# Knowledge Base (/docs/sift-ai/knowledge-base)

Sift allows you to bring your own knowledge into the platform by adding data sources like FAQs, help documentation, and product pages. The Knowledge Base is where you manage the contexts that power SiftGPT's responses and your AI-driven workflows.

## At a glance [#at-a-glance]

| | |
| --- | --- |
| Organize sources into named Contexts | ✅ |
| Add sources by website crawl, file URL, or upload | ✅ |
| Automatic crawling and indexing (website sources) | ✅ |
| Include / Exclude path patterns to scope a website crawl | ✅ |
| Status tracking per source in the Data Sources table | ✅ |
| Re-crawl on demand to refresh content | ✅ |
| Add article titles, audiences, topics, and matching phrases | ✅ |
| Contexts reused in Chat, AI Responses, and Auto-reply Bots | ✅ |

## Creating a context [#creating-a-context]

A **Context** is a named collection of sources - create one first (name only) before you can add any sources to it. Once a context exists, it shows a **Data Sources** panel with an **Add Source** button.

<figure>
  <img src="/assets/Screenshot 2026-04-12 at 1.13.41 AM.png" alt="The Knowledge Base with a list of indexed data sources" />
  <figcaption>The Knowledge Base manages the contexts that power SiftGPT and AI workflows.</figcaption>
</figure>

## Adding a source [#adding-a-source]

Click **"Add Source"** inside a context and pick one of three methods:

- **Website**: crawl a site by root URL.
  - **Name**: A descriptive label for the source (e.g., "Help Center", "Product FAQ").
  - **Source URL**: The root URL of the website or documentation you want Sift to index.
  - **Include Paths** _(optional)_ - Specify URL patterns to include, one per line. Only pages matching these patterns will be crawled.
  - **Exclude Paths** _(optional)_ - Specify URL patterns to skip, one per line. Pages matching these patterns will be excluded from indexing.
- **File URL**: point at a single hosted file (e.g. a PDF) to index directly.
- **Upload**: upload one or more files from your machine to index directly. One upload holds up to 10 files of 25 MB each. Each file becomes its own source, named after the file.

## Indexing and re-crawling [#indexing-and-re-crawling]

<figure>
  <img src="/assets/Screenshot 2026-04-12 at 1.14.34 AM.png" alt="The Add Source form with name, source URL, and include and exclude path fields" width="375" />
  <figcaption>Add a source by entering its name, URL, and optional include/exclude paths.</figcaption>
</figure>

Once added, Sift automatically crawls and indexes the content (website sources) or processes the file directly (File URL / Upload). The Data Sources table shows each source's name, URL, and current status - so you can track when indexing is complete and content is ready to use. Click **"Re-crawl"** anytime to refresh a website source with updated content.

## Where contexts are used [#where-contexts-are-used]

Contexts created here are reusable across the platform - in SiftGPT Chat for grounded responses, in Auto-Reply Bots for answering customer questions, and in AI response workflows for accurate, on-brand automated replies.

## Improve article matching [#improve-article-matching]

Open a website source and find an indexed page in **Page inventory**. Select **Improve matching** to add customer-managed details.

- Use **Article title** for the name that your team searches for.
- Use **Audiences** for the people who should use the article. Examples include `customers`, `drivers`, or `admins`.
- Use **Topics** for a few stable subjects. Examples include `refunds`, `billing`, or `account access`.
- Use **Matching phrases** for real customer wording. Add one phrase per line.

Use a small number of precise labels. Do not copy the full article into these fields. A matching phrase can help Sift find an article. It cannot create a strong match unless the article title or body also supports the customer request.

Sift marks articles with direct support as **Strong match**. It also shows **Related article** when a source may help but needs review. Open a related article and check its content before you use it in a reply. Sift does not show known weak matches.


## Related [#related]

<Cards>
  <Card href="/docs/sift-ai/chat" title="Chat">Ask questions grounded in your indexed content.</Card>
  <Card href="/docs/sift-ai/ai-responses" title="AI Responses">Draft replies that pull from your knowledge base.</Card>
</Cards>