COLLECTION |
Name for the Qdrant collection to be used |
DEEPSEEK_API_KEY |
(Optional) If not provided, an Ollama instance is used |
Both Dockerfile and docker-compose.yml are provided.
To use the agent on the terminal, first install the python requirements (requirements.txt), then
python3 query.py
