Enable LLM-Driven Data Exploration with Presets and 3W Integration

[zRafaF]

Overview

This PR lays the groundwork for integrating Large Language Models (LLMs) into BibMon. It introduces a client that allows users to interact with endpoints for data processing and exploration. As a starting point, we have implemented access to the 3W dataset, enabling the model to infer the most relevant column based on the provided data.

Data tailoring is handled through what we call presets, which are located in bibmon/llm/presets. These presets allow users to customize how data is structured before being sent to the model.

Limitations

This feature acts purely as a client, meaning it requires an external endpoint for model interaction. For instance, you can use OpenAI's API or self-host an alternative model, as we have done.

Direct LLM inference within BibMon could be achieved using tools like llama-cpp-python or similar. However, this approach was avoided to prevent unnecessary complexity and bloat in the library.

While fine-tuning the LLM and creating more precise instructions for a dataset is possible, it requires a detailed data annotation process. Additional information on this can be found in our auxiliary notebook.

Note: This PR is dependent on #50.

Implementation Details (3W Dataset Integration)

Data Preset

The data sent to the model follows this structure:

{
    "event_name": "string",
    "event_description": "string",
    "columns_and_description": "dict",
    "data": [
        {
            "event_name": "string",
            "average_values": "string",
            "standard_deviation": "string",
            "head": "string",
            "tail": "string"
        },
        ...
    ]
}

Model Response Format

The model will respond with the following structure:

{
    "column": "key of the column of interest",
    "extra": "additional information deemed relevant by the model"
}

Usage Example

import bibmon
import bibmon.llm
import bibmon.llm.presets
import bibmon.llm.tools
import bibmon.three_w

# Loading the data
df, ini, class_id = bibmon.load_3w(8, "WELL-00026_20170608230000.parquet")

pre_processed = bibmon.PreProcess(
    f_pp=["remove_empty_variables", "ffill_nan", "remove_frozen_variables"]
).apply(df)

formatted_data = bibmon.three_w.tools.format_for_llm_prediction(
    pre_processed, ini, class_id, 10
)

llm_client = bibmon.llm.Client(
    "http://127.0.0.1:5000/v1",
    "NO_API_KEY",
    "Llama3",
)

res = llm_client.chat_completion_json_tool(
    *bibmon.llm.presets.three_w_find_linked_column(formatted_data)
)

parsed_result = bibmon.llm.tools.parse_chat_completion(res)

print(parsed_result)

Additional Resources

For further information and examples, please refer to our detailed notebook showcasing this feature.