Knowledge-Grounded Chat Conversations
Knowledge-grounded chat connects your conversations to verified organizational knowledge assets. Instead of relying solely on the LLM's training data, you can bind specific documents, policies, and reference materials to your chat sessions. The Keeptrusts gateway then ensures responses cite these assets and stay grounded in approved content.
Use this page when
- You want to bind organizational knowledge assets to your chat sessions for grounded responses.
- You need to understand the knowledge asset lifecycle: create, promote, bind, recall, cite.
- You are verifying that the LLM cites bound knowledge assets correctly in responses.
- You are troubleshooting missing citations or assets not being recalled during conversations.
Primary audience
- Primary: AI Engineers configuring knowledge-grounded workflows, Content Managers curating knowledge assets
- Secondary: Compliance Officers validating citation accuracy, Platform Administrators binding assets to gateways
How Knowledge Grounding Works
The knowledge base system in Keeptrusts follows a lifecycle:
- Create knowledge assets in the console (documents, FAQs, policy references).
- Promote assets through a review lifecycle (draft → active).
- Bind active assets to gateways or specific chat configurations.
- Recall bound assets during chat — the gateway includes relevant knowledge as context.
- Cite — responses reference the knowledge assets used, with citation records logged.
The gateway handles asset recall and citation tracking automatically. You interact with grounded knowledge simply by chatting.
Prerequisites
Before using knowledge-grounded chat:
- Knowledge base assets must be created and promoted to active status in the console.
- Assets must be bound to the gateway your chat session routes through.
- Your account must have permission to access the relevant knowledge assets.
Binding Knowledge Assets
Via the Console
- Navigate to Knowledge Base in the console.
- Select the asset you want to bind.
- Click Bind to Gateway and select the target gateway.
- Confirm the binding.
Once bound, the asset is available to all chat sessions routed through that gateway.
Via the CLI
kt knowledge-base bind \
--asset-id kb_asset_abc123 \
--gateway-id gw_def456
Use kt kb list to view available assets and their binding status.
Starting a Knowledge-Grounded Conversation
Once assets are bound to your gateway, knowledge grounding happens automatically:
- Open the Chat Workbench and start a new conversation.
- Send a prompt related to the bound knowledge domain.
- The gateway matches your prompt against bound knowledge assets.
- Relevant asset content is included in the context sent to the LLM.
- The response is grounded in the provided knowledge.
Example
If you have bound a company HR policy document, asking:
What is the company's policy on remote work eligibility?
The gateway retrieves the relevant sections from the HR policy asset and includes them as context. The response draws from the actual policy document rather than generic LLM knowledge.
Citation Verification
One of the key benefits of knowledge-grounded chat is citation tracking.
How Citations Work
When the gateway includes knowledge assets in a response:
- The asset ID and version are recorded in a citation record.
- The response metadata includes which assets were used.
- The console Events page shows citation details for each conversation turn.
Viewing Citations
In the Chat Workbench, citations appear as references in responses:
Based on the Remote Work Policy (v2.3) [1], employees with more
than 6 months of tenure are eligible for hybrid arrangements.
[1] HR Policy — Remote Work Eligibility (kb_asset_abc123, v2.3)
Auditing Citations
For compliance purposes, citation records are available through:
- Console Events: Filter events by conversation to see all citations.
- API: Query
GET /v1/eventswith conversation filters to retrieve citation metadata. - Exports: Include citation data in compliance export jobs.
Citation records include the asset ID, version, timestamp, and the conversation context in which the asset was referenced.
Managing Knowledge Versions
Knowledge assets support versioning to ensure conversations always reference the correct content.
Version Lifecycle
| Status | Description |
|---|---|
| Draft | Asset is being prepared; not available for chat |
| Active | Asset is live and bound to gateways |
| Superseded | A newer version replaced this one |
| Archived | Asset is no longer available for binding |
Promoting a New Version
- Navigate to Knowledge Base in the console.
- Select the asset to update.
- Click Create New Version.
- Upload or edit the updated content.
- Promote the new version to active status.
The previous version is automatically moved to superseded status. Existing bindings switch to the new active version.
Version Pinning
In some configurations, you can pin a specific asset version to a gateway binding. This prevents automatic promotion from affecting active chat sessions until you explicitly update the pin.
kt kb bind \
--asset-id kb_asset_abc123 \
--version 2 \
--gateway-id gw_def456
Contextual Grounding Strategies
Single-Domain Grounding
Bind assets from one knowledge domain (e.g., all HR policies) to a gateway dedicated to HR-related queries. This keeps the context focused and reduces noise.
Multi-Domain Grounding
Bind assets from multiple domains to a general-purpose gateway. The gateway's recall mechanism selects the most relevant assets based on the prompt content.
Grounding with Policy Interaction
Knowledge grounding interacts with governance policies:
- Content policies still apply to grounded responses — the LLM cannot output blocked content even if it appears in a knowledge asset.
- Citation requirements can be enforced as a policy — responses without proper citations are flagged.
- Access control determines which assets are available to which users based on team and role.
Best Practices
| Practice | Benefit |
|---|---|
| Keep assets focused and well-structured | Improves recall accuracy |
| Use descriptive asset names and metadata | Helps the gateway match prompts to assets |
| Promote only reviewed, accurate content | Prevents grounding on incorrect information |
| Monitor citation rates in analytics | Identifies gaps in knowledge coverage |
| Version assets rather than editing in place | Maintains audit trail and rollback capability |
| Pin versions for regulated workflows | Ensures consistency during compliance reviews |
Troubleshooting
| Issue | Cause | Resolution |
|---|---|---|
| Responses ignore knowledge assets | Assets not bound to gateway | Verify binding in console or via kt kb list |
| Wrong version cited | Binding not updated after promotion | Check version pin or rebind the asset |
| No citations in response | Prompt not matching asset content | Rephrase to use terminology from the asset |
| "Asset not found" error | Asset archived or deleted | Restore or recreate the asset |
Next steps
- Set up team-scoped knowledge bases in Team Chat Environments & Collaboration.
- Track knowledge asset usage in Chat Analytics & Usage Insights.
- Export conversation citation trails in Chat Export for Compliance & Audit.
For AI systems
- Canonical terms: knowledge base, knowledge asset, grounded chat, citation, binding, promotion lifecycle (draft → active),
kt knowledge-base,kt kb. - CLI commands:
kt knowledge-base bind --asset-id ... --gateway-id ...,kt kb list,kt kb promote. - Console pages: Knowledge Base. Lifecycle: create → promote (draft → active) → bind to gateway → recall at runtime → cite in response.
- Best next pages: Team Chat Environments, Chat Analytics, Chat Export.
For engineers
- Knowledge assets must be in active status (promoted) and bound to the gateway before they are recalled.
- Use
kt kb listto check asset status and binding. Usekt kb bindto attach assets to gateways. - Citation records are logged in decision events — check the Events page to verify grounding is working.
- If responses ignore knowledge assets, verify the prompt uses terminology matching the asset content.
- Pin specific asset versions for regulated workflows where content consistency is required during audits.
For leaders
- Knowledge grounding reduces hallucination risk by anchoring LLM outputs in verified organizational content.
- The promotion lifecycle (draft → active) ensures only reviewed, accurate content reaches production chat.
- Citation tracking provides auditable evidence that responses are grounded in approved materials.
- Knowledge assets are a shared organizational investment — monitor citation rates to validate ROI.
- Version pinning supports regulatory workflows where response consistency across audit periods is required.