Week 5 Lab: PDF ingestion

2026-05-06

Lab focus

From Tuesday’s exercise

Today we will use the exercise to practise thinking about PDF ingestion from two angles:

  • What would I want to ingest for my organisation?
  • What information survives extraction?
  • What structure gets lost?
  • What would retrieval see?
  • What would a better output preserve?

Part 1

Your own source

Discuss in pairs or small groups:

  1. What publication or data source did you choose?
  2. Who would use it?
  3. What questions might they ask?
  4. What would be important to preserve?

This is the adoption question: what would need to change for your own organisation?

Part 2

CPI extraction example

Original CPI March 2026 PDF table showing the Total row and percentage change columns

Simple extracted text

Total
100.0000
0.5
4.4

The value is present, but its meaning depends on the surrounding table headings, units and report context.

Part 2 discussion

CPI extraction example

Questions

  1. What is clear from the extracted text?
  2. What depends on the table heading or surrounding context?
  3. Is this enough to answer a simple question?
  4. What structure would make it safer?

Original CPI March 2026 PDF table crop used for discussion

Extracted text

Total
100.0000
0.5
4.4

Part 3

Visual factsheet extraction

Original ICT factsheet showing mobile phone ownership and internet usage indicators in a visual layout

Simple extracted text

Mobile Phone Ownership by Individuals
National
Male
Female
53.7%
54.5%
52.9%
Internet Usage by Individuals
National
Male
35.0%
37.8%
Female
32.2%

Part 3 discussion

Visual factsheet extraction

Questions

  1. What survives?
  2. What grouping is unclear?
  3. What could retrieval or generation misunderstand?

Original ICT factsheet visual layout used for discussion

Extracted text

Mobile Phone Ownership by Individuals
National / Male / Female
53.7% / 54.5% / 52.9%

Internet Usage by Individuals
National / Male / Female
35.0% / 37.8% / 32.2%

A better target

{
  "chunk_type": "table_or_visual_section",
  "title": "Mobile Phone Ownership by Individuals",
  "groups": ["National", "Male", "Female"],
  "values": ["53.7%", "54.5%", "52.9%"],
  "page": 1,
  "source_url": "...",
  "notes": "Values extracted from a factsheet section"
}

The exact schema is less important than the principle: keep labels, values, units and source context together.

Group discussion

Where does the failure happen?

Original CPI March 2026 PDF table crop for group discussion

Extracted text

Total
100.0000
0.5
4.4

Use the example to walk backwards through the pipeline.

Group discussion prompts

Where does the failure happen?

For the CPI example, ask:

  • Was the source collected?
  • Was the text extracted?
  • Was the structure preserved?
  • Would the chunk be meaningful?
  • Would retrieval find it?
  • Would the LLM interpret it correctly?

Original CPI March 2026 PDF table crop for group discussion

Total
100.0000
0.5
4.4

Wrap-up

Main idea

A good ingestion pipeline should make evidence easier to retrieve, verify and cite.

For your own organisation, the first question is not only:

Can we extract the text?

It is also:

What structure and metadata do we need to preserve so the answer can be trusted?