pip install -U parea-ai
or install with Poetry
poetry add parea-ai
You can evaluate any step of your LLM app by wrapping it with a decorator, called trace
, and specifying the evaluation
function(s).
The scores associated with the traces will be logged to the Parea dashboard and/or in a
local CSV file if you don't have a Parea API key.
Evaluation functions receive an argument log
(of type Log) and should return a
float. You don't need to start from scratch, there are pre-defined evaluation
functions for general purpose,
chat, RAG, and summarization apps :)
You can define evaluation functions locally or use the ones you have deployed to Parea's Test Hub. If you choose the latter option, the evaluation happens asynchronously and non-blocking.
A fully locally working cookbook can be found here. Alternatively, you can add the following code to your codebase to get started:
import os
from parea import Parea, InMemoryCache, trace
from parea.schemas.log import Log
Parea(api_key=os.getenv("PAREA_API_KEY"), cache=InMemoryCache()) # use InMemoryCache if you don't have a Parea API key
def locally_defined_eval_function(log: Log) -> float:
...
@trace(eval_func_names=['deployed_eval_function_name'], eval_funcs=[locally_defined_eval_function])
def function_to_evaluate(*args, **kwargs) -> ...:
...
You can run an experiment for your LLM application by defining the Experiment
class and passing it the name, the data and the
function you want to run. You need annotate the function with the trace
decorator to trace its inputs, outputs, latency, etc.
as well as to specify which evaluation functions should be applied to it (as shown above).
from parea import Experiment
Experiment(
name="Experiment Name", # Name of the experiment (str)
data=[{"n": "10"}], # Data to run the experiment on (list of dicts)
func=function_to_evaluate, # Function to run (callable)
)
Then you can run the experiment by using the experiment
command and give it the path to the python file.
This will run your experiment with the specified inputs and create a report with the results which can be viewed under
the Experiments tab.
parea experiment <path/to/experiment_file.py>
Full working example in our docs.
You can iterate on your chains & agents much faster by using a local cache. This will allow you to make changes to your code & prompts without waiting for all previous, valid LLM responses. Simply add these two lines to the beginning your code and start a local redis cache:
from parea import Parea, InMemoryCache
Parea(cache=InMemoryCache())
If you set cache = None
for Parea
, no cache will be used.
You can benchmark your LLM app across many inputs by using the benchmark
command. This will run your the entry point
of your app with the specified inputs and create a report with the results.
parea benchmark --func app:main --csv_path benchmark.csv
The CSV file will be used to fill in the arguments to your function. The report will be a CSV file of all the traces. If you set your Parea API key, the traces will also be logged to the Parea dashboard. Note, for this feature you need to have a redis cache running. Please, raise a GitHub issue if you would like to use this feature without a redis cache.
You can automatically log all your LLM traces to the Parea dashboard by setting the PAREA_API_KEY
environment variable
or specifying it in the Parea
initialization.
This will help you debug issues your customers are facing by stepping through the LLM call traces and recreating the
issue
in your local setup & code.
from parea import Parea
Parea(
api_key=os.getenv("PAREA_API_KEY"), # default value
cache=...
)
import os
from dotenv import load_dotenv
from parea import Parea
from parea.schemas.models import Completion, UseDeployedPrompt, CompletionResponse, UseDeployedPromptResponse
load_dotenv()
p = Parea(api_key=os.getenv("PAREA_API_KEY"))
# You will find this deployment_id in the Parea dashboard
deployment_id = '<DEPLOYMENT_ID>'
# Assuming your deployed prompt's message is:
# {"role": "user", "content": "Write a hello world program using {{x}} and the {{y}} framework."}
inputs = {"x": "Golang", "y": "Fiber"}
# You can easily unpack a dictionary into an attrs class
test_completion = Completion(
**{
"deployment_id": deployment_id,
"llm_inputs": inputs,
"metadata": {"purpose": "testing"}
}
)
# By passing in my inputs, in addition to the raw message with unfilled variables {{x}} and {{y}},
# you we will also get the filled-in prompt:
# {"role": "user", "content": "Write a hello world program using Golang and the Fiber framework."}
test_get_prompt = UseDeployedPrompt(deployment_id=deployment_id, llm_inputs=inputs)
def main():
completion_response: CompletionResponse = p.completion(data=test_completion)
print(completion_response)
deployed_prompt: UseDeployedPromptResponse = p.get_prompt(data=test_get_prompt)
print("\n\n")
print(deployed_prompt)
async def main_async():
completion_response: CompletionResponse = await p.acompletion(data=test_completion)
print(completion_response)
deployed_prompt: UseDeployedPromptResponse = await p.aget_prompt(data=test_get_prompt)
print("\n\n")
print(deployed_prompt)
import os
import openai
from dotenv import load_dotenv
from parea import Parea
load_dotenv()
openai.api_key = os.getenv("OPENAI_API_KEY")
p = Parea(api_key=os.getenv("PAREA_API_KEY"))
x = "Golang"
y = "Fiber"
messages = [{
"role": "user",
"content": f"Write a hello world program using {x} and the {y} framework."
}]
model = "gpt-3.5-turbo"
temperature = 0.0
# define your OpenAI call as you would normally and we'll automatically log the results
def main():
openai.chat.completions.create(model=model, temperature=temperature, messages=messages).choices[0].message.content
Ready-to-use Pull Requests templates and several Issue templates.
- Files such as:
LICENSE
,CONTRIBUTING.md
,CODE_OF_CONDUCT.md
, andSECURITY.md
are generated automatically. - Semantic Versions specification
with
Release Drafter
.
This project is licensed under the terms of the Apache Software License 2.0
license.
See LICENSE for more details.
@misc{parea-sdk,
author = {joel-parea-ai},
title = {Parea python sdk},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/parea-ai/parea-sdk}}
}