Week 4 Lab: StatsChat Overview

2026-04-29

Week 4 lab

StatsChat Overview

Task

Choose one realistic question that you would want a StatsChat-like system to answer in your own organisation.

Today’s lab goal

By the end, you should have one realistic example of:

  • a question a StatsChat-like system could answer
  • the documents it would need

You should also be able to explain:

  • what happens once before users ask questions
  • what happens each time a user asks
  • what you would check before trusting the answer

Quick recap

Two halves of StatsChat

Worked example

Kenya inflation question

In March 2026, what was Kenya’s annual inflation rate, and what drove it?

A good answer would need to check:

Definition

Annual CPI inflation?

Date

March 2026 vs March 2025?

Caveats

What does “drove it” mean?

Your turn

Choose one realistic question

Pick a question from your own organisation or NSO context.

Good examples often ask about:

  • a recent figure or trend
  • a definition or methodology
  • a comparison across time or groups
  • caveats behind a published statistic

Short and realistic is better than ambitious and vague.

Example questions

  • What was the latest figure for ____?
  • How is ____ defined?
  • What changed between [period A] and [period B]?
  • What caveats does the report mention about [topic]?
  • Which official report supports [claim]?

Quick check

Which task happens once before users ask questions?

A. Retrieve relevant chunks for this question
B. Convert source documents into searchable text
C. Generate the answer
D. Decide whether the answer is trustworthy

Suggested answer: B — document preparation is part of the background setup.

Fill in the split

Prepare once

  • What documents are needed?
  • How would they be collected?
  • How would they be processed?
  • What metadata matters?

Each question

  • What does the user ask?
  • What evidence should be retrieved?
  • What should the LLM answer?
  • What references should be shown?

Trust check

Before trusting the answer, what would you inspect?

Definition

Is the concept right?

Date

Is the period right?

Source

Is it official?

Caveats

What might be missing?

Share back

One-minute report out

For your example, share:

  1. your question
  2. the key documents needed
  3. one step that happens once
  4. one step that happens each time
  5. one thing you would check before trusting the answer

Optional extension

General AI search vs StatsChat

For your question:

  • What might a general AI search summary do well?
  • What might be risky about relying on it?
  • What would be better about grounding the answer in official sources?
  • What would still be difficult for StatsChat?

Components of StatsChat

PDF Processing Answering Questions
Done before users ask questions. Done each time a user asks.
  • PDF processing - scraping PDFs, converting PDFs to JSON, chunking and embedding.
  • Answering questions - retrieving relevant evidence and generating an answer with references.