LLM Data Handle

LLM Handler

We have implemented a simple LLM mechanism composed of three parts: Prompt, LLM, and Output Parser. We recommend using Langchain [Introduction | 🦜️🔗 LangChain] to initialize the LLM, which provides a unified interface. Here is a simple example:

from langchain_openai import OpenAI
from rllm.llm import PromptTemplate, LangChainLLM, BaseOutputParser

template = "Please write five random names related to {topic}."
prompt = PromptTemplate(template=template)

class my_parser(BaseOutputParser):
    def parse(self, output: str):
        return output
output_parser = my_parser()

llm = LangChainLLM(OpenAI(openai_api_key="YOUR_API_KEY"), output_parser=output_parser)

output = llm.predict(prompt, topic='dogs')
print(output)

We have also implemented two utility classes: Predictor and Enhancer. The former can annotate data with pseudo-labels, while the latter can generate explanations for the data to obtain information or embed the additional information. The data needs to be organized in the form of a pandas.DataFrame.

Here’s an example of using Predictor:

import pandas as pd

from langchain_openai import OpenAI
from rllm.llm import LangChainLLM, BaseOutputParser, Predictor

# Generate some data
data = {
    "EmployeeID": [123, 456],
    "FirstName": ["John", "Jane"],
    "LastName": ["Doe", "Smith"],
    "BirthDate": ["1980-01-01", "1990-02-02"],
    "Salary": [70000, 80000]
}
df = pd.DataFrame(data)

class my_parser(BaseOutputParser):
    def parse(self, output: str):
        start = output.find("ANSWER:") + len("ANSWER:")
        first_word_after_answer = output[start:].strip().split()[0]
        return first_word_after_answer
output_parser = my_parser()

llm = LangChainLLM(OpenAI(openai_api_key="YOUR_API_KEY"),
                output_parser=output_parser)
predictor = Predictor(llm=llm, type='classification')

# Two parameters should be provided: `scenario` and `labels`
output = predictor(df, scenario='career classification', labels='doctor, engineer')
print(output)

Below is an example demonstrating the usage of Enhancer:

import pandas as pd

from langchain_openai import OpenAI
from rllm.llm import LangChainLLM, BaseOutputParser, Enhancer

data = {
    "title": ["The Shawshank Redemption", "Farewell My Concubine"],
    "year": ["1994", "1993"],
    "director": ["Frank Darabont", "Kaige Chen"]
}
df = pd.DataFrame(data)

llm = LangChainLLM(OpenAI(openai_api_key="YOUR_API_KEY"),
                output_parser=output_parser)
enhancer = Enhancer(llm=llm, type='explanation')

# Parameter `scenario` should be provided!
# If you want to get embedding, you should provide a list-like
# parameter `cols` to identify which columns should be embedded.
output = enhancer(df, scenario='movie explanation')
print(output)

Enhancer

In this section, we will demonstrate how to use the enhancer to augment table information with textual enhancements and encode it into vectors.

First, it is necessary to initialize the large language models for interpreting the tables and performing the encoding:

from langchain_openai import OpenAI, OpenAIEmbeddings
from rllm.llm import LangChainLLM

llm = LangChainLLM(OpenAI(openai_api_key="YOUR_API_KEY"))
llm_embed = LangChainLLM(OpenAIEmbeddings(openai_api_key="YOUR_API_KEY"))

Next, initialize the enhancer instance:

import pandas as pd
from rllm.llm import Enhancer

data = pd.read_csv('data.csv')
scenario = 'Your_task_description'
enhancer = Enhancer(llm=llm, llm_embed=llm_embed, type='explanation|embedding')

Finally, the data can be passed to the enhancer to quickly obtain interpreted and encoded semantic vectors of the tabular data.

import pandas as pd

data = pd.read_csv('data.csv')
scenario = 'Your_task_description'
outputs = enhancer(data.head(10), scenario=scenario)