Llama Researcher

A demo of a personal research assistant using Cloudflare's Workers AI Llama-2 7b and D1. It demonstrates a simple use-case where an AI agent continuously researches on your behalf.

1. You first send an email to the worker (using Workers email routing) explaining topics you are interested in learning about. Llama converts the message to formatted topic labels and stores them in D1.
2. A cron trigger causes a random topic to be retrieved. That topic is converted by the model to several queries which are ultimately used to retrieve research papers from arXiv.
3. Each paper is summarized to be easier to understand (a step that probably is not needed for this API specifically, but possibly useful for other search APIs), formatted and then emailed to you with links to each paper.

Get Started

You'll need a Cloudflare account and a domain which is on Cloudflare to get started. This should work on the free plan(?)

git clone https://github.com/LoganGrasby/LlamaResearcher.git

cd LlamaResearcher

npm i wrangler -g (Ensure the latest version of wrangler is installed)

npm i (Install dependencies)

Configuration

There is a decent amount of configuration required to get started:

1. Configure email routing for a domain you have available on Cloudflare here.

2. Add a custom address (still in the email routing section) that points to this worker. In wrangler.toml this is the variable researcherEmail.

3. Add your personal email as a destination address. In wrangler.toml this is the userEmail variable.

4. Create a new D1 database wrangler d1 create researcher and database table wrangler d1 execute researcher --command "CREATE TABLE IF NOT EXISTS topics (id INTEGER PRIMARY KEY, topic TEXT NOT NULL)".

5. Fill in wrangler.toml with the required emails and database ID. Optionally change the cron trigger.

Usage

Get started by sending an email to the researcher email you configured. Describe practically any topic you're interested in that might be found on arXiv. The format of this email does not matter.

If everything is set up, you'll receive a confirmation email describing the topics llama ultimately picked. In the Cloudflare dashboard, you can check exactly how the topics are stored.

Testing

Test the cron trigger here:

wrangler dev --test-scheduled
curl "http://localhost:8787/__scheduled?cron=*+*+*+*+*"

The model will convert the topic to several additional queries, search for a recent paper with each query, summarize the paper and send you an email with all of the results!