stevekrenzel / pick-ems

An LLM agent built on top of OpenAI that predicts winners for ESPN's Pick-Em's game.

🏈 Pick-Em's LLM Bot

This is an LLM agent built on top of OpenAI that predicts winners for ESPN's Pick-Em's game.

For the 2023 NFL season, this agent won my local group's pick-em's competition and came in the top 15% across all of ESPN's users (~one million users played). The top player across all of ESPN for 2023 correctly chose 195 games (~71%). This agent correctly chose 163 (~60%).

How Does It Work?

At a high-level, the agent crawls a bunch of statistics and news articles on ESPN's website. For each match in a given week it will feed the LLM stats and news relevant to that match and have it predict a winner.

The data retrieval is done by scraping ESPN pages (built using Playwright).

The data analysis is done by the LLM.

The general division here is that anything that can be done more-or-less deterministically in code, we should do in code. And fallback to the LLM for very specific tasks that are fuzzy, non-deterministic, and don't lend themselves to code due to their ambiguous or difficult-to-code nature.

Scrapers

The scrapers know how to retrieve:

• The list of all NFL teams.
• The list of all matches for the week.
• Offense, Defense, Turnover, and Special statistics for all teams.
• Headlines from the front page and the content of the articles for those headlines.

Agents

The agents know how to:

• Analyze News: Take a news article and extract:

  1. The primary team it is written about
  2. A summary of the article
  3. Whether it refers to the NFL, College, or Fantasy Football.

  We need that last bit because on ESPN's NFL site, a lot of articles get written about things that have nothing to do with the NFL regular season and we want to filter that content out prior to giving it to the LLM.

• Predict a Winner: Given offense, defense, special, and turnover stats along with any relevant news we could find (e.g. injuries), for any given match up we give the LLM the stats and news for the two teams, and have it analyze that data to make a prediction about who will win.

Getting Started

To get started, you first need to clone the code and install your dependencies.

$ git clone [e-mail]:stevekrenzel/pick-ems.git
$ npm install

And then update the .env file to have a valid OpenAI API key.

And then you can run the agent with:

$ npm run start

It may take several minutes to complete. It analyzes many articles and matches and each call to the LLM may take upwards of ~30 seconds to complete. So be patient.

Architectural Patterns

There are a few key architectural patterns we use in this repo. One for data retrieval, one for working with LLMs, and one for managing the flow of data between the two.

Data Access

Web scraping can be messy business, so we attempt to hide the browser from the rest of our code as quickly as possible. The end goal is to basically access the content from the various webpages the same way we would access data from an API or a database. To that end, we use a Data Mapper Pattern where each kind of data (e.g. Article, Match, Team, etc...) has a Repo and an Entity. The Entity contains all of the fields of data that we want that domain object to have. The Repo is how we retrieve and access the data.

The repos are structured in such a way that if you just use the objects without peeking behind the scenes, you'd have no idea that you weren't querying a database of some sort (though, an admittedly slow database).

Agents

Bridging the gap between nice well-defined data and fuzzy natural language can be a bit tricky. To help address this, we rely on OpenAI's ability to call tools/functions. We don't actually care about the tool itself. We pass OpenAI exactly one tool, force it to use that tool, and the only bit of the tool we care about are the parameters of the tool. This is the data that we are seeking from the LLM. The tool definition is just a means for us to provide OpenAI a nice JSONSchema that will influence the shape of the data the LLM returns to us.

Note: The LLM may return something that doesn't match our schema, but with GPT-4 this is exceedingly uncommon (at least for our use case).

Each agent is broken up into 3 parts: Prompt, Schema, Tool.

• The Prompt instructs the LLM on what its task is and how it should achieve it.
• The Schema instructs the LLM on what data we expect it to return to us.
• The Tool retrieves any data the LLM might need for its task and calls the LLM with the prompt, data, and schema and wraps it all in a nice function that abstracts away the details. A user can call a tool just like any other function and they can be oblivious to the fact that the string or boolean they got back required the processing power of 1,000 remote GPUs.

Analysis and Conclusions

One other pattern we use, that may be non-obvious, is that we wrap the tool's schema in a parent schema that has a field called analysis and a field called conclusion. The conclusion field isn't particularly notable and simply maps to the tool's schema. The analysis is the notable one here. Importantly, The field is generated first by the LLM and is a way for the LLM to "think" out loud and reference that thinking in the subsequent generation of the data in the conclusion.

A lot of LLM techniques, like Chain-of-Thought (CoT) require that the LLM generates a sequence of tokens that it can then reference on for its final answer. If your Prompt to the Tool asks the LLM to use any strategy like this, and you don't give it a scratchpad of sorts to write in, then it won't be able to use that pattern. So this analysis field is a scratchpad that we offer to the LLM to write whatever it needs to before answering our prompt.

tl;dr; If you ask the LLM to return true or false, and you instruct it to think carefully about the choice beforehand, but you don't give it space to do that thinking then it won't be able to and instead will just return a boolean without thinking carefully about it.

This pattern meaningfully improves the results of the payloads.

Model-View-Controller

Finally, we've got data retrieval and we've got agents, but how do we think about the interplay between the two? We draw inspiration from the Model-View-Controller (MVC) pattern.

In this case, our:

• Models map to Repos
• Views map to Prompts
• Controllers map to Tools

The Tool is the thing that coordinates retrieving data from the Repos and then rendering that data into a Prompt prior to sending it to our "client", the LLM.

The analogy is not perfect, and starts to stretch under scrutiny, but as a rough guide on how to think about the division of labor between the components, I find it useful.