Data, Privacy, and Maintenance

The knowledge base stores hosted copies of imported materials, parsed text, chunks, and retrieval indexes. Whether data leaves the local machine depends on the services chosen for parsing, OCR, embedding, reranking, and chat.

Keeping the original files local does not mean the entire knowledge base workflow is offline. If even one processing step uses a cloud service, it may send the files, snippets, or queries needed to complete the task.

What is stored after import

Content	Purpose	Update method
File copies from files or directories	For the knowledge base to continue processing and displaying	Re-add or replace after the original file is updated
Snapshots of web pages and notes	Preserve the content from the time of import	Re-import after the source is updated
Parsed main text and chunks	Preview and retrieval	Reindex after changing processors or chunking
BM25 keyword index	Exact word search	Rebuilt when reindexing
Vector index	Semantic retrieval	Generated or rebuilt after configuring or changing the embedding model

Items in the materials list are objects hosted and indexed by the knowledge base, not a real-time synchronized view of the original directory.

Which boundaries a single query may pass through

The parsing, vector, reranking, and answer nodes in the diagram may each choose local or cloud services; checking item by item is the only way to determine the data boundaries.

Selected capability	Possible content received
Cloud document processor	File content used for parsing
Cloud embedding model	Material snippets and retrieval queries
Cloud reranking model	Queries and candidate snippets
Cloud chat model	Questions, conversation context, and retrieved snippets
Corresponding local capability	Perform the corresponding processing on this machine

API keys, internal documents, and logs containing personal information should not appear in public screenshots or feedback. De-identify them before deleting or sharing.

Perform a full maintenance check

1

1. Record the current configuration

Record the knowledge base name, file processor, OCR, embedding model, reranking model, and key chunking settings. Use them to verify the environment after migration.

2

2. Clean up duplicates and old versions

Keep only the current version of each policy; if historical auditing is needed, clearly mark the year or version in the title to avoid them competing in retrieval.

3

3. Check abnormal items

Handle materials marked [Error] or stuck on [Processing] for a long time, and spot-check the main text and chunks. If indexing was not completed due to an app interruption, run [Reindex].

4

4. Create an appropriate backup

Open [Settings] → [Data]. Use a full backup that includes the knowledge base data files when migrating devices or preparing to delete materials.

5

5. Restore and verify in the target environment

Do not just confirm that restoration is complete. Check the knowledge base entries, main text, and chunks, and run fixed retrieval test questions.

6

6. Keep a rollback baseline

Validate a new processor or model on a small set of materials first, then reindex in batches. Do not delete the most recent full backup until the new results are confirmed stable.

Full backup and slim backup

Backup type	Included content	Applicable scenario	Limitations
Full backup	Data files such as chats, settings, and knowledge bases	Device migration, protection before deletion, full restore	Larger file size, takes longer
Slim backup	Mostly chat history and settings; skips data files such as images and knowledge bases	Quickly preserves commonly used settings and chats	Cannot restore a complete knowledge base on its own

A slim backup is not a safeguard for restoring a knowledge base before deletion. For important migrations, keep at least one full backup and actually run retrieval tests in the target environment.

How updates and deletions are handled

Source content updated

1. Re-add materials with the same name.
2. Choose [Replace] when you need to overwrite the old version; choose [Keep All] only when you really want them to coexist.
3. Wait for the materials to become ready.
4. Spot-check the main text and chunks.
5. Run fixed retrieval regression questions.

Processing settings updated

If you only change processor, OCR, chunking, or model settings, run [Reindex] on existing items. Changing the configuration itself will not automatically redo old materials.

Delete materials or knowledge base

Deletion removes the content and indexes hosted by the knowledge base, but does not delete the files at the original path or the original notes. Before operating, confirm that the original source is still available, that a full backup exists, and whether any agents are still bound to this knowledge base.

Configuration notes

Item	Recommended starting point	Verification method	Risk
Material version	Keep only the current version for the same purpose	Fixed questions should hit the correct version only	Mixing old and new rules
Cloud services	Confirm item by item according to sensitivity level	Check processor and model settings	Documents or snippets are sent to external services
Backup	Create a full backup before major changes	Check entries and retrieval after restore	Slim backups do not include knowledge base files
Reindex	Process representative materials in batches	Compare the same question set before and after	A one-time rebuild loses a usable baseline

User case

Xiaolin wants to migrate the team policy repository to a new computer. He first records the processor and model settings, creates a full backup, checks the main text and chunks of three materials after restoring on the new computer, and repeats the original five retrieval questions. Only after everything passes does he clean up the old environment.

The acceptance criteria are: the number of materials and titles match, key questions still hit the same source, and the team confirms that all cloud services comply with data requirements.

Fully offline checklist

• Document parsing and OCR use system, local, or self-hosted capabilities.
• The embedding model runs locally.
• Do not use cloud reranking, or use local reranking capability.
• Conversations and agents use a local chat model.
• No MCP, web search, or channels are enabled that send content to external systems.

Common questions

Will the knowledge base update automatically after modifying the original file?

No. Files, web pages, and notes are turned into materials based on the content at import time. Re-add and replace, or reindex as needed.

Is using a local embedding model completely offline?

Not necessarily. If any step in parsing, OCR, reranking, or chat uses a cloud service, it may send the content needed to complete the task.

Can a slim backup restore the knowledge base?

It cannot restore the complete knowledge base files. Use a full backup before migration or deletion, and actually verify materials and retrieval after restoration.

Continue reading

Document parsing and OCR

Understand the differences between local and cloud processors.

Add and organize materials

Replace sources and manage material versions.

Common questions

Quickly locate issues by failure level.