Surveys

A Survey is collection of Questions that can be administered to one or more AI Agents and Language Models at once. Survey questions can be administered asynchronously (by default), or according to rules such as skip and stop logic, and with or without context of other questions in a survey.

Surveys can be used to collect data, generate content or perform other tasks. The results of a survey are stored in a Results object, which can be used to analyze the responses and other components of the survey. By default, the results object also has a cache of the responses that can be reused, which are also added to your local cache or the remote cache if you are running a survey remotely. Learn more about built-in methods for working with Results objects in the Results section and caching responses in the sections on data and Remote Caching.

Key steps

The key steps to creating and conducting a survey are:

1. Create Questions of various types (multiple choice, checkbox, free text, numerical, linear scale, etc.) and combine them in a Survey to administer them together.

2. Optional: Add rules to skip, stop or administer questions based on conditional logic, or provide context of other questions and answers in the survey.

3. Optional: Design personas for AI Agents to answer the questions.

4. Decide whether to use Remote Inference or your own Managing Keys to run the survey.

5. Send the survey to language Models of your choice to generate the responses.

Sending a survey to a language model generates a dataset of Results that includes the responses and other components of the survey. Results can be analyzed and visualized using built-in methods of the Results object.

Key methods

A survey is administered by calling the run() method on the Survey object, after adding any agents, scenarios and models with the by() method, and any rules or memory with the appropriate methods (see examples of each below):

• add_skip_rule() - Skip a question based on a conditional expression (e.g., based on a response to another question).

• add_stop_rule() - End the survey based on a conditional expression.

• add_rule() - Administer a specified question next based on a conditional expression.

• set_full_memory_mode() - Include a memory of all prior questions/answers at each new question in the survey.

• set_lagged_memory() - Include a memory of a specified number of prior questions/answers at each new question in the survey.

• add_targeted_memory() - Include a memory of a particular question/answer at another question in the survey.

• add_memory_collection() - Include memories of a set of prior questions/answers at another question in the survey.

Before running a survey, you can also estimate the cost of running it with the estimate_job_cost() method of a Job object (a survey combined with one or more models). The show_prompts(), show_rules() and show_flow() methods can be used to analyze the structure of a survey and the rules that have been applied to it:

• show_prompts() - Display the user and system prompts for each question in a survey. This is a companion method to the prompts() method of a Job object, which returns a dataset containing the prompts together with information about each question, scenario, agent, model and estimated cost.

• show_rules() - Display a table of the conditional rules that have been applied to a survey.

• show_flow() - Display a graphic of the flow of a survey, showing the order of questions and any rules that have been applied, and any scenarios and/or agent information that has been added.

When you run a survey you can choose to run it remotely at the Expected Parrot server or locally on your own machine. See Remote Inference for more information.

New feature in progress: When you run a job remotely you automatically have access to a universal remote cache of stored responses. Learn more about it in the Remote Caching section.

You can also choose to run a remote survey in the background by passing the background=True parameter to the run() method. This allows you to continue working (or stop working) while your job completes. You can check progress at the progress bar page or by calling the fetch() method at any time.

Piping

You can pipe individual components of questions into other questions, such as inserting the answer to a question in the question text of another question. This is done by using the {{ question_name.answer }} syntax in the text of a question, and is useful for creating dynamic surveys that reference prior answers.

Note that this method is different from memory rules, which automatically inlude the full context of a specified question at a new question in the survey: “Before the question you are now answering, you already answered the following question(s): Question: <question_text> Answer: <answer>”. See examples below.

Constructing a survey

In the examples below we construct a simple survey of questions, and then demonstrate how to run it with various rules and memory options, how to add AI agents and language models, and how to analyze the results.

Defining questions

Questions can be defined as various types, including multiple choice, checkbox, free text, linear scale, numerical and other types. The formats are defined in the Questions module. Here we define some questions by importing question types and creating instances of them:

from edsl import QuestionMultipleChoice, QuestionCheckBox, QuestionLinearScale, QuestionNumerical

q1 = QuestionMultipleChoice(
   question_name = "consume_local_news",
   question_text = "How often do you consume local news?",
   question_options = ["Daily", "Weekly", "Monthly", "Never"]
)

q2 = QuestionCheckBox(
   question_name = "sources",
   question_text = "What are your most common sources of local news? (Select all that apply)",
   question_options = ["Television", "Newspaper", "Online news websites", "Social Media", "Radio", "Other"]
)

q3 = QuestionLinearScale(
   question_name = "rate_coverage",
   question_text = "On a scale of 1 to 10, how would you rate the quality of local news coverage in your area?",
   question_options = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
   option_labels = {1: "Very poor", 10: "Excellent"}
)

q4 = QuestionNumerical(
   question_name = "minutes_reading",
   question_text = "On average, how many minutes do you spend consuming local news each day?",
   min_value = 0, # optional
   max_value = 1440 # optional
)

Adding questions to a survey

Questions are passed to a Survey object as a list of question ids:

from edsl import Survey

survey = Survey(questions = [q1, q2, q3, q4])

Alternatively, questions can be added to a survey one at a time:

from edsl import Survey

survey = Survey().add_question(q1).add_question(q2).add_question(q3).add_question(q4)

Running a survey

Once constructed, a survey can be administered by calling the run() method. If question Scenarios, Agents or Language Models have been specified, they are added to the survey with the by method when running it. (If no language model is specified, the survey will be run with the default model, which can be inspected by running Model().)

For example, here we run the survey with a simple agent persona and specify that GPT-4o should be used. Note that the agent and model can be added in either order, so long as each type of component is added at once (e.g., if using multiple agents or models, pass them as a list to the by() method):

from edsl import Agent, Model

agent = Agent(traits = {"persona": "You are a teenager who hates reading."})

model = Model("gpt-4o")

results = survey.by(agent).by(model).run()

Job status information

When you run a survey, you will see information in the console about the status of the job. When the job completes, you can access the Results object that is generated in your workspace. If you are running a survey remotely, you will also see a link to the results at Coop.

For example, if we ran the survey above remotely we would see a link to the results in the console:

Results: https://www.expectedparrot.com/content/4cfcf0c6-6aff-4447-90cb-cd9e01111a28.

Progress report

While a job is running you can view updates in a Progress Report. If remote inference is activated, a link to a Progress Report will appear automatically. If you are running a survey locally, you can pass run(progress_bar=True) to view a report locally.

Exceptions Report

If any exceptions are generated you can view details about them in an Exceptions Report, which includes informatoin about the questions, agents, scenarios and models that generated exceptions together with error messages and tracebacks. If remote inference is activated, a link to an Exceptions Report will appear automatically. If you are running a survey locally, the report details will appear in your console.

Running jobs in the background

If you are running a survey remotely, you can choose to run it in the background by passing the background=True parameter to the run() method:

results = survey.by(agent).by(model).run(background=True)

This allows you to continue working (or stop working) while your job completes. You can check progress at any time at the progress bar page or by calling the fetch() method, which checks for the results every 1.0 seconds by default or a specified interval:

results = survey.by(agent).by(model).run(background=True)

# check progress every 5 seconds
results.fetch(polling_interval = 5.0)

Optional parameters

There are optional parameters that can be passed to the run() method, including:

• n - The number of responses to generate for each question (default is 1). Example: run(n=5) will administer the same exact question (and scenario, if any) to an agent and model 5 times.

• cache - A boolean value to cache the results of the survey. The default is True; the cache for the survey is automatically added to the Results object that is generated. Example: run(cache=False) will generate fresh responses.

• disable_remote_inference - A boolean value to indicate whether to run the survey locally while remote inference is activated (default is False). Example: run(disable_remote_inference=True).

• remote_inference_results_visibility - A string value to indicate the visibility of the results on the Expected Parrot server, when a survey is being run remotely. Possible values are “public”, “unlisted” or “private” (default is “unlisted”). Visibility can also be modified at the Coop web app. Example: run(remote_inference_results_visibility=”public”).

• progress_bar=True - This parameter can be used to view a Progress Report locally. A link to a Progress Report will automatically be provided when you run a survey remotely.

• background=True - This parameter can be used to run a survey in the background, allowing you to continue working (or stop working) while your job completes.

• polling_interval - This parameter can be used to specify the interval (in seconds) at which to check for results when running a survey in the background. The default is 1.0 seconds. Example: run(polling_interval=5.0) will check for results every 5 seconds.

Survey rules & logic

Rules can be applied to a survey with the add_skip_rule(), add_stop_rule() and add_rule() methods, which take a logical expression and the relevant questions.

Skip rules

The add_skip_rule() method skips a question if a condition is met. The (2) required parameters are the question to skip and the condition to evaluate.

Here we use add_skip_rule() to skip q2 if the response to “consume_local_news” is “Never”. Note that we can refer to the question to be skipped using either the id (“q2”) or question_name (“consume_local_news”):

from edsl import Survey

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_skip_rule(q2, "consume_local_news == 'Never'")

This is equivalent:

from edsl import Survey

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_skip_rule("sources", "consume_local_news == 'Never'")

We can run the survey and verify that the rule was applied:

results = survey.by(agent).by(model).run() # using the agent and model from the previous example
results.select("consume_local_news", "sources", "rate_coverage", "minutes_reading")

This will print the answers, showing “None” for a skipped question (your own results for answers may vary):

answer.consume_local_news	answer.sources	answer.rate_coverage	answer.minutes_reading
Never	None	4	0

Stop rules

The add_stop_rule() method stops the survey if a condition is met. The (2) required parameters are the question to stop at and the condition to evaluate.

Here we use add_stop_rule() to end the survey at q1 if the response is “Never” (note that we recreate the survey to demonstrate the stop rule alone):

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_stop_rule(q1, "consume_local_news == 'Never'")

This time we see that the survey ended when the response to “consume_local_news” was “Never”:

results = survey.by(agent).run()
results.select("consume_local_news", "sources", "rate_coverage", "minutes_reading")

Output:

answer.consume_local_news	answer.sources	answer.rate_coverage	answer.minutes_reading
Never	None	None	None

Other rules

The generalizable add_rule() method is used to specify the next question to administer based on a condition. The (3) required parameters are the question to evaluate, the condition to evaluate, and the question to administer next.

Here we use add_rule() to specify that if the response to “color” is “Blue” then q4 should be administered next:

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_rule(q1, "consume_local_news == 'Never'", q4)

We can run the survey and verify that the rule was applied:

results = survey.by(agent).run()
results.select("consume_local_news", "sources", "rate_coverage", "minutes_reading")

We can see that both q2 and q3 were skipped but q4 was administered (and the response makes sense for the agent):

answer.consume_local_news	answer.sources	answer.rate_coverage	answer.minutes_reading
Never	None	None	0

Show flow

We can call the show_flow() method to display a graphic of the flow of the survey, and verify how any rules were applied. For example, here we show the flow of the survey above with the skip rule applied:

survey = Survey(questions = [q1, q2, q3, q4])

survey = survey.add_skip_rule(q2, "consume_local_news == 'Never'")

survey.show_flow()

Output:

If we add agent or scenario details to the survey questions, the flow diagram will also show this information. Here we modify the survey questions from above to include an agent trait:

from edsl import QuestionMultipleChoice, QuestionCheckBox, QuestionLinearScale, QuestionNumerical, Survey, Agent

agent = Agent(traits = {
   "nickname": "Robin",
   "persona": "You are a teenager who hates reading."
})

# Adding the nickname to the question texts
q1 = QuestionMultipleChoice(
   question_name = "consume_local_news",
   question_text = "Hey {{ agent.nickname }}, how often do you consume local news?",
   question_options = ["Daily", "Weekly", "Monthly", "Never"]
)

q2 = QuestionCheckBox(
   question_name = "sources",
   question_text = "{{ agent.nickname }}, what are your most common sources of local news? (Select all that apply)",
   question_options = ["Television", "Newspaper", "Online news websites", "Social Media", "Radio", "Other"]
)

q3 = QuestionLinearScale(
   question_name = "rate_coverage",
   question_text = "{{ agent.nickname }}, on a scale of 1 to 10, how would you rate the quality of local news coverage in your area?",
   question_options = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
   option_labels = {1: "Very poor", 10: "Excellent"}
)

q4 = QuestionNumerical(
   question_name = "minutes_reading",
   question_text = "{{ agent.nickname }}, on average, how many minutes do you spend consuming local news each day?",
   min_value = 0, # optional
   max_value = 1440 # optional
)

survey = Survey(questions = [q1, q2, q3, q4]).add_stop_rule(q1, "consume_local_news == 'Never'")

job = survey.by(agent)

job.show_flow()

Output:

Conditional expressions

The rule expressions themselves (“consume_local_news == ‘Never’”) are written in Python. An expression is evaluated to True or False, with the answer substituted into the expression. The placeholder for this answer is the name of the question itself. In the examples, the answer to q1 is substituted into the expression “consume_local_news == ‘Never’”, as the name of q1 is “consume_local_news”.

Piping

Piping is a method of explicitly referencing components of a question in a later question. For example, here we use the answer to q0 in the prompt for q1:

from edsl import QuestionFreeText, QuestionList, Survey, Agent

q0 = QuestionFreeText(
   question_name = "color",
   question_text = "What is your favorite color?",
)

q1 = QuestionList(
   question_name = "examples",
   question_text = "Name some things that are {{ color.answer }}.",
)

survey = Survey([q0, q1])

agent = Agent(traits = {"persona": "You are a botanist."})

results = survey.by(agent).run()

results.select("color", "examples")

In this example, q0 will be administered before q1 and the response to q0 is piped into q1. Output:

answer.color	answer.examples
As a botanist, I find myself drawn to the vibrant greens of nature. Green is a color that symbolizes growth, life, and the beauty of plants, which are central to my work and passion.	[‘Leaves’, ‘Grass’, ‘Ferns’, ‘Moss’, ‘Green algae’]

If an answer is a list, we can use the list as the question_options in another question, or index items individually. Here we demonstrate examples of both:

from edsl import QuestionList, QuestionFreeText, QuestionMultipleChoice, Survey, Agent

q_colors = QuestionList(
   question_name = "colors",
   question_text = "What are your 3 favorite colors?",
   max_list_items = 3
)

q_examples = QuestionFreeText(
   question_name = "examples",
   question_text = "Name some things that are {{ colors.answer }}",
)

q_favorite = QuestionMultipleChoice(
   question_name = "favorite",
   question_text = "Which is your #1 favorite color?",
   question_options = [
      "{{ colors.answer[0] }}",
      "{{ colors.answer[1] }}",
      "{{ colors.answer[2] }}",
   ]
)

survey = Survey([q_colors, q_examples, q_favorite])

agent = Agent(traits = {"persona": "You are a botanist."})

results = survey.by(agent).run()

results.select("colors", "examples", "favorite")

Output:

answer.colors	answer.examples	answer.favorite
[‘Green’, ‘Brown’, ‘Yellow’]	Certainly! Here are some things that can be green, brown, or yellow: Green: 1. Leaves - Many plants have green leaves due to chlorophyll, which is essential for photosynthesis. 2. Grass - Typically green, especially when healthy and well-watered. 3. Green Apples - Varieties like Granny Smith are known for their green color. Brown: 1. Tree Bark - The outer layer of trees is often brown, providing protection. 2. Soil - Many types of soil appear brown, indicating organic matter. 3. Acorns - These seeds from oak trees are generally brown when mature. Yellow: 1. Sunflowers - Known for their bright yellow petals. 2. Bananas - Yellow when ripe and ready to eat. 3. Daffodils - These flowers are often a vibrant yellow, heralding spring.	Green

This can also be done with agent traits. For example:

from edsl import Agent, QuestionFreeText

a = Agent(traits = {'first_name': 'John'})

q = QuestionFreeText(
   question_text = 'What is your last name, {{ agent.first_name }}?',
   question_name = "last_name"
)

job = q.by(a)

job.prompts().select('user_prompt')

This code will output the text of the prompt for the question:

What is your last name, John?

We can also show both system and user prompts together with information about the question, agent and model by calling the show_prompts() method:

job.show_prompts()

Output:

user_prompt	system_prompt	interview_index	question_name	scenario_index	agent_index	model	estimated_cost
What is your last name, John?	You are answering questions as if you were a human. Do not break character. You are an agent with the following persona: {‘first_name’: ‘John’}	0	last_name	0	0	gpt-4o	0.0005375

Question memory

When an agent is taking a survey, they can be prompted to “remember” answers to previous questions. This can be done in several ways:

Full memory

The method set_full_memory_mode() gives the agent all of the prior questions and answers at each new question in the survey, i.e., the first question and answer are included in the memory when answering the second question, both the first and second questions and answers are included in the memory when answering the third question, and so on. The method is called on the survey object:

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.set_full_memory_mode()

In the results, we can inspect the _user_prompt for each question to see that the agent was prompted to remember all of the prior questions:

results = survey.by(agent).run()

(
   results
   .select("consume_local_news_user_prompt", "sources_user_prompt", "rate_coverage_user_prompt", "minutes_reading_user_prompt")
)

This will print the prompt that was used for each question, and we can see that each successive prompt references all prior questions and answers that were given:

prompt.consume_local_news_user_prompt	prompt.sources_user_prompt	prompt.rate_coverage_user_prompt	prompt.minutes_reading_user_prompt
How often do you consume local news? Daily Weekly Monthly Never Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	What are your most common sources of local news? 0: Television 1: Newspaper 2: Online news websites 3: Social Media 4: Radio 5: Other Please respond only with a comma-separated list of the code of the options that apply, with square brackets. E.g., [0, 1, 3] After the answer, you can put a comment explaining your choice on the next line.	On a scale of 1 to 10, how would you rate the quality of local news coverage in your area? 1 : Very poor 10 : Excellent Only 1 option may be selected. Respond only with the code corresponding to one of the options. E.g., “1” or “10” by itself. After the answer, you can put a comment explaining why you chose that option on the next line.	On average, how many minutes do you spend consuming local news each day? Minimum answer value: 0 Maximum answer value: 1440 This question requires a numerical response in the form of an integer or decimal (e.g., -12, 0, 1, 2, 3.45, …). Respond with just your number on a single line. If your response is equivalent to zero, report ‘0’. After the answer, put a comment explaining your choice on the next line.

Note that this is slow and token-intensive, as the questions must be answered serially and requires the agent to remember all of the answers to the questions in the survey. In contrast, if the agent does not need to remember all of the answers to the questions in the survey, execution can proceed in parallel.

Lagged memory

The method set_lagged_memory() gives the agent a specified number of prior questions and answers at each new question in the survey; we pass it the number of prior questions and answers to remember. Here we use it to give the agent just 1 prior question/answer at each question:

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.set_lagged_memory(1)

We can inspect each _user_prompt again and see that the agent is only prompted to remember the last prior question/answer:

results = survey.by(agent).run()

(
   results
   .select("consume_local_news_user_prompt", "sources_user_prompt", "rate_coverage_user_prompt", "minutes_reading_user_prompt")
)

This will print the prompts for each question:

prompt.consume_local_news_user_prompt	prompt.sources_user_prompt	prompt.rate_coverage_user_prompt	prompt.minutes_reading_user_prompt
How often do you consume local news? Daily Weekly Monthly Never Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	What are your most common sources of local news? 0: Television 1: Newspaper 2: Online news websites 3: Social Media 4: Radio 5: Other Please respond only with a comma-separated list of the code of the options that apply, with square brackets (e.g., [0, 1, 3]). After the answer, you can put a comment explaining your choice on the next line.	On a scale of 1 to 10, how would you rate the quality of local news coverage in your area? 1 : Very poor 10 : Excellent Only 1 option may be selected. Respond only with the code corresponding to one of the options (e.g., “1” or “10”). After the answer, you can put a comment explaining why you chose that option on the next line.	On average, how many minutes do you spend consuming local news each day? Minimum answer value: 0 Maximum answer value: 1440 This question requires a numerical response in the form of an integer or decimal (e.g., -12, 0, 1, 2, 3.45, …). Respond with just your number on a single line. If your response is equivalent to zero, report ‘0’. After the answer, put a comment explaining your choice on the next line.

Targeted memory

The method add_targeted_memory() gives the agent a targeted prior question and answer when answering another specified question. We pass it the question to answer and the prior question/answer to remember when answering it. Here we use it to give the agent the question/answer to q1 when prompting it to answer q4:

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_targeted_memory(q4, q1)

results = survey.by(agent).run()

(
   results
   .select("consume_local_news_user_prompt", "sources_user_prompt", "rate_coverage_user_prompt", "minutes_reading_user_prompt")
)

Output:

prompt.consume_local_news_user_prompt	prompt.sources_user_prompt	prompt.rate_coverage_user_prompt	prompt.minutes_reading_user_prompt
How often do you consume local news? Daily Weekly Monthly Never Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	What are your most common sources of local news? 0: Television 1: Newspaper 2: Online news websites 3: Social Media 4: Radio 5: Other Please respond only with a comma-separated list of the code of the options that apply, with square brackets (e.g., [0, 1, 3]). After the answer, you can put a comment explaining your choice on the next line.	On a scale of 1 to 10, how would you rate the quality of local news coverage in your area? 1 : Very poor 10 : Excellent Only 1 option may be selected. Respond only with the code corresponding to one of the options (e.g., “1” or “10”). After the answer, you can put a comment explaining why you chose that option on the next line.	On average, how many minutes do you spend consuming local news each day? Minimum answer value: 0 Maximum answer value: 1440 This question requires a numerical response in the form of an integer or decimal (e.g., -12, 0, 1, 2, 3.45, …). Respond with just your number on a single line. If your response is equivalent to zero, report ‘0’. After the answer, put a comment explaining your choice on the next line.

Memory collection

The add_memory_collection() method is used to add sets of prior questions and answers to a given question. We pass it the question to be answered and the list of questions/answers to be remembered when answering it. For example, we can add the questions/answers for both q1 and q2 when prompting the agent to answer q4:

survey = Survey(questions = [q1, q2, q3, q4])
survey = survey.add_memory_collection(q4, [q1, q2])

results = survey.by(agent).run()

(
   results
   .select("consume_local_news_user_prompt", "sources_user_prompt", "rate_coverage_user_prompt", "minutes_reading_user_prompt")
)

Output:

prompt.consume_local_news_user_prompt	prompt.sources_user_prompt	prompt.rate_coverage_user_prompt	prompt.minutes_reading_user_prompt
How often do you consume local news? Daily Weekly Monthly Never Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	What are your most common sources of local news? 0: Television 1: Newspaper 2: Online news websites 3: Social Media 4: Radio 5: Other Please respond only with a comma-separated list of the code of the options that apply, with square brackets (e.g., [0, 1, 3]). After the answer, you can put a comment explaining your choice on the next line.	On a scale of 1 to 10, how would you rate the quality of local news coverage in your area? 1 : Very poor 10 : Excellent Only 1 option may be selected. Respond only with the code corresponding to one of the options (e.g., “1” or “10”). After the answer, you can put a comment explaining why you chose that option on the next line.	On average, how many minutes do you spend consuming local news each day? Minimum answer value: 0 Maximum answer value: 1440 This question requires a numerical response in the form of an integer or decimal (e.g., -12, 0, 1, 2, 3.45, …). Respond with just your number on a single line. If your response is equivalent to zero, report ‘0’. After the answer, put a comment explaining your choice on the next line.

Costs

Before running a survey, you can estimate the cost of running the survey in USD and the number of credits needed to run it remotely at the Expected Parrot server. After running a survey, you can see details on the actual cost of each response in the results. The costs are calculated based on the estimated and actual number of tokens used in the survey and the model(s) used to generate the prompts.

Estimated costs

Before running a survey, you can estimate the cost in USD of running the survey by calling the estimate_job_cost() method on a Job object (a survey combined with one or more models). This method returns a dictionary with the estimated costs and tokens for each model used with the survey. You can also estimate credits needed to run a survey remotely at the Expected Parrot server by passing the job to the remote_inference_cost() method of a Coop client object.

Example:

from edsl import QuestionFreeText, Survey, Agent, Model

q0 = QuestionFreeText(
    question_name = "favorite_flower",
    question_text = "What is the name of your favorite flower?"
)
q1 = QuestionFreeText(
    question_name = "flower_color",
    question_text = "What color is {{ favorite_flower.answer }}?"
)

survey = Survey(questions = [q0, q1])

a = Agent(traits = {"persona":"You are a botanist on Cape Cod."})

m = Model("gpt-4o")

job = survey.by(a).by(m)

estimated_job_cost = job.estimate_job_cost()
estimated_job_cost

Output:

{'estimated_total_cost_usd': 0.0008625000000000001,
'estimated_total_input_tokens': 85,
'estimated_total_output_tokens': 65,
'model_costs': [{'inference_service': 'openai',
   'model': 'gpt-4o',
   'estimated_cost_usd': 0.0008625000000000001,
   'estimated_input_tokens': 85,
   'estimated_output_tokens': 65}]}

To get the estimated cost in credits to run the job remotely:

from edsl import Coop

coop = Coop()

estimated_remote_inference_cost = coop.remote_inference_cost(job) # using the job object from above
estimated_remote_inference_cost

Output:

{'credits': 0.09, 'usd': 0.0008650000000000001}

Details of the calculations for these methods can be found in the credits section.

Actual costs

The actual costs of running a survey are stored in the survey results. Details about the cost of each response can be accessed in the raw_model_response fields of the results dataset. For each question that was run, the following columns will appear in results:

• raw_model_response.<question_name>_cost: The cost in USD for the API call to a language model service provider.

• raw_model_response.<question_name>_one_usd_buys: The number of tokens that can be purchased with 1 USD (for reference).

• raw_model_response.<question_name>_raw_model_response: A dictionary containing the raw response for the question, which includes the input text and tokens, output text and tokens, and other information about the API call. This dictionary is specific to the language model service provider and may contain additional information about the response.

The cost in credits of a response is calculated as follows:

• The number of input tokens is multiplied by the input token rate set by the language model service provider.

• The number of output tokens is multiplied by the output token rate set by the language model service provider.

• The total cost in USD is converted to credits (1 USD = 100 credits).

• The total cost in credits is rounded up to the nearest 1/100th of a credit.

To learn more about these methods and calculations, please see the credits section.

Survey class

A Survey is collection of questions that can be administered to an Agent.

class edsl.surveys.Survey.PseudoIndices(dict=None, /, **kwargs)

Bases: UserDict

property last_item_was_instruction: bool

Return whether the last item added to the survey was an instruction.

This is used to determine the pseudo-index of the next item added to the survey.

Example:

>>> s = Survey.example()
>>> s._pseudo_indices.last_item_was_instruction
False
>>> from edsl.surveys.instructions.Instruction import Instruction
>>> s = s.add_instruction(Instruction(text="Pay attention to the following questions.", name="intro"))
>>> s._pseudo_indices.last_item_was_instruction
True

property max_pseudo_index: float

Return the maximum pseudo index in the survey. >>> Survey.example()._pseudo_indices.max_pseudo_index 2

class edsl.surveys.Survey.Survey(questions: List['QuestionType'] | None = None, memory_plan: 'MemoryPlan' | None = None, rule_collection: 'RuleCollection' | None = None, question_groups: 'QuestionGroupType' | None = None, name: str | None = None, questions_to_randomize: List[str] | None = None)

Bases: SurveyExportMixin, Base

A collection of questions that supports skip logic.

__init__(questions: List['QuestionType'] | None = None, memory_plan: 'MemoryPlan' | None = None, rule_collection: 'RuleCollection' | None = None, question_groups: 'QuestionGroupType' | None = None, name: str | None = None, questions_to_randomize: List[str] | None = None)

Create a new survey.

Parameters:

• questions – The questions in the survey.

• memory_plan – The memory plan for the survey.

• rule_collection – The rule collection for the survey.

• question_groups – The groups of questions in the survey.

• name – The name of the survey - DEPRECATED.

>>> from edsl import QuestionFreeText
>>> q1 = QuestionFreeText(question_text = "What is your name?", question_name = "name")
>>> q2 = QuestionFreeText(question_text = "What is your favorite color?", question_name = "color")
>>> q3 = QuestionFreeText(question_text = "Is a hot dog a sandwich", question_name = "food")
>>> s = Survey([q1, q2, q3], question_groups = {"demographics": (0, 1), "substantive":(3)})

add_instruction(instruction: Instruction | ChangeInstruction) → Survey

Add an instruction to the survey.

Parameters:

instruction – The instruction to add to the survey.

>>> from edsl import Instruction
>>> i = Instruction(text="Pay attention to the following questions.", name="intro")
>>> s = Survey().add_instruction(i)
>>> s._instruction_names_to_instructions
{'intro': Instruction(name="intro", text="Pay attention to the following questions.")}
>>> s._pseudo_indices
{'intro': -0.5}

add_memory_collection(focal_question: QuestionBase | str, prior_questions: List[QuestionBase | str]) → Survey

Add prior questions and responses so the agent has them when answering.

This adds instructions to a survey than when answering focal_question, the agent should also remember the answers to prior_questions listed in prior_questions.

Parameters:

• focal_question – The question that the agent is answering.

• prior_questions – The questions that the agent should remember when answering the focal question.

Here we have it so that when answering q2, the agent should remember answers to q0 and q1:

>>> s = Survey.example().add_memory_collection("q2", ["q0", "q1"])
>>> s.memory_plan
{'q2': Memory(prior_questions=['q0', 'q1'])}

add_question(question: QuestionBase, index: int | None = None) → Survey

Add a question to survey.

Parameters:

• question – The question to add to the survey.

• question_name – The name of the question. If not provided, the question name is used.

The question is appended at the end of the self.questions list A default rule is created that the next index is the next question.

>>> from edsl import QuestionMultipleChoice
>>> q = QuestionMultipleChoice(question_text = "Do you like school?", question_options=["yes", "no"], question_name="q0")
>>> s = Survey().add_question(q)

>>> s = Survey().add_question(q).add_question(q)
Traceback (most recent call last):
...
edsl.exceptions.surveys.SurveyCreationError: Question name 'q0' already exists in survey. Existing names are ['q0'].
...

add_question_group(start_question: QuestionBase | str, end_question: QuestionBase | str, group_name: str) → Survey

Add a group of questions to the survey.

Parameters:

• start_question – The first question in the group.

• end_question – The last question in the group.

• group_name – The name of the group.

Example:

>>> s = Survey.example().add_question_group("q0", "q1", "group1")
>>> s.question_groups
{'group1': (0, 1)}

The name of the group must be a valid identifier:

>>> s = Survey.example().add_question_group("q0", "q2", "1group1")
Traceback (most recent call last):
...
edsl.exceptions.surveys.SurveyCreationError: Group name 1group1 is not a valid identifier.
...
>>> s = Survey.example().add_question_group("q0", "q1", "q0")
Traceback (most recent call last):
...
edsl.exceptions.surveys.SurveyCreationError: ...
...
>>> s = Survey.example().add_question_group("q1", "q0", "group1")
Traceback (most recent call last):
...
edsl.exceptions.surveys.SurveyCreationError: ...
...

add_rule(question: QuestionBase | str, expression: str, next_question: QuestionBase | int, before_rule: bool = False) → Survey

Add a rule to a Question of the Survey.

Parameters:

• question – The question to add the rule to.

• expression – The expression to evaluate.

• next_question – The next question to go to if the rule is true.

• before_rule – Whether the rule is evaluated before the question is answered.

This adds a rule that if the answer to q0 is ‘yes’, the next question is q2 (as opposed to q1)

>>> s = Survey.example().add_rule("q0", "{{ q0 }} == 'yes'", "q2")
>>> s.next_question("q0", {"q0": "yes"}).question_name
'q2'

add_skip_rule(question: QuestionBase | str, expression: str) → Survey

Adds a per-question skip rule to the survey.

Parameters:

• question – The question to add the skip rule to.

• expression – The expression to evaluate.

This adds a rule that skips ‘q0’ always, before the question is answered:

>>> from edsl import QuestionFreeText
>>> q0 = QuestionFreeText.example()
>>> q0.question_name = "q0"
>>> q1 = QuestionFreeText.example()
>>> q1.question_name = "q1"
>>> s = Survey([q0, q1]).add_skip_rule("q0", "True")
>>> s.next_question("q0", {}).question_name
'q1'

Note that this is different from a rule that jumps to some other question after the question is answered.

add_stop_rule(question: QuestionBase | str, expression: str) → Survey

Add a rule that stops the survey. The rule is evaluated after the question is answered. If the rule is true, the survey ends.

Parameters:

• question – The question to add the stop rule to.

• expression – The expression to evaluate.

If this rule is true, the survey ends.

Here, answering “yes” to q0 ends the survey:

>>> s = Survey.example().add_stop_rule("q0", "q0 == 'yes'")
>>> s.next_question("q0", {"q0": "yes"})
EndOfSurvey

By comparison, answering “no” to q0 does not end the survey:

>>> s.next_question("q0", {"q0": "no"}).question_name
'q1'

>>> s.add_stop_rule("q0", "q1 <> 'yes'")
Traceback (most recent call last):
...
edsl.exceptions.surveys.SurveyCreationError: The expression contains '<>', which is not allowed. You probably mean '!='.
...

add_targeted_memory(focal_question: QuestionBase | str, prior_question: QuestionBase | str) → Survey

Add instructions to a survey than when answering focal_question.

Parameters:

• focal_question – The question that the agent is answering.

• prior_question – The question that the agent should remember when answering the focal question.

Here we add instructions to a survey than when answering q2 they should remember q1:

>>> s = Survey.example().add_targeted_memory("q2", "q0")
>>> s.memory_plan
{'q2': Memory(prior_questions=['q0'])}

The agent should also remember the answers to prior_questions listed in prior_questions.

by(*args: 'Agent' | 'Scenario' | 'LanguageModel') → Jobs

Add Agents, Scenarios, and LanguageModels to a survey and returns a runnable Jobs object.

Parameters:

args – The Agents, Scenarios, and LanguageModels to add to the survey.

This takes the survey and adds an Agent and a Scenario via ‘by’ which converts to a Jobs object:

>>> s = Survey.example(); from edsl.agents import Agent; from edsl import Scenario
>>> s.by(Agent.example()).by(Scenario.example())
Jobs(...)

clear_non_default_rules() → Survey

Remove all non-default rules from the survey.

>>> Survey.example().show_rules()
Dataset([{'current_q': [0, 0, 1, 2]}, {'expression': ['True', "q0 == 'yes'", 'True', 'True']}, {'next_q': [1, 2, 2, 3]}, {'priority': [-1, 0, -1, -1]}, {'before_rule': [False, False, False, False]}])
>>> Survey.example().clear_non_default_rules().show_rules()
Dataset([{'current_q': [0, 1, 2]}, {'expression': ['True', 'True', 'True']}, {'next_q': [1, 2, 3]}, {'priority': [-1, -1, -1]}, {'before_rule': [False, False, False]}])

codebook() → dict[str, str]

Create a codebook for the survey, mapping question names to question text.

>>> s = Survey.example()
>>> s.codebook()
{'q0': 'Do you like school?', 'q1': 'Why not?', 'q2': 'Why?'}

dag(textify: bool = False) → DAG

Return the DAG of the survey, which reflects both skip-logic and memory.

Parameters:

textify – Whether to return the DAG with question names instead of indices.

>>> s = Survey.example()
>>> d = s.dag()
>>> d
{1: {0}, 2: {0}}

delete_question(identifier: str | int) → Survey

Delete a question from the survey.

Parameters:

identifier – The name or index of the question to delete.

Returns:

The updated Survey object.

>>> from edsl import QuestionMultipleChoice, Survey
>>> q1 = QuestionMultipleChoice(question_text="Q1", question_options=["A", "B"], question_name="q1")
>>> q2 = QuestionMultipleChoice(question_text="Q2", question_options=["C", "D"], question_name="q2")
>>> s = Survey().add_question(q1).add_question(q2)
>>> _ = s.delete_question("q1")
>>> len(s.questions)
1
>>> _ = s.delete_question(0)
>>> len(s.questions)
0

draw() → Survey

Return a new survey with a randomly selected permutation of the options.

duplicate()

Duplicate the survey.

>>> s = Survey.example()
>>> s2 = s.duplicate()
>>> s == s2
True
>>> s is s2
False

classmethod example(params: bool = False, randomize: bool = False, include_instructions=False, custom_instructions: str | None = None) → Survey

Return an example survey.

>>> s = Survey.example()
>>> [q.question_text for q in s.questions]
['Do you like school?', 'Why not?', 'Why?']

classmethod from_dict(data: dict) → Survey

Deserialize the dictionary back to a Survey object.

Parameters:

data – The dictionary to deserialize.

>>> d = Survey.example().to_dict()
>>> s = Survey.from_dict(d)
>>> s == Survey.example()
True

>>> s = Survey.example(include_instructions = True)
>>> d = s.to_dict()
>>> news = Survey.from_dict(d)
>>> news == s
True

gen_path_through_survey() → Generator[QuestionBase, dict, None]

Generate a coroutine that can be used to conduct an Interview.

The coroutine is a generator that yields a question and receives answers. It starts with the first question in the survey. The coroutine ends when an EndOfSurvey object is returned.

For the example survey, this is the rule table:

>>> s = Survey.example()
>>> s.show_rules()
Dataset([{'current_q': [0, 0, 1, 2]}, {'expression': ['True', "q0 == 'yes'", 'True', 'True']}, {'next_q': [1, 2, 2, 3]}, {'priority': [-1, 0, -1, -1]}, {'before_rule': [False, False, False, False]}])

Note that q0 has a rule that if the answer is ‘yes’, the next question is q2. If the answer is ‘no’, the next question is q1.

Here is the path through the survey if the answer to q0 is ‘yes’:

>>> i = s.gen_path_through_survey()
>>> next(i)
Question('multiple_choice', question_name = """q0""", question_text = """Do you like school?""", question_options = ['yes', 'no'])
>>> i.send({"q0": "yes"})
Question('multiple_choice', question_name = """q2""", question_text = """Why?""", question_options = ['**lack*** of killer bees in cafeteria', 'other'])

And here is the path through the survey if the answer to q0 is ‘no’:

>>> i2 = s.gen_path_through_survey()
>>> next(i2)
Question('multiple_choice', question_name = """q0""", question_text = """Do you like school?""", question_options = ['yes', 'no'])
>>> i2.send({"q0": "no"})
Question('multiple_choice', question_name = """q1""", question_text = """Why not?""", question_options = ['killer bees in cafeteria', 'other'])

get_job(model=None, agent=None, **kwargs)

humanize(project_name: str = 'Project', survey_description: str | None = None, survey_alias: str | None = None, survey_visibility: 'VisibilityType' | None = 'unlisted') → dict

Send the survey to Coop.

Then, create a project on Coop so you can share the survey with human respondents.

move_question(identifier: str | int, new_index: int) → Survey

>>> from edsl import QuestionMultipleChoice, Survey
>>> s = Survey.example()
>>> s.question_names
['q0', 'q1', 'q2']
>>> s.move_question("q0", 2).question_names
['q1', 'q2', 'q0']

next_question(current_question: str | QuestionBase | None = None, answers: dict | None = None) → QuestionBase | EndOfSurvey.__class__

Return the next question in a survey.

Parameters:

• current_question – The current question in the survey.

• answers – The answers for the survey so far

• If called with no arguments, it returns the first question in the survey.

• If no answers are provided for a question with a rule, the next question is returned. If answers are provided, the next question is determined by the rules and the answers.

• If the next question is the last question in the survey, an EndOfSurvey object is returned.

>>> s = Survey.example()
>>> s.next_question("q0", {"q0": "yes"}).question_name
'q2'
>>> s.next_question("q0", {"q0": "no"}).question_name
'q1'

property parameters

Return a set of parameters in the survey.

>>> s = Survey.example()
>>> s.parameters
set()

property parameters_by_question

Return a dictionary of parameters by question in the survey. >>> from edsl import QuestionFreeText >>> q = QuestionFreeText(question_name = “example”, question_text = “What is the capital of {{ country}}?”) >>> s = Survey([q]) >>> s.parameters_by_question {‘example’: {‘country’}}

property question_name_to_index: dict[str, int]

Return a dictionary mapping question names to question indices.

Example:

>>> s = Survey.example()
>>> s.question_name_to_index
{'q0': 0, 'q1': 1, 'q2': 2}

property question_names: list[str]

Return a list of question names in the survey.

Example:

>>> s = Survey.example()
>>> s.question_names
['q0', 'q1', 'q2']

question_names_to_questions() → dict

Return a dictionary mapping question names to question attributes.

questions

A collection of questions that supports skip logic.

Initalization: - questions: the questions in the survey (optional) - question_names: the names of the questions (optional) - name: the name of the survey (optional)

Methods: -

Notes: - The presumed order of the survey is the order in which questions are added.

classmethod random_survey()

run(*args, **kwargs) → Results

Turn the survey into a Job and runs it.

>>> from edsl import QuestionFreeText
>>> s = Survey([QuestionFreeText.example()])
>>> from edsl.language_models import LanguageModel
>>> m = LanguageModel.example(test_model = True, canned_response = "Great!")
>>> results = s.by(m).run(cache = False, disable_remote_cache = True, disable_remote_inference = True)
>>> results.select('answer.*')
Dataset([{'answer.how_are_you': ['Great!']}])

async run_async(model: 'LanguageModel' | None = None, agent: 'Agent' | None = None, cache: 'Cache' | None = None, disable_remote_inference: bool = False, disable_remote_cache: bool = False, **kwargs)

Run the survey with default model, taking the required survey as arguments.

>>> import asyncio
>>> from edsl.questions import QuestionFunctional
>>> def f(scenario, agent_traits): return "yes" if scenario["period"] == "morning" else "no"
>>> q = QuestionFunctional(question_name = "q0", func = f)
>>> s = Survey([q])
>>> async def test_run_async(): result = await s.run_async(period="morning", disable_remote_inference = True, disable_remote_cache=True); print(result.select("answer.q0").first())
>>> asyncio.run(test_run_async())
yes
>>> import asyncio
>>> from edsl.questions import QuestionFunctional
>>> def f(scenario, agent_traits): return "yes" if scenario["period"] == "morning" else "no"
>>> q = QuestionFunctional(question_name = "q0", func = f)
>>> s = Survey([q])
>>> async def test_run_async(): result = await s.run_async(period="evening", disable_remote_inference = True, disable_remote_cache = True); print(result.select("answer.q0").first())
>>> results = asyncio.run(test_run_async())
no

property scenario_attributes: list[str]

Return a list of attributes that admissible Scenarios should have.

Here we have a survey with a question that uses a jinja2 style {{ }} template:

>>> from edsl import QuestionFreeText
>>> s = Survey().add_question(QuestionFreeText(question_text="{{ greeting }}. What is your name?", question_name="name"))
>>> s.scenario_attributes
['greeting']

>>> s = Survey().add_question(QuestionFreeText(question_text="{{ greeting }}. What is your {{ attribute }}?", question_name="name"))
>>> s.scenario_attributes
['greeting', 'attribute']

set_full_memory_mode() → Survey

Add instructions to a survey that the agent should remember all of the answers to the questions in the survey.

>>> s = Survey.example().set_full_memory_mode()

set_lagged_memory(lags: int) → Survey

Add instructions to a survey that the agent should remember the answers to the questions in the survey.

The agent should remember the answers to the questions in the survey from the previous lags.

show_flow(filename: str | None = None) → None

Show the flow of the survey.

show_prompts()

Show the prompts for the survey.

show_rules() → None

Print out the rules in the survey.

>>> s = Survey.example()
>>> s.show_rules()
Dataset([{'current_q': [0, 0, 1, 2]}, {'expression': ['True', "q0 == 'yes'", 'True', 'True']}, {'next_q': [1, 2, 2, 3]}, {'priority': [-1, 0, -1, -1]}, {'before_rule': [False, False, False, False]}])

simulate() → dict

Simulate the survey and return the answers.

table(*fields, tablefmt=None) → Table

to_dict(add_edsl_version=True) → dict[str, Any]

Serialize the Survey object to a dictionary.

>>> s = Survey.example()
>>> s.to_dict(add_edsl_version = False).keys()
dict_keys(['questions', 'memory_plan', 'rule_collection', 'question_groups'])

to_jobs()

Convert the survey to a Jobs object. >>> s = Survey.example() >>> s.to_jobs() Jobs(…)

tree(node_list: List[str] | None = None)

using(obj: 'Cache' | 'KeyLookup' | 'BucketCollection') → Jobs

Turn the survey into a Job and appends the arguments to the Job.

edsl.surveys.Survey.main()

Run the example survey.