Getting started

Natural-language filtering is one of the headline features of PFD Toolkit. The Screener class lets you describe a topic in plain English – e.g. "deaths in police custody" – and have an LLM screen reports, delivering you a curated dataset.

To use the Screener you'll first need to set up an LLM client if you haven't already.

---

A minimal example

First, import the necessary modules, load reports and set up an LLM client:

from pfd_toolkit import load_reports, LLM, Screener

# Grab all reports from 2024
reports = load_reports(start_date="2023-01-01",
                       end_date="2023-12-31")

# Set up your LLM client
llm_client = LLM(api_key=YOUR-API-KEY)

Then define a user_query which describes the reports you're interested in. Pass this query as an argument to screen_reports() and you'll be given a filtered dataset containing only reports which the LLM judged to have matched your query.

user_query = "Deaths in police custody **only**."

screener = Screener(
    llm=llm_client,
    reports=reports
)

police_df = screener.screen_reports(
    user_query=user_query)

police_df will now only contain reports related to your query.

---

Why not just have a "normal" search function?

A keyword search is only as good as the exact words you type. Coroners, however, don't always follow a shared vocabulary. The same idea can surface in wildly different forms:

• Under-staffing might be written as "staff shortages," "inadequate nurse cover," or even "resource constraints."
• Suicide in prisons may masquerade as "self-inflicted injury while remanded," "ligature event in cell," or may not even appear together in the same sentence.

A keyword filter misses these variants unless you identify every synonym in advance. By contrast, an LLM understands the context behind your query and links the phrasing for you, which is exactly what Screener taps into.