Abstract

The following exercise attempts to use LLM to perform information extraction. The webpage https://techcrunch.com/2023/10/27/x-is-launching-new-premium-and-basic-subscription-tiers/ was used as the context to retrieve basic information such as company involved, their urls and topic of the article into a structured, JSON format.

Methodology

Setup

1. Model: TheBloke/Llama-2-13b-Chat-Q4_K_M_GGUF (4-bit quantization, medium sized for balanced quality vs computational intensity)
2. BeautifulSoup - for HTML content parsing
3. LangChain - for inference pipeline and output parsing

##Approach

1. Extract content of the webpage via BSHTMLLoader from LangChain
2. Extract hrefs (URLs) from the HTML using BeautifulSoup
3. Get a list of companies mentioned within the article via Prompt Engineering
4. Get a JSON mapping of companies to company domain through Prompt Engineering by passing in hrefs extracted and the list of companies from step 3.
5. Get the topic of the article as a JSON structure via Prompt Engineering
6. Combine the company JSON and topic JSON into one and parse it as a Pydantic BaseModel Object.

Results

{
"related_companies": [
      { "company_name":"X", "company_domain":"x.com" },
      {"company_name":"Bloomberg", "company_domain":"bloomberg.com" }
   ], 
  "topic":"X is launching two new subscription tiers, including a ‘Premium+’ ad-free plan" 
}