Upload and Manage Documents
Add local or S3 documents to a knowledge base, monitor parsing and indexing, and manage document chunks.
After a document enters a knowledge base, UnieAI Studio stores, parses, chunks, and indexes it in the background. Knowledge-base search and Customize Model RAG can retrieve the document only after its chunks have been indexed.
Permissions and prerequisites
Only the knowledge-base creator and co-editors can upload, re-index, modify, or delete documents. Members with use access can view and use the knowledge base but cannot maintain its content.
The system also checks required models before upload:
- An administrator must configure the KB Chunking task model. File upload is disabled without it.
- The user must have access to the platform's configured VLM. If it is not in My Models, the page prompts the user to add it.
- Uploading an
.mp4video also requires access to the configured speech-to-text model. - Audio parsing requires an available platform speech-to-text service.
When a required model is missing, follow the page prompt to add it to My Models. Platform task-model configuration must be completed by an administrator in Admin Panel.
Supported formats
| Type | Direct-upload extensions | Processing |
|---|---|---|
| Documents | .pdf, .md, .docx, .txt | Reads text according to the format; PDFs, Word files, and layout-rich content may use visual and paged parsing |
| Tables | .csv, .xlsx | Chunks rows; you can choose the number of rows per chunk before upload |
| Presentations | .pptx | Converts slide content and sends it through the active parsing pipeline |
| Images | .jpg, .jpeg, .gif, .png, .jp2, .tiff, .webp, .bmp | Uses a VLM to interpret text, charts, and visual content |
| Video | .mp4 | Extracts video frames and transcribes audio |
| Audio | .mp3, .wav, .m4a, .aac, .ogg, .flac | Transcribes speech into text that can be chunked and indexed |
The knowledge-base page does not currently accept .hwp, legacy .doc, legacy .ppt, or video formats other than .mp4 for direct upload.
Upload files from your computer
Choose a personal space or organization in the sidebar, open My knowledge base, and select the target knowledge base.
Select Add new → Upload files, then choose one or more supported files. You can also drag files directly onto the knowledge-base page.
For every CSV or Excel file, configure Rows per chunk. The minimum is 1; leave it empty to use the platform default.
Review the upload queue and select Start. Choosing files adds them to the queue but does not upload them immediately.
The queue shows transfer progress for each file. Select Stop to abort the active upload queue, or Clear before starting to remove queued files. Large files automatically use multipart upload and require no additional configuration.
File and storage limits
Limits come from platform and user quotas. The page shows storage available to the current knowledge base.
| Limit | Behavior |
|---|---|
| Per-file size | A file over the current per-file limit is not added to the queue |
| Total knowledge-base storage | The batch is blocked if existing documents plus queued files exceed the total limit |
| Maximum PDF pages | A PDF over the platform page limit is blocked; if the browser cannot determine the count, the server checks it again |
| Audio size | Audio files are also subject to the platform audio-upload limit |
These are not fixed 50 MB or 200 MB limits. An administrator must change the corresponding user or platform quota when a different limit is required.
Sync from S3
If the platform has enabled Data Hub connectors, select Add new → Sync from S3 to ingest a larger document collection.
Select an existing S3 connection, or create a new AWS or S3-compatible storage connection.
Choose a folder in the bucket, or select the bucket root.
Choose what happens when an object is deleted from the source: mark it removed in the knowledge base, or permanently delete it there as well.
Create the binding and select Sync now. The page reports scanned, ingested, deduplicated, failed, skipped, and deleted counts.
Unchanged content is deduplicated by content hash. Added or changed unstructured documents enter the knowledge-base parsing pipeline; structured or unsupported objects may be reported as skipped. Open the object details to inspect the outcome for each source object.
If the page reports that connectors are disabled, ask an administrator to enable Data Hub connectors. Removing an S3 binding is not the same as deleting the knowledge base.
Parsing pipelines
The system uses a pipeline appropriate to each source type rather than applying the same OCR process to every document:
- Text documents are read and normalized.
- CSV and Excel files are chunked by the configured row count.
- PDFs and images can use page rendering, VLM parsing, figure detection, and figure captioning.
- Office content such as PowerPoint may be converted to a parser-friendly format first.
- Audio is transcribed; video frames are extracted and its audio is transcribed.
- Parsed content is chunked, optionally contextualized, and written to the retrieval index.
Creators and co-editors can choose a pipeline under Settings → Parsing:
| Mode | Description |
|---|---|
| Default (platform setting) | Inherits the administrator's default pipeline |
| Concurrent mode | Uses the concurrent document-processing flow |
| Page mode | Splits PDF or image pages into resumable jobs; better suited to large or many-page documents |
The same page includes Create an independent Chunk for each Figure, which is off by default. When enabled, every detected Figure remains available in the page content and also receives its own searchable chunk. This is useful when users need to retrieve one image directly by its title, caption, or visible content.
Pipeline and Figure Chunk changes apply to new uploads and re-indexed documents. They do not automatically alter documents that have already finished parsing. If the first PDF page-preview render fails, the system automatically retries with a lighter render. A page is marked failed only after every attempt fails, preventing an empty “image unavailable” result from being accepted as a completed page.
Monitor processing
| Status group | Common states | Meaning |
|---|---|---|
| Waiting | Pending, Uploaded, Queued | The document exists and is waiting for a worker |
| Processing | Running, Processing, Parsing | Downloading, rendering, transcribing, parsing, chunking, or indexing is in progress |
| Usable | Indexed, Ready, Succeeded, Ready Partial | Indexing completed; partial readiness may mean some pages did not succeed |
| Failed | Failed, Error | A processing stage failed and needs investigation |
| Cancelled | Cancelled | A user cancelled active processing |
An active document shows its stage, percentage, page progress, estimated token count, and elapsed time. Select the status to open the live event stream and processing timeline. Creators and co-editors can also select Cancel while processing is active.
If the live connection drops, on-screen updates pause while the page attempts to reconnect. The background job has not necessarily stopped.
View and edit chunks
After indexing completes, hover over the document row and select View chunks. The Chunks page displays text, page, image, figure, or media source details according to the document type.
Creators and co-editors can:
- Edit chunk text and select Save.
- Expand Metadata to inspect sources, page numbers, and chunking strategy.
- Open History to see versions created by manual edits and restores.
- Preview an older version and select Restore this version.
- Delete eligible figure chunks.
Every manual edit and restore preserves a chunk version. Chunk versions are separate from whole-knowledge-base snapshots used for Marketplace review.
Change CSV or Excel chunking
CSV and Excel document rows include Tabular chunk settings:
- Enter a new number of rows per chunk.
- Save the setting.
- If the document is already indexed, the system queues rechunking. Otherwise, it only saves the setting.
A smaller row count creates more focused retrieval units. A larger value preserves more adjacent data but also puts more content into each result.
Parsing failures and re-indexing
For a failed document, select its failed status or information icon to inspect:
- The status log and failed stage
- Parser errors and processing summary
- Missing VLM access, a model missing from My Models, or an unconfigured platform task model
- Quota, file size, page count, corrupt file, or unsupported-format errors
After fixing the problem, select the re-index icon on the document row. Re-indexing processes the original file already stored by the platform and rebuilds its index. It does not read a newer copy from your computer.
If the source content changed, upload the new file. A same-name upload does not overwrite the old document: the system creates a unique display name such as (1) or (2) and records the files as one document series. CM search uses only the newest indexed document in that series. While the new version is still processing, or if it fails, the previous available version remains searchable. Older documents stay in the list for review or deletion.
Delete a document
Hover over the document row, select the delete icon, and confirm. Deletion removes the document record, parsed content, retrieval vectors, and stored platform object. It cannot be undone. If a model still needs the content, upload it again or resync it from S3.