Search your own documents

The useful part of local document AI is not asking a model to remember everything in a 500-page report. It is giving the application a collection of documents, letting its search index find the most relevant passages, and then asking the language model to answer from those passages.

This is often called RAG, or retrieval-augmented generation. You do not need the acronym to use it. Think of it as find the relevant evidence first, then ask the model to work with that evidence.

Before working with Māori or organisational information, read Before you use AI with Māori information. Local processing reduces some disclosure risks but does not replace Māori data governance, organisational approvals, privacy, security or source-context decisions.

Use public material for the first technical run

For a first workspace, choose two or three substantial public reports that contain overlapping places or topics. Waitangi Tribunal reports, council reports, environmental assessments and public planning material can work well because there is enough text to make retrieval useful and the source can be checked independently.

This recommendation is about testing the software with material that can be independently verified. It does not mean that everything inside a public document is culturally unrestricted or appropriate to extract, combine, geocode or republish. Public sources can still contain whakapapa-linked kōrero, sensitive locations, disputed evidence, approximate geography or information provided for a particular kaupapa.

If the exercise moves beyond technical testing into substantive Māori research, reconsider the source collection, authority, intended use and outputs. Keep the original context with extracted material and use the wider Māori Data Sovereignty for GIS guidance.

Do not begin with a folder containing hundreds of files. A small collection makes it easier to see whether the system is retrieving the right evidence.

Create a workspace

In AnythingLLM Desktop:

1. Create a new workspace with a clear name, such as Wairarapa public reports test.
2. Add the PDF files to the workspace document collection.
3. Allow the application to parse and embed the documents.
4. Check that each file appears as available to the workspace rather than merely attached to one chat.
5. Keep the original PDFs in a normal source folder outside the application as well.

Indexing creates derived data from the documents. In a local setup that index should be stored on the same computer. See Install local AI for the AnythingLLM storage folders.

Ask a question that has a checkable answer

Start with something narrow enough that you can verify it manually. For example:

Using only the documents in this workspace, find passages that discuss Lake Wairarapa and access to, use of, or relationships with the lake.

For each relevant passage return:
- document title
- page number if available
- a short description of what the passage says
- the source wording needed to verify your summary

Do not add facts that are not in the supplied documents. If the evidence is not present, say not found.

The exact question is less important than the discipline. Tell the model what evidence it may use and require a source reference.

Ask across several documents

Once a single-document question works, ask the same topic across the whole workspace:

Compare how the supplied documents refer to the Ruamāhanga River.

List each document separately. Preserve Māori spelling as written in the source. Identify any different spellings or naming forms. Give a page or section reference for every finding. Do not decide which spelling is authoritative.

This is where document retrieval becomes more useful than an ordinary PDF search. The model can group related passages and give you a structured starting point, while you still retain the source references needed to check its work.

Open the source

Do not stop at the generated answer. Open the cited page or passage and read enough surrounding text to understand what the reference actually means.

A place name beside a discussion of land loss, whakapapa or customary use can have a very different meaning from the same name in a modern planning table. Retrieval can separate a sentence from the context that gives it significance.

For research-grade extraction, record the source context in your output table rather than only the name.

Use follow-up questions carefully

Useful follow-ups include:

Which of those references describe a physical place rather than an iwi, hapū, organisation or person?

Show me the evidence for row 4. Do not use your general knowledge.

Which candidate place names have uncertain spelling or could refer to more than one feature?

Return the results as CSV fields, but do not add coordinates.

The follow-up should narrow or test the evidence. It should not gradually encourage the model to fill missing information from memory.

Recognise retrieval failure

A fluent answer can hide a poor search result. Warning signs include:

• the cited passage does not support the answer
• the application returns the same few chunks for unrelated questions
• a name visible in the PDF cannot be found by the model
• page references are missing or inconsistent
• tables are flattened into confusing text
• the answer contains details that do not appear anywhere in the source collection
• scanned pages produce empty or broken text

When this happens, inspect the extracted document text. If the text is missing, fix the PDF or OCR before changing the prompt. If the text is present but retrieval misses it, try a more direct term, re-index after changing the embedding model, or reduce the document collection while testing.

Open WebUI's current RAG troubleshooting guidance makes the same basic point: many apparent model failures are actually extraction or context problems.

Search by Māori text as well as English

For mixed English and te reo Māori documents, test retrieval with several forms of the same concept. A multilingual embedding model can help find related text across languages, but it is still worth asking direct name and phrase searches when precision matters.

Do not assume that “100+ languages” in a model card means strong performance for every form of te reo Māori. Low-frequency names, historical spellings and OCR-damaged macrons remain difficult.

Keep a small evidence log

For each useful extraction, retain at least:

Field	Example purpose
source_document	identifies the report
source_page	lets another person reopen the evidence
source_text	preserves the wording used for the extraction
ai_task	records what you asked the model to do
model	records the local model used
embedding_model	records how retrieval was created
review_status	separates unreviewed output from checked information
reviewer_notes	records corrections or uncertainty

This is more useful than preserving a huge chat transcript with no clear relationship to the final data.

Next, see Working with PDFs before attempting large historical collections, then Find ingoa wāhi.

Sources

• AnythingLLM – using documents in chat
• AnythingLLM – desktop storage
• LM Studio – offline document chat
• Open WebUI – Retrieval Augmented Generation
• Open WebUI – RAG troubleshooting
• Te Mana Raraunga – Principles of Māori Data Sovereignty
• Te Kāhui Raraunga – Māori AI Governance

Last verified: 2 September 2026