LLMTextualAnswer

Python package for finding textual answers via LLMs. This is a Python port of the Wolfram Language LLMTextualAnswer function, focused on building prompts, wiring LangChain models, and parsing structured outputs.

Install

pip install LLMTextualAnswer

Usage

Question answering

from LLMTextualAnswer import LLMTextualAnswer
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini")

text = (
    "Born and raised in the Austrian Empire, Tesla studied engineering and physics "
    "in the 1870s without receiving a degree."
)

questions = ["Where born?"]

result = LLMTextualAnswer(
    text,
    questions,
    llm=llm,
    form=dict,
)

print(result)

Classification

Here is a list of workflow construction specifications:

queries = [
    'Make a classifier with the method RandomForest over the data dfTitanic; show precision and accuracy; plot True Positive Rate vs Positive Predictive Value.',
    'Make a recommender over the data frame dfOrders. Give the top 5 recommendations for the profile year:2022, type:Clothing, and status:Unpaid',
    'Create an LSA object over the text collection aAbstracts; extract 40 topics; show statistical thesaurus for "notebook", "equation", "changes", and "prediction"',
    'Compute quantile regression for dfTS with interpolation order 3 and knots 12 for the probabilities 0.2, 0.4, and 0.9.'
]

Here are possible workflows names:

workflows = ['Classification', 'Latent Semantic Analysis', 'Quantile Regression', 'Recommendations']

For each workflow spec give the corresponding (most likely) workflow name:

for q in queries:
    print("Spec  : " + q)
    print("Class : " + llm_classify(q, workflows, llm = llm, form=dict) + "\n")

# Spec  : Make a classifier with the method RandomForest over the data dfTitanic; show precision and accuracy; plot True Positive Rate vs Positive Predictive Value.
# Class : Classification
# 
# Spec  : Make a recommender over the data frame dfOrders. Give the top 5 recommendations for the profile year:2022, type:Clothing, and status:Unpaid
# Class : Recommendations
# 
# Spec  : Create an LSA object over the text collection aAbstracts; extract 40 topics; show statistical thesaurus for "notebook", "equation", "changes", and "prediction"
# Class : Latent Semantic Analysis
# 
# Spec  : Compute quantile regression for dfTS with interpolation order 3 and knots 12 for the probabilities 0.2, 0.4, and 0.9.
# Class : Quantile Regression
# 

Notes

• For more detailed examples see the notebook "Basic-usage.ipynb.
• llm_textual_answer and llm_classify accept "LangChain" chat/text models that support .invoke.
• Use prompt_style="chat" or prompt_style="text" if auto-detection is not desired.
• When you want only the prompt template, pass form="StringTemplate".