johnjosephhorton

Research workflow representations

Three research tasks implemented with:

• EDSL
• the OpenAI Python SDK directly
• LangChain with langchain-openai

The comparison illustrates how each approach represents questions, scenarios, conditional survey flow, repetitions, models, and result normalization. It is

johnjosephhorton

import textwrap

from edsl.agents import Agent
from edsl.questions import (
    QuestionFreeText,
    QuestionLinearScale,
    QuestionMultipleChoice,
)
from edsl.surveys import Survey
from edsl.macros import Macro

johnjosephhorton

from edsl import (
    QuestionLinearScale,
    ScenarioList,
    Scenario,
    Scenario,
    Survey,
    QuestionCompute,
)

johnjosephhorton

#!/usr/bin/env python3
"""
Grid Image Tool

Adds red horizontal and vertical lines to an image and labels each cell with letters.
Dependencies are declared inline for use with uv run.

Usage: uv run grid_image.py input.png --num-vertical 5 --num-horizontal 3
"""

johnjosephhorton

# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "pip",
#   "edsl @ git+https://github.com/expectedparrot/edsl.git@improved_report",
# ]
# ///

from textwrap import dedent
from edsl import QuestionFreeText, QuestionList, Agent

johnjosephhorton

library(ggplot2)
library(dplyr)
library(tidyr)
library(gridExtra)
library(purrr)

# Function to compute medians and bootstrap CIs for a single tau value
get_medians <- function(data, simtot = 150, tau = 0.5) {
  # Data without errors
  data0 <- subset(data, all_errors == 0)

johnjosephhorton

from edsl import QuestionList, ScenarioList, QuestionFreeText
import random 

q = QuestionList(question_text = "Generate a list of 10 fruits", question_name = "fruits")
fruits = q()

def list_to_pref(items):
    return  " > ".join(items)

preferences = [list_to_pref(random.sample(fruits, len(fruits))) for _ in range(20)]

johnjosephhorton

import textwrap
from edsl import QuestionMultipleChoice, Scenario

pay_offs = {
    ('red', 'red'):(1, -1),
    ('red', 'black'):(-1, 1),
    ('black', 'red'):(1, -1),
    ('black', 'black'): (-1, 1)
}
instructions = textwrap.dedent("""\

johnjosephhorton

"""
This gives some examples of how to use SQL and ggplot2 together in the Results class.
Note that rather than call LLMs, we are created an agent with a direct-answering capability. 
After we create the agent, we create a `Results` object. 

"""
import random

from edsl import QuestionNumerical, QuestionFreeText
from edsl import Agent

johnjosephhorton

(venv) john@Johns-MacBook-Air agent_dialog % python -i AgentDialog.py
John: Hello Robin, I was thinking about going hiking today. Are you interested in joining me?
Robin: Hello John, hiking sounds like a great idea! I was actually hoping to go biking today, but a hike could be a nice compromise. Where were you thinking of hiking?
John: That's great to hear, Robin! I was thinking of hitting the trails at the Green Mountain Forest. There are some fantastic views and the terrain is just right for a good day's hike. How does that sound to you?
Robin: Green Mountain Forest sounds wonderful, John! I've heard the trails there are quite scenic. And you're right, it seems like the perfect place for a good day's hike. Let's do it! What time were you thinking of starting?
John: I'm glad you're on board, Robin! How about we start early to make the most of the day? Meeting at 8:00 AM would give us plenty of time to enjoy the trails. Does that work for you?
Robin: That sounds like a plan, John! Meeting at 8:00 AM works for