BIA bob is a Jupyter-based assistant for interacting with data using large language models which generate Python code for Bio-Image Analysis (BIA).
It can make use of OpenAI's chatGPT, Google's Gemini, Anthropic's Claude, Github Models Marketplace, Helmholtz' blablador and Ollama.
You need an OpenAI API account or a Google Cloud account or a Helmholtz ID account to use it.
Using it with Ollama is free but requires running an Ollama server locally.
bob can write short Python code snippets and entire Jupyter notebooks for your image / data analysis workflow.
Caution
When using the OpenAI, Google Gemini, Anthropic, Github Models or any other endpoint via BiA-Bob, you are bound to the terms of service of the respective companies or organizations. The prompts you enter are transferred to their servers and may be processed and stored there. Make sure to not submit any sensitive, confidential or personal data. Also using these services may cost money.
You can initialize bob like this:
from bia_bob import bob
You can ask Bob to generate code like this:
%bob Load blobs.tif and show it
It will then respond with a Python code snippet that you can execute (see full example):
Use %bob if you want to write in the same line and %%bob if you want to write below.
If you want to continue using a variable in the next code cell, you need to specify the name of the variable in the following prompt.
When asking Bob explicitly to generate a notebook, it will put a new notebook file in the current directory with the generated code (See full example). You can then open it in Jupyter lab.
You can also ask Bob to modify an existing notebook, e.g. to introduce explanatory markdown cells (See full example):
Furthermore, one can translate Jupyter notebooks to other languages, e.g. by prompting %bob translate the filename.ipynb to <language>.
You can also ask Bob to write a prompt for you. This can be useful to explore potential strategies for analyzing image data. Note: It might be necessary to modify those prompts, especially when suggested analysis workflows are long and complicated. Shorten suggested prompts to the minimal necessary steps to answer your scientific question. (See full example).
You can add additional information from given Python variables into your prompt using the {variable_name} syntax.
With this, the content of the variable will become part of the prompt (full example).
If you are not sure what generated code does, you can ask Bob to explain it to you:
Bob can fix simple bugs in code you executed. Just add %%fix on top of the cell right after the error happened.
Using the %%doc magic, you can generate documentation for a given code cell.
Using the %%acc magic, you can replace common image processing functions with GPU-accelerated functions. It is recommended to check if the image processing results remain the same. You can see an example in this notebook.
You can use bia-bob from the terminal. This is recommended for creating notebooks for example like this:
bia-bob Please create a Jupyter Notebook that opens blobs.tif, segments the bright objects and shows the resulting label image on top of the original image with a curtain.
This can also be used to create other files, e.g. CSV files.
bia-bob is a research project aiming at streamlining the design of image analysis workflows. Under the hood it uses
artificial intelligence / large language models to generate text and code fulfilling the user's requests.
Users are responsible to verify the generated code according to good scientific practice. Some general advice:
- If you do not understand what a generated code snippet does, ask
%%bob explain this code in detail to a Python beginner:before executing the code. - After Bob generated a data analysis workflow for you, ask
%%bob How could I verify this analysis workflow ?. It is good scientific practice to measure the quality of segmentation results for example, or to measure the difference of automated quantitative measurements, in comparison to manual analysis. - If you are not sure if an image analysis workflow is valid, consider asking human experts. E.g. reach out via https://image.sc .
When sending a request to the LLM service providers, bia-bob sends the following information:
- The content of the cell you were typing
- If you entered variables using the
{variable}syntax, the content of these variables - All available variable, function and module names in the current Python environment (not variable values)
- A selected list of Python libraries that are installed
- The conversation history of the current session
- Optional: The image mentioned in the first line of the cell
This information is necessary to enable bia-bob to generate code that runs in your environment. If you want to know exactly what is sent to the server, you can activate verbose mode like this:
from bia_bob._machinery import Context
Context.verbose = True
If you want to ask bob a question, you need to put a space before the ?.
%bob What do you know about blobs.tif ?
You can install bia-bob using conda/pip. It is recommended to install it into a conda/mamba environment. If you have never used conda before, please read this guide first.
It is recommended to install bia-bob in a conda-environment together with useful tools for bio-image analysis.
conda env create -f https://github.com/haesleinhuepf/bia-bob/raw/main/environment.yml
You can then activate this environment...
conda activate bob_env
OR install bob into an existing environment:
pip install bia-bob
For using LLMs from remote service providers, you need to set an API key.
API Keys are short cryptic texts such as "proj_sk_asdasdasd" which allow you to log into a remote service without entering your username and password. Many online serives require using API keys for billing; others enable you to use their free services only after obtaining an API key.
This also means that you should not share your API key with others.
In the following sections, you find links to a couple of LLM services providers that are compatible with bia-bob.
After obtaining the key, you need to add it to the enviroment variables of your computer.
On Windows, you can do this by 1) searching for "env" in the start menu, 2) clicking on "Edit the system environment variables",
3) clicking on "Environment Variables", 4) clicking on "New" in the "System variables" section and adding a new variable with the name specified below (e.g. OPENAI_API_KEY) and the value of your API key.
On Linux and MacOS, this is typically done by modifying a hidden .bashrc or .zshrc file in the home directory, e.g. like this:
echo "export OPENAI_API_KEY='yourkey'" >> ~/.zshrc
Note: After setting the environment variables, you need to restart your terminal and/or Jupyter Lab to make them work.
See also further instructions on this page.
Create an OpenAI API Key and add it to your environment variables named OPENAI_API_KEY as explained on
You can then initialize Bob like this (optional, as that's default):
from bia_bob import bob
bob.initialize("gpt-4o-2024-08-06", vision_model="gpt-4o-2024-08-06")
Create an Anthropic API Key and add it to your environment variables named ANTHROPIC_API_KEY.
You can then initialize Bob like this:
from bia_bob import bob
bob.initialize(model="claude-3-5-sonnet-20240620", vision_model="claude-3-5-sonnet-20240620")
You can also apply for an API Key from the German Artificial Intelligence Service Center for Sensible and Critical Infrastructures who operates the ChatAI service.
You can store it in an environment variable named OPENAI_API_KEY and use initialize bob like this:
from bia_bob import bob
bob.initialize(endpoint="https://chat-ai.academiccloud.de/v1",
model="meta-llama-3.1-70b-instruct")
If you are using the models from Github Models Marketplace, please create an GITHUB API key (with default settings) and store it for accessing the models in an environment variable named GH_MODELS_API_KEY.
You can then access the models like this:
bob.initialize(
endpoint='github_models',
model='Phi-3.5-mini-instruct')
If you are using the models hosted on Microsoft Azure, please store your API key for accessing the models in an environment variable named AZURE_API_KEY.
You can then access the models like this:
bob.initialize(
endpoint='azure',
model='Phi-3.5-mini-instruct')
Alternatively, you can specify the endpoint directly, too:
bob.initialize(
endpoint='https://models.inference.ai.azure.com',
model='Phi-3.5-mini-instruct')
Custom endpoints can be used as well if they support the OpenAI API. Examples are blablador and ollama. An example is shown in this notebook:
For this, just install the openai backend as explained above (tested version: 1.5.0).
- If you want to use ollama and e.g. the
codellamamodel, you must runollama servefrom a separate terminal and then initialize bob like this:
bob.initialize(endpoint='ollama', model='codellama')
- If you want to use blablador, which is free for German academics, just get an API key as explained on
this page and store it in your environment as
BLABLADOR_API_KEYvariable.
bob.initialize(
endpoint='blablador',
model='Mistral-7B-Instruct-v0.2')
- Custom end points can be used as well, for example like this:
bob.initialize(
endpoint='http://localhost:11434/v1',
model='codellama')
Create a Google API key and store it in the environment variable GOOGLE_API_KEY.
pip install google-generativeai>=0.7.2
You can then initialize Bob like this:
from bia_bob import bob
bob.initialize("gemini-1.5-pro-002")
Note: This method is deprecated. Use gemini 1.5 as shown above.
pip install google-cloud-aiplatform
(Recommended google-cloud-aiplatform version >= 1.38.1)
To make use of the Google Cloud API, you need to create a Google Cloud account here and a project within the Google cloud (for billing) here. You need to store authentication details locally as explained here. This requires installing Google Cloud CLI. In very short: run the installer and when asked, activate the "Run gcloud init" checkbox. Or run 'gcloud init' from the terminal yourself. Restart the terminal window. After installing Google Cloud CLI, start a terminal and authenticate using:
gcloud auth application-default login
Follow the instructions in the browser. Enter your Project ID (not the name). If it worked the terminal should approximately look like this:
If you want to contribute to bia-bob, you can install it in development mode like this:
git clone https://github.com/haesleinhuepf/bia-bob.git
cd bia-bob
pip install -e .
If you are maintainer of a Python library and want to make BiA-bob aware of functions in your library, you can extend Bob's knowledge using entry-points. Add this to your library setup.cfg:
[options.entry_points]
bia_bob_plugins =
plugin1 = your_library._bia_bob_plugins:list_bia_bob_plugins
In the above mentioned _bia_bob_plugins.py define this function (and feel free to rename the function and the Python file):
def list_bia_bob_plugins():
"""List of function hints for bia_bob"""
return """
* Computes the sum of a and b
your_library.compute_sum(a:int,b:int) -> int
* Determines the difference between a and b
your_library.compute_difference(a:int, b:int) -> int
"""
Note that the syntax should be pretty much as shown above: A bullet point with a short description and a code-snippet just below.
You can also generate the list_bia_bob_plugins function as demonstrated in this notebook.
Please only list the most important functions. If the list of all plugins extending BiA-Bob becomes too long, the prompt will exceed the maximum prompt length.
List of known Python libraries that provide extensions to Bob:
(Feel free to extend this list by sending a pull-request)
There are similar projects:
- jupyter-ai
- chatGPT-jupyter-extension
- chapyter
- napari-chatGPT
- bioimageio-chatbot
- Claude Engineer
- BioChatter
- aider
- OpenDevin
- Devika
If you encounter any problems or want to provide feedback or suggestions, please create a thread on image.sc along with a detailed description and tag @haesleinhuepf .