vianu-fraudcrawler

vianu-fraudcrawler

Intelligent Market Monitoring

The pipeline for monitoring the market has the folling main steps:

• search for a given term using SerpAPI
• get product information using ZyteAPI
• assess relevance of the found products using an OpenAI API

Installation

python3.11 -m venv .venv
source .venv/bin/activate
pip install vianu-fraudcrawler

Usage

.env file

Make sure to create an .env file with the necessary API keys and credentials (c.f. .env.example file).

Run demo pipeline

python -m fraudcrawler.launch_demo_pipeline

Customize the pipeline

Start by initializing the client

from fraudcrawler import FraudCrawlerClient

# Initialize the client
client = FraudCrawlerClient()

For setting up the search we need 5 main objects.

search_term: str

The search term for the query (similar to search terms used within major search providers).

language: Language

The language used in SerpAPI ('hl' parameter), as well as for the optional search term enrichement (e.g. finding similar and related search terms). language=Language('German') creates an object having a language name and a language code as: Language(name='German', code='de').

location: Location

The location used in SerpAPI ('gl' parameter). location=Location('Switzerland') creates an object having a location name and a location code as Location(name='Switzerland', code='ch').

deepness: Deepness

Defines the search depth with the number of results to retrieve and optional enrichment parameters.

prompts: List[Prompt]

The list of prompts to classify a given product with (multiple) LLM calls. Each prompt object has a name, a context (used for defining the user prompt), a system_prompt (for defining the classification task), allowed_classes (a list of possible classes) and optionally default_if_missing (a default class if anything goes wrong).

from fraudcrawler import Language, Location, Deepness, Prompt
# Setup the search
search_term = "sildenafil"
language = Language(name="German")
location = Location(name="Switzerland")
deepness = Deepness(num_results=50)
prompts = [
    Prompt(
        name="relevance",
        context="This organization is interested in medical products and drugs.",
        system_prompt=(
            "You are a helpful and intelligent assistant. Your task is to classify any given product "
            "as either relevant (1) or not relevant (0), strictly based on the context and product details provided by the user. "
            "You must consider all aspects of the given context and make a binary decision accordingly. "
            "If the product aligns with the user's needs, classify it as 1 (relevant); otherwise, classify it as 0 (not relevant). "
            "Respond only with the number 1 or 0."
        ),
        allowed_classes=[0, 1],
    )
]

(Optional) Add search term enrichement. This will find related search terms (in a given language) and search for these as well.

from fraudcrawler import Enrichment
deepness.enrichment = Enrichment(
    additional_terms=5,
    additional_urls_per_term=10
)

(Optional) Add marketplaces where we explicitely want to look for (this will focus your search as the :site parameter for a google search)

from fraudcrawler import Host
marketplaces = [
    Host(name="International", domains="zavamed.com,apomeds.com"),
    Host(name="National", domains="netdoktor.ch, nobelpharma.ch"),
]

(Optional) Exclude urls (where you don't want to find products)

excluded_urls = [
    Host(name="Compendium", domains="compendium.ch"),
]

And finally run the pipeline

# Execute the pipeline
client.execute(
    search_term=search_term,
    language=language,
    location=location,
    deepness=deepness,
    prompts=prompts,
    # marketplaces=marketplaces,    # Uncomment this for using marketplaces
    # excluded_urls=excluded_urls   # Uncomment this for using excluded_urls
)

This creates a file with name pattern <search_term>_<language.code>_<location.code>_<datetime[%Y%m%d%H%M%S]>.csv inside the folder data/results/.

Once the pipeline terminated the results can be loaded and examined as follows:

df = client.load_results()
print(df.head(n=10))

If the client has been used to run multiple pipelines, an overview of the available results (for a given instance of FraudCrawlerClient) can be obtained with

client.print_available_results()

Contributing

see CONTRIBUTING.md

Async Setup

The following image provides a schematic representation of the package's async setup.