A data source tells the platform where your documents live and how to read them. It is the first step in building a knowledge base, the searchable index that grounds a model in your own content. This post covers the four connector types you can choose from and the file formats the platform can index.
A data source only stores the location and credentials. Nothing is copied or indexed until you attach the data source to a knowledge base, which is where the documents get chunked and embedded.
The four data source types
When you add a data source, the Data source type dropdown offers four connectors. Pick the one that matches where your files already sit, so you do not have to move them first.
| Connector | What it connects to | When to use it |
|---|---|---|
| Google Drive Folder | A folder in Google Drive, including its files and subfolders. | When your documents are kept in Google Workspace and shared through Drive. |
| Confluence Space | A space in Atlassian Confluence, with its pages and attachments. | When your team writes and stores knowledge as Confluence wiki pages. |
| Amazon S3 Compatible Store | A bucket in any S3-compatible object store, such as AWS S3 or a self-hosted gateway. | When files live in object storage, or when you want a private, self-hosted bucket. |
| SharePoint Enterprise Site | A SharePoint site in Microsoft 365, including its document libraries. | When content sits in SharePoint or is shared through Microsoft Teams. |
Each connector asks for a few details: the location (a folder, space, bucket, or site) and credentials with read access. For the S3 connector you can also set a prefix to limit indexing to a sub-path instead of the whole bucket.
Supported file formats
Whatever connector you choose, the platform indexes the same set of file formats. Files in other formats are skipped, so it helps to know what counts before you point a source at a large folder.
| Format | What it is | Typical content |
|---|---|---|
| Portable Document Format. | Reports, manuals, guides, and product sheets. Text-based PDFs work best. | |
| .html | Web page markup. | Saved web pages, exported articles, and help-center content. |
| .md | Markdown text. | Documentation, README files, and engineering notes. |
| .pptx | PowerPoint presentation. | Slide decks, training material, and briefings. |
| .docx | Word document. | Reports, policies, letters, and proposals. |
| .txt | Plain text. | Notes, transcripts, logs, and exported records. |
| .csv | Comma-separated values. | Tabular data, exports, and simple record sets. |
A scanned PDF is an image of text, not text. Run it through an OCR step first so the words become readable, then index it. The same applies to images and audio, which need OCR or speech-to-text before they can join a knowledge base.
How a data source fits the pipeline
The flow is short and worth keeping in mind. First you register a data source with one of the four connectors. Then you create a knowledge base and attach the source to it, which kicks off chunking and embedding. After that, a chat model can retrieve the right passages from the knowledge base and answer questions grounded in your documents.
- Use a connector that matches where your files already live.
- Keep the source scoped. A folder, space, or prefix is easier to manage than an entire account.
- Give each source a least-privilege, read-only credential so a rotated key affects only that one source.
- Check that your important files are in a supported format before indexing.
Data sources are the on-ramp for retrieval. Connect the place your documents live, keep the scope tight, and confirm the formats are supported. The knowledge base then turns that content into answers.
Reference note for the Private AI series. Vendor product names belong to their respective owners.
About The Author
DrJha