Auto Taxonomy Creation

Build a site taxonomy from a list of keywords, provided via CSV file upload, or by connecting to a Google Search Console property. Uses OpenAI for taxonomy creation.

Demo

Try out a working Streamlit demo on Huggingface here.

Proudly open sourced by Locomotive Agency.

Installation

pip install git+https://github.com/locomotive-agency/taxonomyml.git

Usage

Example with CSV

from taxonomyml import create_taxonomy

filename = "domain_data.csv"
brand_terms = ['brand', 'brand', 'brand']
website_subject = "This website is about X"

taxonomy, df, samples = create_taxonomy(
    filename,
    text_column="keyword",
    search_volume_column="search_volume",
    website_subject=website_subject,
    cross_encoded=True,
    min_df=5,
    brand_terms=brand_terms,
    openai_api_key="sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
)

df.to_csv("taxonomy.csv", index=False)

Example with Search Console (GSC)

import os
from taxonomyml import create_taxonomy
from taxonomyml.lib import gsc, gauth
os.environ["OPENAI_API_KEY"] = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

auth_manager = gauth.GoogleServiceAccManager(
    scopes=["https://www.googleapis.com/auth/webmasters.readonly"],
    credentials_path="service_account.json",
)
gsc_client = gsc.GoogleSearchConsole(auth_manager=auth_manager)

brand_terms = ["brand"]
website_subject = "This website is about X"
prop = "https://www.example.com/"

taxonomy, df, samples = create_taxonomy(
    prop,
    gsc_client=gsc_client,
    days=30,
    website_subject=website_subject,
    min_df=2,
    brand_terms=brand_terms,
    limit_queries_per_page=5
)

df.to_csv("domain_taxonomy.csv", index=False)

Example with GSC (Service Account with Subject)

import os
from taxonomyml import create_taxonomy
from taxonomyml.lib import gsc, gauth
os.environ["OPENAI_API_KEY"] = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

auth_manager = gauth.GoogleServiceAccManager(
    scopes=["https://www.googleapis.com/auth/webmasters.readonly"],
    credentials_path="service_account.json",
    subject="emailuser@domain.com",
)
gsc_client = gsc.GoogleSearchConsole(auth_manager=auth_manager)

brand_terms = ["brand"]
website_subject = "This website is about X"
prop = "https://www.example.com/"

taxonomy, df, samples = create_taxonomy(
    prop,
    gsc_client=gsc_client,
    days=30,
    website_subject=website_subject,
    brand_terms=brand_terms,
)

df.to_csv("domain_taxonomy.csv", index=False)

Parameters

Important

These are the most important parameters. If you are using a CSV, search_volume_column and text_column are required.

• data (Union[str, pd.DataFrame]): GSC Property, CSV Filename, or pandas dataframe.
• text_column (str, optional): Name of the column with the queries. Defaults to None.
• search_volume_column (str, optional): Name of the column with the search volume. Defaults to None.
• website_subject (str, required): Provides GPT with context about the website.
• openai_api_key (str | None, optional): OpenAI API key. Defaults to environment variable OPENAI_API_KEY if not provided.
• brand_terms (List[str], optional): List of brand terms to remove from queries. Defaults to None.
• days (int, optional): Number of days back to pull data from Search Console. Defaults to 30.

Taxonomy Fine-tuning

These parameters allow you to fine-tune the selection of topics sent to OpenAI.

• ngram_range (tuple, optional): Ngram range to use for scoring. Defaults to (1, 5).
• min_df (int, optional): Minimum document frequency to use for scoring. Defaults to 5.
• limit_queries_per_page (int, optional): Number of queries per page to use for clustering. Defaults to 5.
• debug_responses (bool, optional): Whether to print debug responses. Defaults to False.

Matching Fine-tuning

These parameters control the matching back of taxonomy categories to your original data.

• cluster_embeddings_model (Union[str, None], optional): Name of the cluster embeddings model. Defaults to 'local'.
• cross_encoded (bool, optional): Whether to use cross encoded matching. Defaults to False. Improves matching, but takes longer.
• percentile_threshold (int, optional): The percentile threshold to use for good matches. Defaults to 50.
• std_dev_threshold (float, optional): The standard deviation threshold to use for good matches. Defaults to 0.1.

Example of ClusterTopics

from taxonomyml.lib.clustering import ClusterTopics
model = ClusterTopics()

corpus = ['The sky is blue and beautiful.', 'Love this blue and beautiful sky!', 'The quick brown fox jumps over the lazy dog.']
categories = ['weather', 'sports', 'news']
model.fit_pairwise(corpus, categories)

2023-08-02 06:51:39.209 | INFO     | lib.clustering:fit_pairwise:615 - Similarity threshold: 0.06378791481256485
2023-08-02 06:51:39.211 | INFO     | lib.clustering:fit_pairwise:621 - Most similar: 0
2023-08-02 06:51:39.212 | INFO     | lib.clustering:fit_pairwise:623 - Most similar score: 0.407695472240448
2023-08-02 06:51:39.213 | INFO     | lib.clustering:fit_pairwise:625 - Standard deviation: 0.16589634120464325
2023-08-02 06:51:39.214 | INFO     | lib.clustering:fit_pairwise:621 - Most similar: 0
2023-08-02 06:51:39.215 | INFO     | lib.clustering:fit_pairwise:623 - Most similar score: 0.37827810645103455
2023-08-02 06:51:39.216 | INFO     | lib.clustering:fit_pairwise:625 - Standard deviation: 0.15577822923660278
2023-08-02 06:51:39.218 | INFO     | lib.clustering:fit_pairwise:621 - Most similar: 1
2023-08-02 06:51:39.219 | INFO     | lib.clustering:fit_pairwise:623 - Most similar score: 0.0687832236289978
2023-08-02 06:51:39.220 | INFO     | lib.clustering:fit_pairwise:625 - Standard deviation: 0.0209036972373724

>>> ([0, 0, -1], ['weather', 'weather', '<outlier>'])