> ## Documentation Index
> Fetch the complete documentation index at: https://docs.periskope.app/llms.txt
> Use this file to discover all available pages before exploring further.
# Knowledge Base
> Train your AI Agent with FAQs, documents, and self-learned context for accurate, reliable responses
The Knowledge Base is where you teach your AI Agent what to know and how to answer customer questions. Access it through **AI** > **Knowledge Base**.
Every entry — FAQs, self-learned Q\&A pairs, and documents — lives in one list. Use the tabs above the list to filter by type:
| Tab | Shows |
| - | - |
| **All** | Everything in the knowledge base |
| **FAQ** | Entries you wrote manually |
| **Self-Learned** | Q\&A pairs the AI discovered automatically from real conversations |
| **Documents** | Uploaded PDF reference material |
Next to the search bar, a second row of filters — **All**, **Active**, **Inactive**, **Needs Review** — narrows the list by status, with a live count next to each. **Needs Review** is where self-learned entries land while they're waiting for approval.
Use **Add New** to create a **FAQ** or upload a **Document (PDF)**. To have the AI read from a live system instead, see [Connectors](/ai/agent/connectors). Select one or more entries to reveal a bulk action bar for activating, deactivating, or deleting them at once.
The **AI Training** button opens a small settings panel with two switches:
| Setting | What it controls |
| - | - |
| **AI Training** | Whether the AI auto-learns from team responses in conversations (the weekly [Self-Training](/ai/agent/self-training) run) |
| **Require Approval** | Whether self-learned entries need review before the AI can use them |
***
## FAQs
FAQs are essentially **prompts for the AI Agent**. Click **Add New** > **FAQ** (or open an existing entry) to get the full-page editor. Each FAQ entry has three fields:
* **Question**: The customer query — add multiple variations so the AI can match different phrasings
* **Answer**: The response the AI should provide, with an optional file/link attachment
* **Instructions**: Special behavioral guidance for this specific FAQ (optional)
The editor also shows the entry's status (Active/Inactive/Needs Review) with a switch to toggle it, and prev/next arrows to move through the rest of the list without going back to the Knowledge Base page.
### Writing Effective Questions
Add multiple variations of the same question so the AI can identify it regardless of how customers phrase it:
```text theme={null}
How much does your service cost?
What are your pricing plans?
Can you tell me about your subscription fees?
How much does it cost per month?
```
The more variations you add, the more reliably the AI will match incoming questions to the right FAQ entry.
### Writing Effective Answers
Keep answers concise but complete. Use bullet points for multi-step or multi-item responses. Include links where relevant.
```text theme={null}
Our service has three tiers:
- Basic: $10/month — includes core features
- Pro: $25/month — includes advanced analytics and integrations
- Enterprise: Custom pricing — contact sales@company.com
```
### Using Instructions for Controlled Responses
The **Instructions** field lets you add behavioral rules specific to this FAQ — without affecting the rest of the AI's behavior.
**Basic instructions:**
```text theme={null}
Ask for the user's account email before providing specific account details.
```
**Conditional instructions:**
```text theme={null}
If the user mentions an error code, ask for a screenshot before troubleshooting.
If the user asks about pricing, confirm their region first.
```
**Multi-turn / escalation instructions:**
```text theme={null}
If they are facing issues adding a new user and appear to have hit their user limit,
explain that they may need to upgrade their plan. Ask if that's the case, and if so,
direct them to Settings > Manage Plans or notify the sales team to step in.
```
FAQs are evaluated as prompts by the AI model. Writing clear, specific instructions produces far better results than vague guidance. If an FAQ isn't working as expected, try being more explicit in the Instructions field.
### FAQ Best Practices
* **Include multiple question variations** — different phrasings, abbreviations, and common misspellings
* **Keep answers factual** — the AI will not fabricate information, but vague answers lead to vague responses
* **Use instructions for edge cases** — handle follow-up logic, escalations, and conditional behavior here
* **Review regularly** — add new FAQs for recurring questions you see in conversations
***
## Documents
Documents let you upload large bodies of reference content that the AI searches when answering questions. This is ideal for:
* Product documentation
* Company policies and procedures
* Pricing sheets
* Onboarding guides
* Technical specifications
### Recommended Document Formats
**1. Q\&A Format (Most Effective)**
Structure content as question-answer pairs. This directly mirrors how the AI processes queries and produces the most reliable results.
```text theme={null}
Q: How do I reset my password?
Ans: Go to the login page and click "Forgot Password". Enter your email address
and you'll receive a reset link within 5 minutes.
Q: What payment methods do you accept?
Ans: We accept Visa, Mastercard, American Express, PayPal, and bank transfers
for enterprise accounts.
```
**2. Paragraph Structure**
Group related information into focused paragraphs with clear headings:
```text theme={null}
[Product Features]
Our platform includes real-time analytics, custom reporting, and team
collaboration tools. All features are accessible from the dashboard.
[Security]
We implement industry-standard encryption for all data. Servers are hosted
in ISO-certified data centers with 24/7 monitoring and SOC 2 compliance.
```
**3. List Structure**
Use bullet points for step-by-step instructions or feature lists:
```text theme={null}
[Getting Started]
- Sign up with your email address
- Verify your email through the confirmation link
- Complete your profile setup
- Connect your first integration
- Invite team members
```
The AI searches documents for relevant sections when answering questions. Well-structured documents with clear headings and labeled sections produce better results than unformatted walls of text.
***
## Self-Learned Context
The AI Agent automatically learns from real conversations. Every Sunday, [Self-Training](/ai/agent/self-training) reviews your team's resolved chats from the past week, extracts the answers they gave, and drafts new FAQ entries here — you can also trigger a run manually for any week from the Self-Training page. Filter the Knowledge Base list to the **Self-Learned** tab to see only these entries.
This means your knowledge base improves over time without manual effort — every good answer your team gives becomes training data.
### Require Approval
Open the **AI Training** panel above the list and toggle **Require Approval** to control what happens to newly self-learned entries:
| Require Approval | Behavior |
| - | - |
| **On** | New self-learned Q\&A pairs are held out with a **Needs Review** status until you approve them — they aren't used by the AI yet |
| **Off** | New self-learned Q\&A pairs go **Active** immediately and are used by the AI right away |
Turning it on is recommended when:
* You want tight quality control over what the AI learns
* Your team's responses may include information not appropriate for all customers
* You're in an early stage of AI deployment
Turning it off suits mature deployments where your team consistently gives accurate, appropriate answers.
To review pending items, filter the list to **Needs Review**, open an entry, and use its status switch to approve (activate) or leave it inactive. You can also select multiple entries and use the bulk **Activate**/**Deactivate**/**Delete** actions.
***
## Knowledge Base Optimization Tips
**Quality over quantity.** Start with a small set of high-quality FAQs covering your most common questions. A well-written FAQ with good instructions outperforms ten vague ones.
**Test with real questions.** After adding FAQs, test them by sending actual customer questions through the AI and reviewing the responses in [Logs](/ai/agent/logs).
**Fill gaps over time.** Monitor AI conversations to identify questions the AI is handling poorly or escalating too often. Add or refine FAQs for those topics.
**Avoid conflicting content.** If multiple FAQ entries or documents cover the same topic with different answers, the AI may give inconsistent responses. Keep your knowledge base clean.
**Review self-learned entries regularly.** Self-learned context is valuable but requires curation. Schedule a regular review of pending entries to keep your knowledge base growing accurately.
**Let the Org Assistant do the audit.** Instead of scanning entries manually, open the [Org Assistant](/ai/agent/ai-assistant#keep-the-knowledge-base-up-to-date) on this page and ask it to find gaps, flag outdated articles, or spot duplicates — it searches your actual FAQs, documents, and self-learned entries and reports back.
## Next Steps
* [**Personalization**](/ai/agent/personalization) - Define the AI's role, personality, and instructions
* [**Built-in Tools**](/ai/agent/built-in-tools) - Configure what actions the AI can take
* [**Custom Tools**](/ai/agent/custom-tools) - Connect the AI to your own APIs
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.
Every entry — FAQs, self-learned Q\&A pairs, and documents — lives in one list. Use the tabs above the list to filter by type:
| Tab | Shows |
| - | - |
| **All** | Everything in the knowledge base |
| **FAQ** | Entries you wrote manually |
| **Self-Learned** | Q\&A pairs the AI discovered automatically from real conversations |
| **Documents** | Uploaded PDF reference material |
Next to the search bar, a second row of filters — **All**, **Active**, **Inactive**, **Needs Review** — narrows the list by status, with a live count next to each. **Needs Review** is where self-learned entries land while they're waiting for approval.
Use **Add New** to create a **FAQ** or upload a **Document (PDF)**. To have the AI read from a live system instead, see [Connectors](/ai/agent/connectors). Select one or more entries to reveal a bulk action bar for activating, deactivating, or deleting them at once.
The **AI Training** button opens a small settings panel with two switches:
| Setting | What it controls |
| - | - |
| **AI Training** | Whether the AI auto-learns from team responses in conversations (the weekly [Self-Training](/ai/agent/self-training) run) |
| **Require Approval** | Whether self-learned entries need review before the AI can use them |
***
## FAQs
FAQs are essentially **prompts for the AI Agent**. Click **Add New** > **FAQ** (or open an existing entry) to get the full-page editor. Each FAQ entry has three fields:
* **Question**: The customer query — add multiple variations so the AI can match different phrasings
* **Answer**: The response the AI should provide, with an optional file/link attachment
* **Instructions**: Special behavioral guidance for this specific FAQ (optional)
The editor also shows the entry's status (Active/Inactive/Needs Review) with a switch to toggle it, and prev/next arrows to move through the rest of the list without going back to the Knowledge Base page.
### Writing Effective Questions
Add multiple variations of the same question so the AI can identify it regardless of how customers phrase it:
```text theme={null}
How much does your service cost?
What are your pricing plans?
Can you tell me about your subscription fees?
How much does it cost per month?
```
The more variations you add, the more reliably the AI will match incoming questions to the right FAQ entry.
### Writing Effective Answers
Keep answers concise but complete. Use bullet points for multi-step or multi-item responses. Include links where relevant.
```text theme={null}
Our service has three tiers:
- Basic: $10/month — includes core features
- Pro: $25/month — includes advanced analytics and integrations
- Enterprise: Custom pricing — contact sales@company.com
```
### Using Instructions for Controlled Responses
The **Instructions** field lets you add behavioral rules specific to this FAQ — without affecting the rest of the AI's behavior.
**Basic instructions:**
```text theme={null}
Ask for the user's account email before providing specific account details.
```
**Conditional instructions:**
```text theme={null}
If the user mentions an error code, ask for a screenshot before troubleshooting.
If the user asks about pricing, confirm their region first.
```
**Multi-turn / escalation instructions:**
```text theme={null}
If they are facing issues adding a new user and appear to have hit their user limit,
explain that they may need to upgrade their plan. Ask if that's the case, and if so,
direct them to Settings > Manage Plans or notify the sales team to step in.
```
The AI Agent automatically learns from real conversations. Every Sunday, [Self-Training](/ai/agent/self-training) reviews your team's resolved chats from the past week, extracts the answers they gave, and drafts new FAQ entries here — you can also trigger a run manually for any week from the Self-Training page. Filter the Knowledge Base list to the **Self-Learned** tab to see only these entries.
This means your knowledge base improves over time without manual effort — every good answer your team gives becomes training data.
### Require Approval
Open the **AI Training** panel above the list and toggle **Require Approval** to control what happens to newly self-learned entries:
| Require Approval | Behavior |
| - | - |
| **On** | New self-learned Q\&A pairs are held out with a **Needs Review** status until you approve them — they aren't used by the AI yet |
| **Off** | New self-learned Q\&A pairs go **Active** immediately and are used by the AI right away |
Turning it on is recommended when:
* You want tight quality control over what the AI learns
* Your team's responses may include information not appropriate for all customers
* You're in an early stage of AI deployment
Turning it off suits mature deployments where your team consistently gives accurate, appropriate answers.
To review pending items, filter the list to **Needs Review**, open an entry, and use its status switch to approve (activate) or leave it inactive. You can also select multiple entries and use the bulk **Activate**/**Deactivate**/**Delete** actions.
***
## Knowledge Base Optimization Tips
**Quality over quantity.** Start with a small set of high-quality FAQs covering your most common questions. A well-written FAQ with good instructions outperforms ten vague ones.
**Test with real questions.** After adding FAQs, test them by sending actual customer questions through the AI and reviewing the responses in [Logs](/ai/agent/logs).
**Fill gaps over time.** Monitor AI conversations to identify questions the AI is handling poorly or escalating too often. Add or refine FAQs for those topics.
**Avoid conflicting content.** If multiple FAQ entries or documents cover the same topic with different answers, the AI may give inconsistent responses. Keep your knowledge base clean.
**Review self-learned entries regularly.** Self-learned context is valuable but requires curation. Schedule a regular review of pending entries to keep your knowledge base growing accurately.
**Let the Org Assistant do the audit.** Instead of scanning entries manually, open the [Org Assistant](/ai/agent/ai-assistant#keep-the-knowledge-base-up-to-date) on this page and ask it to find gaps, flag outdated articles, or spot duplicates — it searches your actual FAQs, documents, and self-learned entries and reports back.
## Next Steps
* [**Personalization**](/ai/agent/personalization) - Define the AI's role, personality, and instructions
* [**Built-in Tools**](/ai/agent/built-in-tools) - Configure what actions the AI can take
* [**Custom Tools**](/ai/agent/custom-tools) - Connect the AI to your own APIs
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.