Scholaris is a Python package that sets up a research assistant on your local computer, leveraging function calling capabilities. Designed specifically for health and life sciences, it helps researchers gain insights from scholarly articles by integrating with the Ollama Python library.
- Local setup: No dependency on cloud-hosted LLMs for inference.
- Extract data from local files: Built-in tools to extract data from PDFs, py files, and plain text or markdown files.
- External data retrieval: Built-in tools to make API calls to external data sources, such as OpenAlex, Semantic Scholar, or NCBI’s Any ID converter.
- Customizable and extensible architecture: Easily extend the functionality of the assistant by adding new tools with only a few lines of code.
Warning
This Python package is under active development and is not yet ready for production use. Report any issues or feature requests on the GitHub repository.
Step 1. Download Ollama and follow the instructions to install Ollama for your operating system. Then, pull and run llama3.1 (parameters: 8B, quantization: Q4_0, size: 4.7 GB) according to the ollama documentation.
Step 2. Go to your terminal and setup a new virtual environment, such as with Conda:
Tip
Quick command line instructions on how to install Miniconda, a free minimal installer for Conda, can be found here.
$ conda create -n scholaris-env python=3.12Step 3. Activate the virtual environment:
$ conda activate scholaris-envStep 4. Install latest scholaris Python package:
$ pip install scholaris$ pip install git+https://github.com/nicomarr/scholaris.gitStep 1. Open a Jupyter notebook, IPython environment, or start Python from the terminal and import the scholaris core module:
from scholaris.core import *Step 2. Initialize the
Assistant
class:
Caution
Make sure the Ollama app is installed on your local computer and Llama 2 7B has been downloaded and is running before initializing the assistant. Otherwise, the initialization will abort.
Note
If no additional arguments are passed, the assistant is initialized with Llama 3.1 8B, a set of core functions and a default system message. During initialization, messages are printed to indicate whether credentials such as API keys and email are loaded from the environment variables (more on that below), and whether a local data directory already exists or has been created.
Initialize the
Assistant
class with the default parameters:
assistant = Assistant()Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
Tip
For the examples shown below, a local data directory had already been created in the parent directory prior to initialization, and a PDF file was downloaded and saved to the local data directory for demonstration purposes. You can download the same PDF file, or additional/other files by running the following commands in a Jupyter notebook cell. Alternatively, you can also download and add files manually to the local data directory after initialization.
!mkdir -p ../data
pdf_urls = [
"https://df6sxcketz7bb.cloudfront.net/manuscripts/144000/144499/jci.insight.144499.v2.pdf",
# Add more URLs here as needed
]
for url in pdf_urls:
!curl -o ../data/$(basename {url}) {url}To see where data are stored afer initialization, simply call the
dir_path attribute of the assistant object, like so:
print(assistant.dir_path)Explicitly set or change the model by passing the model name as an
argument to the
Assistant
class:
assistant = Assistant(model="llama3.1:latest")Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
Note
Download and select a model that supports tool calling. At the time of writing, the following models are supported:
- llama3.1
- llama3.2
- qwen2.5
- mistral-nemo
- nemotron-mini
- command-r
- command-r-plus
For more information, read the following blog post.
To chat with the assistant, simply call the chat() method with your
prompt as input. You also can store the response in a variable, but this
is optional. By default, the assistant will stream the response and
store the conversation history.
response = assistant.chat("Briefly tell me about the tools you have available.")I can summarize research articles from various sources, provide information on scientific topics, and help with general knowledge queries. I have access to a range of tools that allow me to extract specific details from PDF documents, convert IDs between different formats (e.g., PubMed ID, DOI), and query OpenAlex and Semantic Scholar APIs for article metadata. Additionally, I can describe the purpose and content of local Python files. However, my capabilities are limited to the information provided by the available tools, so if a tool's content is empty or unavailable, I may not be able to provide a response.
Tip
You can also access the assistant’s responses from the message
attribute, like so:
assistant.messages[-1]["content"] # Access the last messageBy setting show_progress=True, you can see the step-by-step progress
of the fuction calling. This includes the tool choice, selected
arguments, if applicable, and the output of the called function that is
being passed back to the LLM to generate the final response.
response = assistant.chat("Which PDF files do you have access to in the local data directory", show_progress=True)Selecting tools...
[{'function': {'name': 'get_file_names', 'arguments': {'ext': 'pdf'}}}]
Calling get_file_names() with arguments {'ext': 'pdf'}...
Generating final response...
I have access to a single PDF file named "jci.insight.144499.v2.pdf" in the local data directory.
By default, streaming is enabled. If you like to disable streaming, set
stream=False. This will store the entire conversation history in the
messages attribute, which can be accessed as shown above.
response = assistant.chat("Does this document have a title?", stream_response=False)Extracting titles and first authors: 100%|██████████| 1/1 [00:07<00:00, 7.72s/it]
Yes, the PDF file "jci.insight.144499.v2.pdf" has a title: "Distinct antibody repertoires against endemic human coronaviruses in children and adults".
Show the conversation history by calling the assistant’s
show_conversation_history() method:
assistant.show_conversation_history()User: Briefly tell me about the tools you have available.
Assistant response: I can summarize research articles from various sources, provide information on scientific topics, and help with general knowledge queries. I have access to a range of tools that allow me to extract specific details from PDF documents, convert IDs between different formats (e.g., PubMed ID, DOI), and query OpenAlex and Semantic Scholar APIs for article metadata. Additionally, I can describe the purpose and content of local Python files. However, my capabilities are limited to the information provided by the available tools, so if a tool's content is empty or unavailable, I may not be able to provide a response.
User: Which PDF files do you have access to in the local data directory
Assistant response: I have access to a single PDF file named "jci.insight.144499.v2.pdf" in the local data directory.
User: Does this document have a title?
Assistant response: Yes, the PDF file "jci.insight.144499.v2.pdf" has a title: "Distinct antibody repertoires against endemic human coronaviruses in children and adults".
Tip
The show_conversation_history() method can be called with the
show_function_calls=True argument to display the function calls made
by the assistant during the conversation, and the output of the
function calls. This can be useful for understanding the assistant’s
responses, and for debugging purposes.
assistant.show_conversation_history(show_function_calls=True)Clear the conversation history by calling the assistant’s
clear_conversation_history() method:
assistant.clear_conversation_history()Show the model by printing the assistant object:
assistantAssistant, powered by llama3.1
Or show the model by accessing the assistant’s model attribute:
assistant.model'llama3.1:latest'
Show the system messages by accessing the assistant’s sys_message attribute:
print(assistant.sys_message)You are an AI assistant specialized in analyzing research articles.
Your role is to provide concise, human-readable responses based on information from tools and conversation history.
Key instructions:
1. Use provided tools to gather information before answering.
2. Interpret tool results and provide clear, concise answers in natural language.
3. If you can't answer with available tools, state this clearly.
4. Don't provide information if tool content is empty.
5. Never include raw JSON, tool outputs, or formatting tags in responses.
6. Format responses as plain text for direct human communication.
7. Use clear formatting (e.g., numbered or bulleted lists) when appropriate.
8. Provide article details (e.g., DOI, citation count) in a conversational manner.
Act as a knowledgeable research assistant, offering clear and helpful information based on available tools and data.
By default, the assistant has access to a single directory, called
data. Within this directory, the assistant can list and read the
following file formats and extensions: pdf, txt, md, markdown, csv, and
py. If not already present, the directory is created when the assistant
is initialized. If you want to change the directory name, you can do so
by passing the desired directory name as an argument to the
Assistant
class when it is initialized. For example, to create a directory called
proprietary_data, you would initialize the assistant as follows:
assistant = Assistant(dir_path="../proprietary_data")By default, the assistant can call a set of core tools or functions
which are passed to the
Assistant
as a dictionary when it is initialized. With these tools or functions,
the assistant will be able to get a list of file names in a specific
data directory, can extract content from these files, and summarize
them. In addition, the assistant can make API calls to external data
sources, such as OpenAlex or Semantic
Scholar, to retrieve information about
a large number of scholarly articles. The tools available to the
assistant can be viewed by accessing the assistant’s list_tools()
method as follows:
assistant.list_tools()get_file_names
extract_text_from_pdf
get_titles_and_first_authors
summarize_local_document
describe_python_code
id_converter_tool
query_openalex_api
query_semantic_scholar_api
respond_to_generic_queries
describe_tools
Tip
You can learn more details about the core tools by visiting the Source
Code page, which lists each function and provides a brief description
of its purpose, functionality, required arguments, and usage (the
docstring). This information helps you understand the available tools
and how the LLM uses them. Alternatively, you can execute the
assistant.pprint_tools() or assistant.get_tools_schema() methods.
Some tools take optional authentication parameters, such as an API key
or email. For example, the query_semantic_scholar tool takes an
optional API key to access the Semantic Scholar API, which will increase
the API rate limit. Request a Semantic Scholar API Key
here. Similarly,
the query_openaplex_api tool takes an optional email parameter to
access the OpenAlex API, which is recommended as a best practice and
kindly requested by the API
provider.
The best way to pass these parameters is to set them as environment
variables, with the following key names: SEMANTIC_SCHOLAR_API_KEY and
EMAIL. The Assistant class will automatically read these environment
variables when initialized and pass them to the tools that require them.
Alternatively, you can pass the Semantic Scholar API key and your email
by simply adding the authentication argument when initializing the
Assistant class, as shown below:
authentication = {
"SEMANTIC_SCHOLAR_API_KEY": "your_api_key",
"EMAIL": "[email protected]"
}
assistant = Assistant(authentication=authentication)A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
If you want to change the core functions, you can do so by passing the desired core functions as an argument to the Assistant class when it is initialized. For example, to limit the assitant’s ability to respond to generic questions and to access external data by making requests to the OpenAlex and Semantic Scholar API’s, you would initialize the assistant as follows:
assistant = Assistant(tools = {
"query_openalex_api": query_openalex_api,
"query_semantic_scholar_api": query_semantic_scholar_api,
"respond_to_generic_queries": respond_to_generic_queries,
})Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
Note
The research assistent is set up so that it has to use a tool to generate a final response to a user’s prompt. This is to ensure that the assistant is primarily providing information which is relevant for health and life sciences. Otherwise it will abort the conversation, like so:
assistant = Assistant(tools = {})
assistant.chat("What is the capital of France?")Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
No tools provided! Please add tools to the assistant.
No tool calls found in the response. Adding an empty tool_calls list to the conversation history. Aborting...
Tip
When working in a Jupyter notebook or another iPython environment, you
can quickly display details of a class method or fucntion by using
special syntax. Type the name of the method or function, followed by a
?, or type ?? to get more detailed information (i.e., the
docstring and basic information, or the source code, respectively).
For example, to get information about the chat method, you can type
the following:
assistant.chat?
assistant??You can define new functions to be used by the assistant as tools. To
simplify this process, a decorator called @json_schema_decorator is
provided so it is not necessary to define the schema for the function.
The schema is automatically generated based on the function’s annotation
and docstring.
Tip
- Use type hints in the function signature to define the input and output types.
- Use Google style docstrings, as shown below, to describe the function’s purpose and the expected input and output.
- Use the
@json_schema_decoratordecorator to automatically generate the schema for the function. - Ensure the output is a string (such as a JSON-formatted string) that can be passed back to the LLM to generate the final response.
Important
It’s crucial to understand that this metadata (function name, type hints, and docstring) is all the information the LLM has access to when deciding which function to call and how to use it. The LLM does not have access to or information about the actual source code or implementation of the functions (unless explicitly provided). Therefore, the metadata must be comprehensive and accurate to ensure proper function selection and usage by the LLM.
The following example shows how to define a new tool to multiply two numbers, which takes as input two integers or strings that can be converted to integers, and returns the product of the two numbers as a string:
from typing import Union
@json_schema_decorator
def multiply_two_numbers(a: Union[int, str], b: Union[int, str]) -> str:
"""
A function to multiply two numbers.
Args:
a: First number, can be an integer or a string representation of an integer.
b: Second number, can be an integer or a string representation of an integer.
Returns:
str: The product of the two numbers, as a string.
Raises:
ValueError: If the inputs cannot be converted to integers.
"""
try:
int_a = int(a)
int_b = int(b)
return str(int_a * int_b)
except ValueError:
return "Error: Inputs must be integers or string representations of integers."Note
Type hints are recommended but not required, unless you want to define
optional arguments. In that case, you can use the Optional type hint
from the typing module. This will determine which arguments are
included in the required list values, as shown below.
To ensure the JSON schema is generated correctly, you can call the
json_schema attribute of the function:
multiply_two_numbers.json_schema{'type': 'function',
'function': {'name': 'multiply_two_numbers',
'description': 'A function to multiply two numbers.',
'parameters': {'type': 'object',
'properties': {'a': {'type': 'object',
'description': 'First number, can be an integer or a string representation of an integer.'},
'b': {'type': 'object',
'description': 'Second number, can be an integer or a string representation of an integer.'}},
'required': ['a', 'b']}}}
Alternatively, you can use the
generate_json_schema
function:
generate_json_schema(multiply_two_numbers)You can add new tools by passing a dictionary of new tools to the
Assistant
class when it is initialized. Use the add_tools argument to add new
tools to the assistant. This will merge the new tools with the existing
tools. For example, to add the new tool called multiply_two_integers
to the assistant, you would initialize the assistant as follows:
assistant = Assistant(add_tools = {"multiply_two_numbers": multiply_two_numbers})Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/data already exists for storing data files. No of files: 1
You can confirm that the new tool has been added to the list of existing
tools by using the list_tools() method:
assistant.list_tools()get_file_names
extract_text_from_pdf
get_titles_and_first_authors
summarize_local_document
describe_python_code
id_converter_tool
query_openalex_api
query_semantic_scholar_api
respond_to_generic_queries
describe_tools
multiply_two_numbers
response = assistant.chat("What is the product of 4173 and 351?", show_progress=True)Selecting tools...
[{'function': {'name': 'multiply_two_numbers', 'arguments': {'a': '4173', 'b': '351'}}}]
Calling multiply_two_numbers() with arguments {'a': '4173', 'b': '351'}...
Generating final response...
The product of 4173 and 351 is 1,464,723.
You can add new methods to the
Assistant
class by using the add_to_class() decorator function, like so:
@add_to_class(Assistant)
def new_method(self):
# Method implementation
passThis Python package has been developed using
nbdev. To contribute to this package, install
nbdev and follow the nbdev
documentation to set up
your development environment.
Many thanks to the developers of Ollama and the Ollama Python Library for providing the core functionality that Scholaris is built upon, and thanks to all the providers of open-source and open-weight models. Special thanks to the developers of nbdev for making it easy to develop and document this package, and for many insightful tutorials and inspirations!