> For a complete page index, fetch https://docs.synthflow.ai/llms.txt. For full documentation content, fetch https://docs.synthflow.ai/llms-full.txt.
# Knowledge Base
> Create knowledge bases, add PDFs, documents, URLs, run website crawls, import from Zendesk, attach bases to agents, and inspect knowledge base searches in call logs.
A knowledge base holds documents and pages your agents can query during calls using retrieval-augmented generation (RAG). This page covers how to create a knowledge base, describe when the agent should search it, add sources (including [scanning URLs](#scan-urls) and [connecting external knowledge](#connect-external-knowledge-bases) such as Zendesk), attach a base to an agent, and review searches in [call logs](/logs#actions-tab).
## How to create a knowledge base
In the dashboard, open **Knowledge Bases** and create a new knowledge base. Give it a clear **name**, then define **when the agent should search the knowledge base**.
That condition should describe the situations or user questions that should trigger a lookup—for example, when the caller asks how to use your product or requests policy details. **Do not** use this field for general prompting (such as objection-handling or tone instructions); it only controls *when* a search runs, not how the agent speaks.
**Best practices:**
* Be explicit about the scenarios that should trigger a search.
* Keep each source focused on a single topic; split large PDFs when it helps retrieval.
* Review and update content regularly so agents do not rely on outdated material.
After you save the knowledge base, add sources (see below), then attach it to an agent.
From the knowledge base screen, add any combination of:
* **PDFs** from your device.
* **Blank documents** to type or paste content when you add the source.
* **URLs** to [fetch and analyze](#scan-urls) your webpage.
* **Integrations** to import from a [connected service.](#connect-external-knowledge-bases)
## Scan websites \[#scan-urls]
Here you’ll bring public web pages into a knowledge base: each page is fetched, analyzed, and stored for retrieval. Stop at one URL (**single page**) or use **Website crawl** with a root URL and crawl settings to add linked pages too.
### How to scan multiple pages
From your knowledge base, click **Add item** → **Paste from URL** → **Website Crawl**. Use the table below to understand what each field is for and how it can help you.
Once ready to scan, the import runs in the background and may take a few minutes depending on site size and your crawl settings; when it finishes, open that source to review imported pages and extracted text.
| Field or setting | What it is for |
| :-------------------------- | :-------------------------------------------------------------------------------------------------------------------------- |
| **Name** | Internal label for this crawl source inside your knowledge base. |
| **Root URL** | The URL Synthflow starts from when discovering pages (for example, `https://docs.synthflow.ai`). |
| **Max Pages (limit)** | Maximum number of pages to import from this crawl. |
| **Sitemap Handling** | Whether sitemap URLs are included with link discovery, skipped, or used as the only URL list. |
| **Sync Schedule** | How often Synthflow automatically re-runs the crawl to refresh imported content. |
| **Include / exclude paths** | Path rules to pull in or block parts of the site (for example excluding `/blog` or `/changelog`). |
| **URL parameters** | How query strings on URLs are handled, to avoid importing many near-duplicate pages. |
| **Subdomains** | Whether the crawler may follow links to other subdomains of the same site. |
| **Crawl depth** | How many link levels from the root URL to follow; lower values keep the crawl narrow, higher values cover more of the site. |
### Best practices
* Start with a lower page limit while testing, then increase gradually.
* Use include/exclude rules to avoid irrelevant sections (for example `/blog` or `/changelog`).
* Keep URL-parameter ignoring enabled to reduce duplicate page imports.
* Use a lower crawl depth for focused imports and higher depth for full-site coverage.
## Connect external knowledge bases \[#connect-external-knowledge-bases]
Synthflow can pull content into a knowledge base from **supported integrations**. Available integrations:
* Zendesk
### Zendesk \[#import-from-zendesk]
Use the Zendesk integration to import your Help Center articles directly into a knowledge base. This is useful when your support documentation already lives in Zendesk and you want your agent to answer from the same source of truth.
The import runs asynchronously. Depending on the number of Zendesk articles, syncing can take a few minutes.
In your knowledge base, choose **Add item** → **Import from Integration**, or start from the **Import from Integration** tile on the empty state. Select **Zendesk**, then sign in with OAuth using your **`Client ID`**, **`Client Secret`**, and **`Subdomain`**.
Once the connection succeeds, pick how often to sync (**Daily**, **Weekly**, or **Monthly**) and click **Import resources**. The source shows **In Progress** while Help Center articles are indexed, then **Synced** when the run finishes.
The Zendesk integration crawls your Zendesk Help Center articles and imports them into your knowledge base. Once synced, your agent can retrieve this content during RAG search.
## Attach to an agent
An agent **only** runs knowledge base searches when a **knowledge base is attached to that agent**. Bases you create elsewhere in the workspace are not used until you link them.
To attach one, open the agent in the [agent editor](/the-agent-editor), go to **Knowledge & Memory**, and choose the knowledge base you want. Save your changes, then publish a [new version](/version-control) so that agent can retrieve from it on live calls.
## Track activity in call logs
When an agent looks up your knowledge base during a call, open your [call logs](/logs#actions-tab) to see the lookup.
Some relevant information you may find interesting:
* Which knowledge base was used
* The situation you set for when the agent should search
* How the caller’s wording compares to the search the agent actually ran
* How many matches came back and how long the lookup took
* Whether the lookup succeeded
Together, this helps you confirm a search ran when you expected and spot problems early.
## FAQ
The knowledge base stores information the agent can reference during a conversation. It suits large or changing content you do not want to paste into the prompt. At answer time the agent searches for relevant snippets; it does not load the entire knowledge base into context at once.
Retrieval-Augmented Generation (RAG) lets the model use your uploaded or linked content when generating a reply, instead of relying only on built-in training data.
The agent matches the conversation to content in the knowledge base using semantic search over the material you added.
Usually the agent paraphrases. If a passage is clear and well structured, it may repeat short phrases exactly when appropriate.
If conflicting information exists across multiple documents (e.g., different figures in different files), the agent's response may be inconsistent and depend on which document it selects.
To prevent this, you can:
* Remove or update outdated content on a regular basis.
* Ensure consistency across documents.
* Test agent responses to verify accuracy.
You can add Knowledge Base content by uploading PDFs, creating documents directly, pasting a single URL, running a website crawl, or importing from supported integrations (for example, Zendesk). See [Scan URLs](#scan-urls) and [Connect external knowledge bases](#connect-external-knowledge-bases) on this page.
Yes, you can mix sources such as PDFs, single-page URLs, website crawls, integration imports, and documents created in Synthflow. The agent will search through all available content regardless of source.
`Paste from URL` can import a **single page** from one URL, or you can switch to **Website crawl** to index multiple pages—see [Scan URLs](#scan-urls) on this page.
Complex layouts (columns, heavy images, unusual formatting) may not extract cleanly. If uploads fail or look wrong, try a simpler PDF or plain text.
The agent retrieves relevant sections rather than reading a file end to end. Clear headings and smaller, topic-focused files improve results.
You cannot manually rank documents. Improve outcomes with clearer writing, tighter scope per file, and removing noise.
Open the call in your [call logs](/logs#actions-tab) and select the **Actions** tab. If the agent ran a lookup, you will see an action for that knowledge base search (often titled like your knowledge base). If nothing appears there, no knowledge base search was recorded for that call.
Refine the search condition, improve source structure and headings, and test alternate caller phrasing. Use [call logs](/logs#actions-tab) to confirm whether a search ran and what query was used.