Promptimize is a prompt engineering evaluation and testing toolkit.
It accelerates and provides structure around prompt engineering at scale with confidence, bringing some of the ideas behind test-driven development (TDD) to engineering prompts.
With promptimize, you can:
- Define your "prompt cases" (think "test cases" but specific to evaluating prompts) as code and associate them with evaluation functions
- Generate prompt variations dynamically
- Execute and rank prompts test suites across different engines/models/temperature/settings and compare results, brining the hyperparameter tuning mindset to prompt engineering
- Get reports on your prompts' performance as you iterate. Answer question around how different prompt suites are performing against one-another. Which individual cases or categories of cases improved? regressed?
- Minimize API calls! only re-assess what changed as you change it
- Perform human if and where needed, introspected failed cases, overriding false negatives
In essence, promptimize provides a programmatic way to execute and fine-tune your prompts and evaluation functions in Python, allowing you to iterate quickly and with confidence.
# Brining some "prompt generator" classes - note that you can derive and extend those
from promptimize.prompts import PromptCase
# Bringing some useful eval function that help evaluating and scoring responses
# eval functions have a handle on the prompt object and are expected
# to return a score between 0 and 1
from promptimize import evals
# Promptimize will scan the target folder and find all Prompt objects
# and derivatives that are in the python modules
simple_prompts = [
# Prompting "hello there" and making sure there's "hi" or "hello"
# somewhere in the answer
PromptCase("hello there!", lambda x: evals.any_word(x, ["hi", "hello"])),
PromptCase(
"name the top 50 guitar players!", lambda x: evals.all_words(x, ["frank zappa"])
),
]
$ promptimize run --help
Usage: promptimize run [OPTIONS] PATH
run some prompts
Options:
-v, --verbose Trigger more verbose output
-f, --force Force run, do not skip
-h, --human Human review, allowing a human to review and force
pass/fail each prompt case
-r, --repair Only re-run previously failed
-x, --dry-run DRY run, don't call the API
--shuffle Shuffle the prompts in a random order
-s, --style [json|yaml] json or yaml formatting
-m, --max-tokens INTEGER max_tokens passed to the model
-l, --limit INTEGER limit how many prompt cases to run in a single
batch
-t, --temperature FLOAT max_tokens passed to the model
-e, --engine TEXT model as accepted by the openai API
-k, --key TEXT The keys to run
-o, --output PATH
-s, --silent
Let's run those examples and produce a report ./report.yaml
$ promptimize run examples/ --output ./report.yaml
💡 ¡promptimize! 💡
# ----------------------------------------
# (1/2) [RUN] prompt: prompt-115868ef
# ----------------------------------------
key: prompt-115868ef
user_input: hello there!
prompt_hash: 115868ef
response: Hi there! How are you doing today?
execution:
api_call_duration_ms: 883.8047981262207
run_at: '2023-04-25T02:21:40.443077'
score: 1.0
# ----------------------------------------
# (2/2) [RUN] prompt: prompt-5c085656
# ----------------------------------------
key: prompt-5c085656
user_input: name the top 10 guitar players!
prompt_hash: 5c085656
response: |-
1. Jimi Hendrix
2. Eric Clapton
{{ ... }}
11. Carlos Santana
weight: 2
execution:
api_call_duration_ms: 2558.135747909546
run_at: '2023-04-25T02:21:43.007529'
score: 0.0
# ----------------------------------------
# Suite summary
# ----------------------------------------
suite_score: 0.3333333333333333
git_info:
sha: 2cf28498ba0f
branch: main
dirty: true
Thousands of product builders are currently trying to figure out how to bring the power of AI into the products and experiences they are building. The probabilistic (often semi-random, sometimes hectic) nature of LLMs makes this a challenge.
Prompt engineering is a huge piece of the puzzle in terms of how to do this right, especially given the complexity, risks, and drawbacks around model tuning.
We believe product builders need to tame AI through proper, rigorous prompt engineering. This allows making the probabilistic nature of AI more deterministic, or somewhat predictable, and allows builders to apply a hyperparameter tuning-type mindset and approach to prompt engineering.
Any prompt-generator logic that's going to be let loose in the wild inside a product should be thoroughly tested and evaluated with "prompt cases" that cover the breath of what people may do in a product.
In short, Promptimize allows you to test prompts at industrial scale, so that you can confidently use them in the products you are building.
- Prompt: A Prompt instance is a certain test case, a single prompt with an associated set of evaluation functions to rate its success.
- Evaluation: An evaluation function that reads the response and returns
a success rate between
0
and1
. - Suite: A Suite is a collection of Prompt; it's able to run things, accumulate results, and print reports about its collection of use cases.
- Report: a report is the compiled results of running a certain prompt
Suite
or set of suites. Reports can be consumed, compared, and expanded.
- Configuration as code: All prompt cases, suites, and evaluations are defined as code, which makes it easy to dynamically generate all sorts of use cases and suites.
- Expressive: a clean DSL that's to-the-point -> user prompt + assertions.
the actually prompt creation logic lives in the derivative class of
PromptCase
, so that we can have clean, dense files that contain niceSuite
s - Support the iteration mindset: making it easy for people to try things, get suggestions from the AI, adapt, compare, and push forward
- Extensibility: the toolkit is designed to be extremely hackable and extensible. Hooks, extensions, high API surface.
- AI-powered: the framework offers ways to expand your suites based on the examples that exists. Use AI to generate more prompt cases!
Listing out a few features you should know about that you can start using as your suites of prompts become larger / more complex
- evaluation functions are assumed to return a value between 0 and 1. contrarily to unit tests, prompt cases aren't boolean
- prompts can be assigned a
weight
(default 1) this enables you to define which prompts are more important than others for reporting purposes and suite evaluation - prompts can be assigned a
category
, this can be used in the reporting. That helps understanding which categories are performing better than others, or are most affected by iterations - The
Prompt
classpre_run
andpost_run
hooks if you want to do post-processing for instance. An example of that would be if you do a prompt that expects GPT to generate code, and you'd like actually say run that code and test it. In our SQL implementation, we run the SQL against the database for instance and get a pandas dataframe back, and allow doing assertions on the dataframe itself
To install the Promptimize package, use the following command:
pip install promptimize
First you'll need an openai API key, let's set it as an env var
export OPENAI_API_KEY=sk-{{ REDACTED }}
Find the examples executed below here
# Clone the repo
git clone git@github.com:preset-io/promptimize.git
cd promptimize
# NOTE: CLI is `promptimize`, but `p9e` is a shorter synonym, can be used interchangibly
# First let's run some of the examples
p9e run ./examples
# Now the same but with verbose output
p9e run ./examples --verbose --output ./report.yaml
How does promptimize relate to langchain
?
We think langchain is amazing and promptimize uses langchain under the
hood to interact with openai, and has integration with langchain
(see LangchainPromptCase
, and the upcoming LangchainChainPromptCase
and LangchainAgntPromptCase
)
While you don't have to use
langchain, and could use promptimize on top of any python prompt generation
whether it'd be another library or some home grown thing.
Where is promptimize
coming from!? I'm (Maxime Beauchemin) a startup
founder at Preset working on brining AI to BI
(data exploration,
and visualization). At Preset, we use promptimize
to generate
complex SQL based on natural language, and to suggest charts to users. We
derive the SimpleQuery
class to make it fitted to our specific use
cases in our own prompt engineering repo. It's not my first open source project
as the creator of
Apache Superset and
Apache Airflow
This project is in its super early stages as of 0.2.0
, and contributions,
contributors, and maintainers are highly encouraged. While it's a great time
to onboard and influence the direction of the project, things are still
evolving quickly. To get involved, open a GitHub issue
or submit a pull request!