Find ingoa wāhi

Local AI can be very useful for finding possible place names across hundreds or thousands of pages. The useful word is possible.

A language model is helping you locate candidate ingoa wāhi and organise the evidence. It is not the authority for the name, the spelling, the feature, the location or the meaning.

Ask for evidence, not just a list

A weak prompt is:

List all Māori place names in this report.

It encourages a confident list with little way to check where each item came from.

A stronger extraction request is:

Using only the supplied documents, identify candidate geographic place names.

For every candidate:
- preserve the name exactly as written in the source
- do not add or remove macrons
- do not modernise the spelling
- identify the source document
- give the page number or section if available
- include enough surrounding context to check what the name refers to
- suggest a feature type only when the text supports it
- mark uncertainty explicitly

Do not treat iwi, hapū, people, organisations or waka as places unless the source clearly also uses the same name for a geographic feature.
Do not invent coordinates.
Do not replace a historical name with a modern name.
If uncertain, keep the record and mark it uncertain rather than guessing.

Use a review table

A useful first output is:

Place as written	Possible normalised name	Type	Source	Page	Context	Confidence
source spelling	leave blank until checked	river, lake, locality, maunga, other	document title	page	short evidence note	high, medium, low

The possible normalised name field is deliberately separate. The first column is evidence. The second is interpretation.

Preserve the source spelling

Suppose an older source prints a name without a macron. Do not overwrite the source field after finding a modern official spelling. Keep both:

place_as_written = Ruamahanga River
possible_normalised_name = Ruamāhanga River

Then record why the normalised form was proposed and which source supports it.

The same rule applies to transcription errors, older orthography and English forms. Normalisation should never destroy the evidence that shows how the source actually recorded the name.

Expect false positives

Māori names occur in many roles. A model may mistake:

• an iwi name for a locality
• a hapū name for a stream or settlement
• a person's name for a place
• a marae name for the wider location
• a waka name for a geographic feature
• an organisation named after a place for the place itself

The reverse also happens. A name used for both a people and a place may be removed when it should have been retained.

This is why the source context belongs in the extraction table.

Expect omissions

The model may miss names because:

• the embedding search did not retrieve the relevant page
• OCR damaged the word
• the name occurs only once
• the name is embedded in a table, footnote or map
• a long list was truncated
• the model did not recognise an unfamiliar historical form as geographic

For important work, do not treat one AI pass as exhaustive. Run at least one direct search for obvious place terms and sample sections manually.

Useful cross-check words include river, awa, lake, roto, mount, maunga, stream, road, bay, block, pā, valley, forest, district and known names relevant to the kaupapa. These are search aids, not a complete Māori geographic vocabulary.

Macrons need their own check

Models often preserve Unicode Māori text well when the source text is clean. OCR and copy/paste pipelines are less dependable. A missing macron can originate in the historical document, OCR, PDF text layer, embedding pipeline or model output.

When spelling matters, compare the generated value with the actual source page. Do not assume the newest-looking spelling is automatically the correct form for the historical context.

Confidence is about the extraction

A confidence field should describe how confident you are that the source is referring to a geographic place, not whether the name is culturally or officially authoritative.

For example:

• high when the text explicitly says “the Ruamāhanga River” and the page is clear
• medium when the same name appears in a travel description but the feature type is implied
• low when OCR is damaged or the same term could be a people, organisation or place

Authority is a separate question handled during review.

Keep more provenance than you think you need

For a GIS-ready extraction, retain fields such as:

record_id
place_as_written
possible_normalised_name
feature_type_candidate
source_document
source_page
source_section
source_context
text_quality
ai_model
embedding_model
extraction_date
extraction_confidence
review_status
review_notes

Do not add latitude and longitude at this stage.

A useful second-pass prompt

After the first extraction, ask the model to challenge its own table:

Review the candidate place-name table against the supplied source passages.

Identify rows that may actually be:
- iwi
- hapū
- people
- organisations
- events
- document headings rather than geographic places

Do not delete them. Add a review_note explaining the uncertainty.
Also identify any row where the proposed normalised spelling is not directly supported by a cited source.

This will not catch every error, but it changes the model's role from confident extractor to assistant reviewer.

Do not ask for coordinates yet

A local model may know approximate coordinates for well-known places from its training data. That is not a suitable geographic source. Even when the numbers look plausible, a coordinate can refer to the wrong feature, a town centroid, an arbitrary search result or a completely invented location.

Move candidate names to Check the places before adding geometry.

For the wider naming practice, including official, local and historical names, see the existing Place names guide and Who gets to name a place?.

Last reviewed: 22 August 2026