scholaris 0.0.1

A Python package to set up an advanced research assistant leveraging function calling capabilities with LLMs on your local computer

Scholaris

[!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.

Installation

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. Setup a new virtual environment, such as with Conda. 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=3.12

Step 3. Activate the virtual environment:

$ conda activate scholaris-env

Step 4. Install latest scholaris Python package from the GitHub repository:

$ pip install git+https://github.com/nicomarr/scholaris.git

or from pypi:

$ pip install scholaris

Step 5. Install the following dependencies:

$ pip install ollama
$ pip install PyPDF2
$ pip install requests
$ pip install tqdm

Getting started

Import the scholaris core module. If you work from a Jupyter notebook environment within a different directory, you may need to add the parent directory of the current working directory to the Python system path as shown below.

from scholaris.core import *

Make sure the Ollama app is installed on your local computer and Llama 3.1 8B has been downloaded and is running. Then, initialize the Assistant class with Llama 3.1 8B, the core functions and 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, like so:

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

[!NOTE]

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)

You can also explicitly set the model by passing the model name as an argument to the Assistant class. If you do not pass a model name, the default model is Llama 3.1 8B. An alternative model that supports tool calling and can be run on a standard laptop is Mistral’s NeMo 12 B model. To use the latter model, change the attribute in the Assistant class to “mistral-nemo”. For more information, read the following blog post.

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

How to use

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("Tell me about the tools you have available.")

I'm happy to chat with you! I have a range of tools at my disposal, designed to assist with various tasks.

Firstly, I can summarize local documents (PDFs, markdown files, or text documents) using the `summarize_local_document` tool. This function will extract key information and provide a concise summary of the content.

I also have access to the `describe_tools` tool, which provides an overview of my available tools and capabilities. Feel free to ask me about specific tools if you're unsure what they do!

Additionally, I can help with tasks related to Python code using the `describe_python_code` tool. This function will provide a description of the purpose and structure of Python code in local files.

If you need information on research articles, I have access to tools that can retrieve metadata from OpenAlex (the `query_openalex_api` tool) or Semantic Scholar (the `query_semantic_scholar_api` tool). These functions use article titles, PubMed IDs, or digital object identifiers as query parameters.

I also have a tool for converting various types of research IDs (ID_converter_tool), which can convert between PMIDs, PMCIDs, DOIs, and manuscript IDs.

Lastly, I can respond to generic questions using the `respond_to_generic_queries` function if no other specific tools are available.

That's an overview of my capabilities! If you have any specific questions or requests, feel free to ask.

[!TIP]

You can also access the assistant’s responses from the message attribute, like so:

assistant.messages[-1]["content"] # Access the last message

response == assistant.messages[-1]["content"]

True

By 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...
Based on the available tools, I have access to 1 PDF file:

*jci.insight.144499.v2.pdf*

This file is located in the local data directory and has a .pdf extension. If you'd like me to summarize this document or describe its content, please let me know!

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,  7.59s/it]



Yes, this document has a title.

The title of the PDF file *jci.insight.144499.v2.pdf* is:

"Distinct antibody repertoires against endemic human coronaviruses in children and adults"

This information was retrieved using the `get_titles_and_first_authors` tool.

Conversation history

Show the conversation history by calling the assistant’s show_conversation_history() method:

assistant.show_conversion_history()

User: Tell me about the tools you have available.

Assistant response: I'm happy to chat with you! I have a range of tools at my disposal, designed to assist with various tasks.

Firstly, I can summarize local documents (PDFs, markdown files, or text documents) using the `summarize_local_document` tool. This function will extract key information and provide a concise summary of the content.

I also have access to the `describe_tools` tool, which provides an overview of my available tools and capabilities. Feel free to ask me about specific tools if you're unsure what they do!

Additionally, I can help with tasks related to Python code using the `describe_python_code` tool. This function will provide a description of the purpose and structure of Python code in local files.

If you need information on research articles, I have access to tools that can retrieve metadata from OpenAlex (the `query_openalex_api` tool) or Semantic Scholar (the `query_semantic_scholar_api` tool). These functions use article titles, PubMed IDs, or digital object identifiers as query parameters.

I also have a tool for converting various types of research IDs (ID_converter_tool), which can convert between PMIDs, PMCIDs, DOIs, and manuscript IDs.

Lastly, I can respond to generic questions using the `respond_to_generic_queries` function if no other specific tools are available.

That's an overview of my capabilities! If you have any specific questions or requests, feel free to ask.

User: Which PDF files do you have access to in the local data directory

Assistant response: Based on the available tools, I have access to 1 PDF file:

*jci.insight.144499.v2.pdf*

This file is located in the local data directory and has a .pdf extension. If you'd like me to summarize this document or describe its content, please let me know!

User: Does this document have a title?

Assistant response: Yes, this document has a title.

The title of the PDF file *jci.insight.144499.v2.pdf* is:

"Distinct antibody repertoires against endemic human coronaviruses in children and adults"

This information was retrieved using the `get_titles_and_first_authors` tool.

[!NOTE]

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_conversion_history(show_function_calls=True)

Clear the conversation history by calling the assistant’s clear_conversation_history() method:

assistant.clear_conversion_history()

Assistant parameters

Show the model by printing the assistant object:

assistant

Assistant, 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.

Local data access

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")

Loaded Semantic Scholar API key from the environment variables.
Loaded email address from the environment variables.
A local directory /Users/user2/GitHub/scholaris/proprietary_data already exists for storing data files. No of files: 0

Tools

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.

Authentication for tool calling and API access (optional)

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": "[e-mail]"
}
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...

:::{.callout-tip} When working in a Jupyter notebook environment, you can quickly display details of the class methods by using special syntax. Type the name of the object, followed by a ?, or type ?? to get more detailed information (i.e., the source code), like so:

assistant.chat?
assistant??

Developer Guide

Defining new tools

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]

Here are key points to consider when defining a new tool:

• Use type hints in the function signature to define the input and output types.
• Use docstrings, as shown below, to describe the function’s purpose and the expected input and output.
• Use the @json_schema_decorator decorator 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 called multiply_two_integers that takes two integers as input and returns a string as output:

@json_schema_decorator
def multiply_two_integers(a: int, b: int) -> int:
    """
    A function to multiply two numbers.

    Args:
        a (int): First integer.
        b (int): Second integer.

    Returns:
        str: The product of the two numbers, as a string.
    """

    if not isinstance(a, int) or not isinstance(b, int):
        return "Error: Inputs must be integers."
    
    return str(a * b)

To ensure the JSON schema is generated correctly, you can call the json_schema attribute of the function:

multiply_two_integers.json_schema

{'type': 'function',
 'function': {'name': 'multiply_two_integers',
  'description': 'A function to multiply two numbers.',
  'parameters': {'type': 'object',
   'properties': {'a': {'type': 'integer', 'description': 'First integer.'},
    'b': {'type': 'integer', 'description': 'Second integer.'}},
   'required': ['a', 'b']}}}

Alternatively, you can use the generate_json_schema function:

generate_json_schema(multiply_two_integers)

{'type': 'function',
 'function': {'name': 'multiply_two_integers',
  'description': 'A function to multiply two numbers.',
  'parameters': {'type': 'object',
   'properties': {'a': {'type': 'integer', 'description': 'First integer.'},
    'b': {'type': 'integer', 'description': 'Second integer.'}},
   'required': ['a', 'b']}}}

Adding tools

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_integers": multiply_two_integers})

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_integers

assistant.chat("What is the product of 4173 and 351?", show_progress=True)

Selecting tools...

[{'function': {'name': 'multiply_two_integers', 'arguments': {'a': '4173', 'b': '351'}}}]
Calling multiply_two_integers() with arguments {'a': '4173', 'b': '351'}...

Generating final response...
To calculate the product, I'll treat the inputs as integers.

The product of 4173 and 351 is 1,464,303.

"To calculate the product, I'll treat the inputs as integers.\n\nThe product of 4173 and 351 is 1,464,303."

Adding methods

You can add new methods to the Assistant class by using the add_to_class() decorator function. For example, the clear_conversion_history method was added to the Assistant class as follows:

@add_to_class(Assistant)
def clear_conversion_history(self):
    """Clear the conversation history."""
    self.messages = [{'role': "system", 'content': self.sys_message},]

Contributing to Scholaris

This 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.

Acknowledgements

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!