Install this package with pip install forecasting-tools
Demo website: https://mokoresearch.streamlit.app/ Demo repo: https://github.com/Metaculus/metac-bot-template
This repository contains forecasting and research tools built with Python and Streamlit. The project aims to assist users in making predictions, conducting research, and analyzing data related to hard to answer questions (especially those from Metaculus).
Here are the tools most likely to be useful to you:
- 🎯 Forecasting Bot: General forecaster that integrates with the Metaculus AI benchmarking competition and provides a number of utilities. You can forecast with a pre-existing bot or override the class to customize your own (without redoing all the API code, etc)
- 🔌 Metaculus API Wrapper: for interacting with questions and tournaments
- đź“Š Benchmarking: Randomly sample quality questions from Metaculus and run you bot against them so you can get an early sense of how your bot is doing by comparing to the community prediction and expected log scores.
- 🔍 Smart Searcher: A custom AI-powered internet-informed llm powered by Exa.ai and GPT. Its a better (but slightly more expensive) alternative to Perplexity.ai that is configurable, more accurate, able to decide on filters, able to link to exact paragraphs, etc.
- 🔑 Key Factor Analysis: Key Factors Analysis for scoring, ranking, and prioritizing important variables in forecasting questions
Here are some other features of the project:
- Base Rate Researcher: for calculating event probabilities (still experimental)
- Niche List Researcher: for analyzing very specific lists of past events or items (still experimental)
- Fermi Estimator: for breaking down numerical estimates (still experimental)
- Monetary Cost Manager: for tracking AI and API expenses
All the examples below are in a Jupyter Notebook called README.ipynb which you can run locally to test the package (make sure to run the first cell though).
If you decide you want to join the Metaculus AI Benchmarking Tournament, it is recommended that your start here. This repo is very easy to set up and allows you to get going in 30min.
Join the discord for updates and to give feedback (btw feedback is very appreciated, even just a quick "I did/didn't decide to use tool X for reason Y, though am busy and don't have time to elaborate" is helpful to know)
Note: This package is still in an experimental phase. The goal is to keep the package API fairly stable, though no guarantees are given at this phase. There will be special effort to keep the ForecastBot and TemplateBot APIs consistent.
The package comes with two major pre-built bots:
- MainBot: The more sophisticated and expensive bot that does deeper research.
- TemplateBot: A simpler bot that models the Metaculus templates that's cheaper, easier to start with, and faster to run.
They both have roughly the same parameters. See below on how to use the TemplateBot to make forecasts.
from forecasting_tools import TemplateBot, MetaculusApi
# Initialize the bot
bot = TemplateBot(
research_reports_per_question=3, # Number of separate research attempts per question
predictions_per_research_report=5, # Number of predictions to make per research report
publish_reports_to_metaculus=False, # Whether to post the forecasts to Metaculus
folder_to_save_reports_to="logs/forecasts/", # Where to save detailed reports (file saving environment variable must be set)
skip_previously_forecasted_questions=False
)
TOURNAMENT_ID = MetaculusApi.CURRENT_QUARTERLY_CUP_ID
reports = await bot.forecast_on_tournament(TOURNAMENT_ID)
# Print results (if the tournament is not active, no reports will be returned)
for report in reports:
print(f"\nQuestion: {report.question.question_text}")
print(f"Prediction: {report.prediction}")from forecasting_tools import (
TemplateBot,
BinaryQuestion,
MetaculusApi,
DataOrganizer
)
# Initialize the bot
bot = TemplateBot(
research_reports_per_question=3,
predictions_per_research_report=5,
publish_reports_to_metaculus=False,
)
# Get and forecast a specific question
question1 = MetaculusApi.get_question_by_url(
"https://www.metaculus.com/questions/578/human-extinction-by-2100/"
)
question2 = BinaryQuestion(
question_text="Will YouTube be blocked in Russia?",
background_info="...", # Or 'None'
resolution_criteria="...", # Or 'None'
fine_print="...", # Or 'None'
id_of_post=0, # The ID and state only matters if using Metaculus API calls so feel free to set it to 0
)
reports = await bot.forecast_questions([question1, question2])
# Print results
for report in reports:
print(f"Question: {report.question.question_text}")
print(f"Prediction: {report.prediction}")
shortened_explanation = report.explanation.replace('\n', ' ')[:100]
print(f"Reasoning: {shortened_explanation}...")
# You can also save and load questions and reports
file_path = "temp/reports.json"
DataOrganizer.save_reports_to_file_path(reports, file_path)
loaded_reports = DataOrganizer.load_reports_from_file_path(file_path)Question: Will humans go extinct before 2100?
Prediction: 0.03
Reasoning: # SUMMARY *Question*: Will humans go extinct before 2100? *Final Prediction*: 3.0% *Total Cost*: $0...
Question: Will YouTube be blocked in Russia?
Prediction: 0.7
Reasoning: # SUMMARY *Question*: Will YouTube be blocked in Russia? *Final Prediction*: 70.0% *Total Cost*: $0...
The bot will:
- Research the question
- Generate multiple independent predictions
- Combine these predictions into a final forecast
- Save detailed research and reasoning to the specified folder
- Optionally post the forecast to Metaculus (if
publish_reports_to_metaculus=True)
Note: You'll need to have your environment variables set up (see the section below)
Generally all you have to do to make your own bot is inherit from the TemplateBot and override any combination of the 3 forecasting methods and the 1 research method. This saves you the headache of interacting with the Metaculus API, implementing aggregation of predictions, creating benchmarking interfaces, etc. Below is an example. It may also be helpful to look at the TemplateBot code (forecasting_tools/forecasting/forecast_bots/template_bot.py) for a more complete example.
from forecasting_tools import (
TemplateBot,
MetaculusQuestion,
BinaryQuestion,
MultipleChoiceQuestion,
NumericQuestion,
ReasonedPrediction,
PredictedOptionList,
NumericDistribution,
SmartSearcher,
MetaculusApi,
GeneralLlm,
PredictionExtractor
)
from forecasting_tools.ai_models.ai_utils.ai_misc import clean_indents
class MyCustomBot(TemplateBot):
async def run_research(self, question: MetaculusQuestion) -> str:
searcher = SmartSearcher(
num_searches_to_run=2,
num_sites_per_search=10
)
prompt = clean_indents(
f"""
Analyze this forecasting question:
1. Filter for recent events in the past 6 months
2. Don't include domains from youtube.com
3. Look for current trends and data
4. Find historical analogies and base rates
Question: {question.question_text}
Background Info: {question.background_info}
Resolution Criteria: {question.resolution_criteria}
"""
)
report = await searcher.invoke(prompt)
return report
async def _run_forecast_on_binary(
self, question: BinaryQuestion, research: str
) -> ReasonedPrediction[float]:
prompt = f"Please make a prediction on the following question: {question.question_text}. The last thing you write is your final answer as: 'Probability: ZZ%', 0-100"
reasoning = await GeneralLlm(model="metaculus/gpt-4o", temperature=0).invoke(prompt)
prediction = PredictionExtractor.extract_last_percentage_value(
reasoning, max_prediction=1, min_prediction=0
)
return ReasonedPrediction(
prediction_value=prediction, reasoning=reasoning
)
async def _run_forecast_on_multiple_choice(
self, question: MultipleChoiceQuestion, research: str
) -> ReasonedPrediction[PredictedOptionList]:
...
async def _run_forecast_on_numeric(
self, question: NumericQuestion, research: str
) -> ReasonedPrediction[NumericDistribution]:
...
custom_bot = MyCustomBot()
question = MetaculusApi.get_question_by_url(
"https://www.metaculus.com/questions/578/human-extinction-by-2100/"
)
report = await custom_bot.forecast_question(question)
print(f"Question: {report.question.question_text}")
print(f"Prediction: {report.prediction}")
print(f"Research: {report.research[:100]}...")Question: Will humans go extinct before 2100?
Prediction: 0.05
Research: # RESEARCH
To analyze the question of whether humans will go extinct before 2100, we need to consider...
You can create your own custom bot through the package in your own repo. An example can be found at this repo which you can fork and edit.
Whether running locally or through Github actions, you will need to set environment variables. All environment variables you might want are in .env.template. Generally you only need the METACULUS_TOKEN if running the Template. Having an EXA_API_KEY (see www.exa.ai) or PERPLEXITY_API_KEY (see www.perplexity.ai) is needed for searching the web. Make sure to put these variables in your .env file if running locally and in the Github actions secrets if running on Github actions.
Below is an example of how to run the benchmarker
from forecasting_tools import Benchmarker, TemplateBot, BenchmarkForBot
class CustomBot(TemplateBot):
...
# Run benchmark on multiple bots
bots = [TemplateBot(), CustomBot()] # Add your custom bots here
benchmarker = Benchmarker(
forecast_bots=bots,
number_of_questions_to_use=2, # Recommended 100+ for meaningful results
file_path_to_save_reports="benchmarks/"
)
benchmarks: list[BenchmarkForBot] = await benchmarker.run_benchmark()
# View results
for benchmark in benchmarks[:2]:
print("--------------------------------")
print(f"Bot: {benchmark.name}")
print(f"Score: {benchmark.average_inverse_expected_log_score}") # Lower is better
print(f"Num reports in benchmark: {len(benchmark.forecast_reports)}")
print(f"Time: {benchmark.time_taken_in_minutes}min")
print(f"Cost: ${benchmark.total_cost}")--------------------------------
Bot: Benchmark for TemplateBot
Score: 0.771654061941104
Num reports in benchmark: 2
Time: 0.5136146903038025min
Cost: $0.0332325
--------------------------------
Bot: Benchmark for CustomBot
Score: 0.7516094970590276
Num reports in benchmark: 2
Time: 0.4508770227432251min
Cost: $0.03279
The ideal number of questions to get a good sense of whether one bot is better than another can vary. 100+ should tell your something decent. See this analysis for exploration of the numbers. With two few questions, the results could just be statistical noise.
If you use the average inverse expected log score, then a lower score is better. The scoring measures the expected value of your score without needing an actual resolution by assuming that the community prediction is the 'true probability'. This correlates very closely with the baseline score on Metaculus (see analysis in scripts/simulate_a_tournament.ipynb)
As of Feb 17, 2024 the benchmarker automatically selects a random set of questions from Metaculus that:
- Are binary questions (yes/no)
- Are currently open
- Will resolve within the next year
- Have at least 40 forecasters
- Have a community prediction
- Are not part of a group question
As of last edit there are plans to expand this to numeric and multiple choice, but right now it just benchmarks binary questions.
You can grab these questions without using the Benchmarker by running the below
from forecasting_tools import MetaculusApi
questions = MetaculusApi.get_benchmark_questions(
num_of_questions_to_return=100,
)You can also save/load benchmarks to/from json
from forecasting_tools import BenchmarkForBot
# Load
file_path = "benchmarks/benchmark.json"
benchmarks: list[BenchmarkForBot] = BenchmarkForBot.load_json_from_file_path(file_path)
# Save
new_benchmarks: list[BenchmarkForBot] = benchmarks
BenchmarkForBot.save_object_list_to_file_path(new_benchmarks, file_path)
# To/From Json String
single_benchmark = benchmarks[0]
json_object: dict = single_benchmark.to_json()
new_benchmark: BenchmarkForBot = BenchmarkForBot.from_json(json_object)Once you have benchmark files in your directory you can run streamlit run scripts/benchmark_displayer.py to get a UI with the benchmarks. This will allow you to see metrics, code of the bots, the actual bot responses, etc. It will pull in any files in your directory that have the word "bench" in it and are json.
The Metaculus API wrapper helps interact with Metaculus questions and tournaments. Grabbing questions returns a pydantic object, and supports important information for Binary, Multiple Choice, Numeric,and Date questions.
from forecasting_tools import MetaculusApi, ApiFilter, DataOrganizer
from datetime import datetime
# Get a question by post id
question = MetaculusApi.get_question_by_post_id(578)
print(f"Question found with url: {question.page_url}")
# Get a question by url
question = MetaculusApi.get_question_by_url("https://www.metaculus.com/questions/578/human-extinction-by-2100/")
print(f"Question found with url: {question.page_url}")
# Get all open questions from a tournament
questions = MetaculusApi.get_all_open_questions_from_tournament(
tournament_id=MetaculusApi.CURRENT_QUARTERLY_CUP_ID
)
print(f"Num tournament questions: {len(questions)}")
# Get questions matching a filter
api_filter = ApiFilter(
num_forecasters_gte=40,
close_time_gt=datetime(2023, 12, 31),
close_time_lt=datetime(2024, 12, 31),
scheduled_resolve_time_lt=datetime(2024, 12, 31),
allowed_types=["binary", "multiple_choice", "numeric", "date"],
allowed_statuses=["resolved"],
)
questions = await MetaculusApi.get_questions_matching_filter(
api_filter=api_filter,
num_questions=50, # Remove this field to make it not error if you don't get 50 questions. However it will only go through one page of questions which may miss questions matching the ApiFilter since some filters are handled locally.
randomly_sample=False
)
print(f"Num filtered questions: {len(questions)}")
# Load and save questions/reports
file_path = "temp/questions.json"
DataOrganizer.save_questions_to_file_path(questions, file_path)
questions = DataOrganizer.load_questions_from_file_path(file_path)
# Get benchmark questions
benchmark_questions = MetaculusApi.get_benchmark_questions(
num_of_questions_to_return=20
)
print(f"Num benchmark questions: {len(benchmark_questions)}")
# Post a prediction
MetaculusApi.post_binary_question_prediction(
question_id=578, # Note that the question ID is not always the same as the post ID
prediction_in_decimal=0.012 # Must be between 0.01 and 0.99
)
print("Posted prediction")
# Post a comment
MetaculusApi.post_question_comment(
post_id=578,
comment_text="Here's example reasoning for testing... This will be a private comment..."
)
print("Posted comment")Question found with url: https://www.metaculus.com/questions/578
Question found with url: https://www.metaculus.com/questions/578
Num tournament questions: 11
Num filtered questions: 50
Num benchmark questions: 20
Posted prediction
Posted comment
The Smart Searcher acts like an LLM with internet access. It works a lot like Perplexity.ai API, except:
- It has clickable citations that highlights and links directly to the paragraph cited using text fragments
- You can ask the AI to use filters for domain, date, and keywords
- There are options for structured output (Pydantic objects, lists, dict, list[dict], etc.)
- Concurrent search execution for faster results
- Optional detailed works cited list
searcher = SmartSearcher(
temperature=0,
num_searches_to_run=2,
num_sites_per_search=10, # Results returned per search
include_works_cited_list=False # Add detailed citations at the end
)
response = await searcher.invoke(
"What is the recent news for Apple?"
)
print(response)Example output:
Recent news about Apple includes several significant developments:
Expansion in India: Apple is planning to open four more stores in India, with two in Delhi and Mumbai, and two in Bengaluru and Pune. This decision follows record revenues in India for the September 2024 quarter, driven by strong iPhone sales. Tim Cook, Apple's CEO, highlighted the enthusiasm and growth in the Indian market during the company's earnings call [1][4][5][8].
Product Launches: Apple is set to launch new iMac, Mac mini, and MacBook Pro models with M4 series chips on November 8, 2024. Additionally, the Vision Pro headset will be available in South Korea and the United Arab Emirates starting November 15, 2024. The second season of the Apple TV+ sci-fi series "Silo" will also premiere on November 15, 2024 [2][12].
... etc ...
You can also use structured outputs by providing a Pydantic model (or any other simpler type hint) and using the schema formatting helper:
from pydantic import BaseModel, Field
from forecasting_tools import SmartSearcher
class Company(BaseModel):
name: str = Field(description="Full company name")
market_cap: float = Field(description="Market capitalization in billions USD")
key_products: list[str] = Field(description="Main products or services")
relevance: str = Field(description="Why this company is relevant to the search")
searcher = SmartSearcher(temperature=0, num_searches_to_run=4, num_sites_per_search=10)
schema_instructions = searcher.get_schema_format_instructions_for_pydantic_type(Company)
prompt = f"""Find companies that are leading the development of autonomous vehicles.
Return as a list of companies with their details. Remember to give me a list of the schema provided.
{schema_instructions}"""
companies = await searcher.invoke_and_return_verified_type(prompt, list[Company])
for company in companies:
print(f"\n{company.name} (${company.market_cap}B)")
print(f"Relevance: {company.relevance}")
print("Key Products:")
for product in company.key_products:
print(f"- {product}")The schema instructions will format the Pydantic model into clear instructions for the AI about the expected output format and field descriptions.
The Key Factors Researcher helps identify and analyze key factors that should be considered for a forecasting question. As of last update, this is the most reliable of the tools, and gives something useful and accurate almost every time. It asks a lot of questions, turns search results into a long list of bullet points, rates each bullet point on ~8 criteria, and returns the top results.
from forecasting_tools import (
KeyFactorsResearcher,
BinaryQuestion,
ScoredKeyFactor
)
# Consider using MetaculusApi.get_question_by_id or MetaculusApi.get_question_by_url instead
question = BinaryQuestion(
question_text="Will YouTube be blocked in Russia?",
background_info="...", # Or 'None'
resolution_criteria="...", # Or 'None'
fine_print="...", # Or 'None'
id_of_post=0, # The ID and state only matters if using Metaculus API calls so feel free to set it to 0
)
# Find key factors
key_factors = await KeyFactorsResearcher.find_and_sort_key_factors(
metaculus_question=question,
num_key_factors_to_return=5, # Number of final factors to return
num_questions_to_research_with=26 # Number of research questions to generate
)
print(ScoredKeyFactor.turn_key_factors_into_markdown_list(key_factors))Example output:
- The Russian authorities have slowed YouTube speeds to near unusable levels, indicating a potential groundwork for a future ban. Source Published on 2024-09-12
- Russian lawmaker Alexander Khinshtein stated that YouTube speeds would be deliberately slowed by up to 70% due to Google's non-compliance with Russian demands, indicating escalating measures against YouTube. Source Published on 2024-07-25
- The press secretary of President Vladimir Putin, Dmitry Peskov, denied that the authorities intended to block YouTube, attributing access issues to outdated equipment due to sanctions. Source Published on 2024-08-17
- YouTube is currently the last Western social media platform still operational in Russia, with over 93 million users in the country. Source Published on 2024-07-26
- Russian users reported mass YouTube outages amid growing official criticism, with reports of thousands of glitches in August 2024. Source Published on 2024-08-09
The simplified pydantic structure of the scored key factors is:
class ScoredKeyFactor():
text: str
factor_type: KeyFactorType (Pro, Con, or Base_Rate)
citation: str
source_publish_date: datetime | None
url: str
score_card: ScoreCard
score: int
display_text: strThe Base Rate Researcher helps calculate historical base rates for events. As of last update, it gives decent results around 50% of the time. It orchestrates the Niche List Researcher and the Fermi Estimator to find base rate.
from forecasting_tools import BaseRateResearcher
# Initialize researcher
researcher = BaseRateResearcher(
"How often has Apple been successfully sued for patent violations?"
)
# Get base rate analysis
report = await researcher.make_base_rate_report()
print(f"Historical rate: {report.historical_rate:.2%}")
print(report.markdown_report)The Niche List Researcher helps analyze specific lists of events or items. The researcher will:
- Generate a comprehensive list of potential matches
- Remove duplicates
- Fact check each item against multiple criteria
- Return only validated items (unless include_incorrect_items=True)
from forecasting_tools import NicheListResearcher
researcher = NicheListResearcher(
type_of_thing_to_generate="Times Apple was successfully sued for patent violations between 2000-2024"
)
fact_checked_items = await researcher.research_niche_reference_class(
return_invalid_items=False
)
for item in fact_checked_items:
print(item)The simplified pydantic structure of the fact checked items is:
class FactCheckedItem():
item_name: str
description: str
is_uncertain: bool | None = None
initial_citations: list[str] | None = None
fact_check: FactCheck
type_description: str
is_valid: bool
supporting_urls: list[str]
one_line_fact_check_summary: str
class FactCheck(BaseModel):
criteria_assessments: list[CriteriaAssessment]
is_valid: bool
class CriteriaAssessment():
short_name: str
description: str
validity_assessment: str
is_valid_or_unknown: bool | None
citation_proving_assessment: str | None
url_proving_assessment: str | None:The Fermi Estimator helps break down numerical estimates using Fermi estimation techniques.
from forecasting_tools import Estimator
estimator = Estimator(
type_of_thing_to_estimate="books published worldwide each year",
previous_research=None # Optional: Pass in existing research
)
size, explanation = await estimator.estimate_size()
print(f"Estimate: {size:,}")
print(explanation)Example output (Fake data with links not added):
I estimate that there are 2,750,000 'books published worldwide each year'.
Facts:
- Traditional publishers release approximately 500,000 new titles annually in English-speaking countries [1]
- China publishes around 450,000 new books annually [2]
- The global book market was valued at $92.68 billion in 2023 [3]
- Self-published titles have grown by 264% in the last 5 years [4]
- Non-English language markets account for about 50% of global publishing [5]
Estimation Steps and Assumptions:
- Start with traditional English publishing: 500,000 titles
- Add Chinese market: 500,000 + 450,000 = 950,000
- Account for other major languages (50% of market): 950,000 * 2 = 1,900,000
- Add self-published titles (estimated 45% of total): 1,900,000 * 1.45 = 2,755,000
Background Research: [Additional research details...]
The GeneralLlm class is a wrapper around around litellm's acompletion function that adds some functionality like retry logic, calling the metaculus proxy, and cost callback handling. Litellm supports every model, most every parameter, and acts as one interface for every provider. See the litellm's acompletion function for a full list of parameters. Not all models will support all parameters. Additionally the Metaculus proxy doesn't support all models.
from forecasting_tools import GeneralLlm
result = await GeneralLlm(model="gpt-4o").invoke(prompt)
result = await GeneralLlm(model="claude-3-5-sonnet-20241022").invoke(prompt)
result = await GeneralLlm(model="metaculus/claude-3-5-sonnet-20241022").invoke(prompt) # Adding 'metaculus' Calls the Metaculus proxy
result = await GeneralLlm(model="gemini/gemini-pro").invoke(prompt)
result = await GeneralLlm(
model="perplexity/sonar-pro",
temperature=0.5,
max_tokens=1000,
top_p=0.95,
frequency_penalty=0,
presence_penalty=0,
stop=["\n\n"],
response_format=...
).invoke(prompt)Additionally GeneralLlm provides some interesting structured response options. It will call the model a number of times until it gets the response type desired. The type validation works for any type including pydantic types like list[BaseModel] or nested types like list[tuple[str,dict[int,float]]].
from forecasting_tools import GeneralLlm
from pydantic import BaseModel
class President(BaseModel):
name: str
age_at_death: int
biggest_accomplishment: str
model = GeneralLlm(model="gpt-4o")
pydantic_model_explanation = model.get_schema_format_instructions_for_pydantic_type(President)
prompt = clean_indents(f"""
You are a historian helping with a presidential research project. Please answer the following question:
Who is Abraham Lincoln?
Please provide the information in the following format:
{pydantic_model_explanation}
""")
verified_president = await model.invoke_and_return_verified_type(
prompt,
President,
allowed_invoke_tries_for_failed_output=2
)
regular_output = await model.invoke(prompt)
print(f"President is pydantic: {isinstance(verified_president, President)}")
print(f"President: {verified_president}")
print(f"Regular output: {regular_output}")President is pydantic: True
President: name='Abraham Lincoln' age_at_death=56 biggest_accomplishment="Preserving the Union during the American Civil War and issuing the Emancipation Proclamation, which began the process of freedom for America's slaves."
Regular output: ```json
{
"name": "Abraham Lincoln",
"age_at_death": 56,
"biggest_accomplishment": "Preserving the Union during the American Civil War and issuing the Emancipation Proclamation, which began the process of freedom for America's slaves."
}
```
You can also do some basic code execution.
from forecasting_tools import GeneralLlm, clean_indents
model = GeneralLlm(model="gpt-4o")
code = clean_indents("""
Please run a fermi estimate for the number of books published worldwide each year.
Generate only code and give your reasoning in comments. Do not include any other text, only code.
Assign your answer to the variable 'final_result'.
""")
result, code = await model.invoke_and_unsafely_run_and_return_generated_code(
code,
expected_output_type=float,
allowed_invoke_tries_for_failed_output=2
)
print(f"Result: {result}")
print(f"Code:\n{code}")Result: 10010000.0
Code:
# Fermi estimate for the number of books published worldwide each year
# Estimate the number of countries in the world
num_countries = 200 # Rough estimate of the number of countries
# Estimate the number of books published per country per year
# Assume a small country might publish around 100 books per year
# Assume a large country might publish around 100,000 books per year
# Use an average of these estimates for a rough calculation
avg_books_per_country = (100 + 100000) / 2
# Calculate the total number of books published worldwide each year
total_books_worldwide = num_countries * avg_books_per_country
# Assign the result to the variable 'final_result'
final_result = total_books_worldwide
The MonetaryCostManager helps to track AI and API costs. It tracks expenses and errors if it goes over the limit. Leave the limit empty to disable the limit. Any call made within the context of the manager will be logged in the manager. CostManagers are async safe and can nest inside of each other. Please note that cost is only tracked after a call finishes, so if you concurrently batch 1000 calls, they will all go through even if the cost is exceeded during the middle of the execution of the batch.
from forecasting_tools import MonetaryCostManager
from forecasting_tools import (
ExaSearcher, SmartSearcher, GeneralLlm
)
max_cost = 5.00
with MonetaryCostManager(max_cost) as cost_manager:
prompt = "What is the weather in Tokyo?"
result = await GeneralLlm(model="gpt-4o").invoke(prompt)
result = await SmartSearcher(model="claude-3-5-sonnet-20241022").invoke(prompt)
result = await ExaSearcher().invoke(prompt)
# ... etc ...
current_cost = cost_manager.current_usage
print(f"Current cost: ${current_cost:.2f}")The environment variables you need can be found in .env.template. Copy this template as .env and fill it in. As of last update, you only strictly need OPENAI_API_KEY and EXA_API_KEY.
Dev containers are reliable ways to make sure environments work on everyone's machine the first try and so you don't have to spend hours setting up your environment (especially if you have Docker already installed). If you would rather just use poetry, without the dev container, you can skip to "Alternatives to Docker". Otherwise, to get your development environment up and running, you need to have Docker Engine installed and running. Once you do, you can use the VSCode dev container pop-up to automatically set up everything for you.
For Windows and Mac, you will download Docker Desktop. For Linux, you will download Docker Engine. (NOTE: These instructions might be outdated).
First download and setup Docker Engine using the instructions at the link below for your OS:
- Windows: windows-install
- Mac: mac-install
- Linux: install
- Do not install Docker Desktop for Linux, rather, select your Linux distribution on the left sidebar and follow the distribution specific instructions for Docker engine. Docker Desktop runs with a different environment in Linux.
- Remember to follow the post-installation steps for Linux: linux-postinstall
Once Docker is installed, when you open up the project folder in VSCode, you will see a pop up noting that you have a setup for a dev container, and asking if you would like to open the folder in a container. You will want to click "open in container". This will automatically set up everything you need and bring you into the container. If the Docker process times out in the middle of installing python packages you can run the .devcontiner/postinstall.sh manually. You may need to have the VSCode Docker extension and/or devcontainer extension downloaded in order for the pop up to appear.
Once you are in the container, poetry should have already installed a virtual environment. For VSCode features to use this environment, you will need to select the correct python interpreter. You can do this by pressing Ctrl + Shift + P and then typing Python: Select Interpreter. Then select the interpreter that starts with .venv.
A number of vscode extensions are installed automatically (e.g. linting). You may need to wait a little while and then reload the window (once for things to install, and a second for needed reload after installation). You can install personal vscode extensions in the dev environment.
There are many ways to manager Docker containers, but generally if you download the vscode Docker extension, you will be able to stop/start/remove all containers and images.
If you choose not to run Docker, you can use poetry to set up a local virtual environment. If you are on Ubuntu, you should be able to just read through and then run .devcontainer/postinstall.sh. If you aren't on Ubuntu, check out the links in the postinstall file for where install instructions for dependencies were originally found. You may also want to take a look at VSCode extensions that would be installed (see the list in the .devcontainer/devcontainer.json file) so that some VSCode workplace settings work out of the box (e.g. automatic Black Formatting).
Given you have forked, if you have a custom bot you don't want committed to the repository when you add code the the package, its probably best to create a branch locally that holds your custom branch in order to leave the main branch open to pull updates or push changes you want to add. If you add all your custom files in a new folder (e.g. custom) then you shouldn't have merge conflicts, and can more easily remove the folder when ready to pull or push.
If you want to add your bot to the package, feel free to add your bot under forecasting/forecast_bots/community and other helper classes under folders that make sense.
You should be able to change up run_bot.py and merge, and a github action will make sure it doesn't overwrite the main repo's run_bot.py (though not a hard fix if not).
You can run any front end folder in the front_end directory by executing streamlit run front_end/Home.py. This will start a development server for you that you can run.
Streamlit makes it very easy to publish demos. If you change up the bot(s) within the front_end, you can quickly deploy a free demo others can play with (look up streamlit cloud).
This repository uses pytest tests are subdivided into folders 'unit_tests', 'low_cost_or_live_api', 'expensive'. Unit tests should always pass, while the other tests are for sanity checking. The low cost folder should be able to be run on mass without a huge cost to you. Do not run pytest without specifying which folder you want or else you will incur some large expenses from the 'expensive' folder.
Also it's helpful to use the log file that is automatically placed at logs/latest.log. You may need to set FILE_WRITING_ALLOWED=TRUE in .env for this to work.
- Fork the Repository: Fork the repository on GitHub. Clone your fork locally:
git clone [email protected]:your-username/forecasting-tools.git - Set Up Development Environment: Follow the "Local Development" section in the README to set up your environment
- Come up with an improvement: Decide on something worth changing. Perhaps, you want to add your own custom bot to the forecasting_bots folder. Perhaps you want to add a tool that you think others could benefit from. As long as your code is clean, you can probably expect your contribution to be accepted, though if you are worried about adoption, feel free to chat on our discord or create an issue.
- Make a pull request:
- Make changes
- Push your changes to your fork
- Make sure you rebase with the upstream main branch before doing a PR (
git fetch upstreamandgit rebase upstream/main) - Go to your fork in github, and choose the branch that you have that has your changes
- You should see a 'Contribute' button. Click this and make a pull request.
- Fill out the pull request template with a description of what changed and why and Url for related issues
- Request review from maintainers
- Respond to any feedback and make requested changes
-
Code Style
- Code is automatically formatted on commit using precommit (this should be automatically installed when you start the devcontainer). Formatting is done using Black, Ruff, and some other tools.
- Use type hints for all function parameters and return values
- Use descriptive variable names over comments
- Follow existing patterns in the codebase
-
Testing
- Add tests where appropriate for new functionality. We aren't shooting for full code coverage, but you shouldn't make none.
- Run unit tests locally before merging to check if you broke anything. See the 'Testing' section.
- Join our Discord for questions
- Open an issue for bugs or feature requests
Thank you for helping improve forecasting-tools!