pythonlearner1025 / mariana

deep research agent

mariana - deep search agent

Search agent that does deep search to answer your question.

The core idea is that quality of search should scale with the value of answering a query (quantified by $ spend budget)

I'm betting that for queries like:

• which chemical compounds might have similar effects as ketamine? (drug discovery)
• How does the basal ganglia work and what can AI researchers learn from it? (AI research)
• Who's built zero-knowledge proof projects? (recruiting)

current search engines are bad at answering them, and the value of answering queries justifies long wait times & search cost.

Experimental, only supports arxiv search using Exa & arxiv API currently for research related queries.

how to run

1. create .env file and set your OpenAI, Anthropic, Groq api keys
2. python main.py --query "why do primate brains have a hippocampus and why might AI systems want one?" --budget 5 --agent openai

example reports

Check out some example reports Mariana generated in /reports. The query was:
Why do primate brains have a hippocampus and why might AI systems want one

observations

• claude-3-opus writes best reports, followed by gpt-4-turbo-2024-04-09
• quality of report with first 3 pages of papers in context >> only abstract of papers in context

TODO

• OpenAI & Groq integration
• logging
• more useful than one-shot generation / existing survey papers?
• how to measure quality vs. $ burn?
• continued learning: use memory & improve report
• clean pdf readings