Knowledge grounding
Evaluate against your business context
Knowledge grounding lets Kits use your product docs, pricing rules, policies, ICP, and playbooks when evaluating conversations. The result is a signal that reflects your business, not generic model assumptions.
Knowledge bases
Grounded check
pricing_discussed_correctly
Evaluate the buyer's pricing question against the approved pricing policy attached to this Kit.
Result
value: "partially_correct"
Definition
Grounding keeps evaluation specific
Many conversation signals depend on company-specific meaning. Knowledge grounding gives Semarize the approved context needed to judge those signals consistently.
Product context
Use product docs, feature lists, release notes, and integration docs when a check depends on product accuracy.
Playbook context
Evaluate discovery, qualification, objection handling, and escalation against the methodology your team actually uses.
Policy context
Ground compliance, security, and legal checks in approved policy language instead of generic interpretation.
Customer context
Keep vertical, ICP, segment, and account-specific terminology available when it changes how a signal should be read.
Scoped knowledge bases
Group documents by domain or team so a Kit only uses the context that belongs to that workflow.
Signals-only output
Grounding improves evaluation, but public run output remains structured Brick results with values, reasons, confidence, and evidence.
How it works
Upload context. Attach it to Kits.
Knowledge grounding is intentionally tied to the framework layer. A Kit decides which context its Bricks should use.
Create a knowledge base
Group source material around a clear domain, such as product documentation, sales playbook, security policy, or pricing rules.
Upload source documents
Add the documents that define how the business should interpret a conversation signal.
Attach to a Kit
Link the knowledge base to the Kit that needs it, rather than attaching context loosely to every individual check.
Run grounded checks
When the Kit runs, Bricks can evaluate against the attached context and still return normal structured output.
Examples
When grounding changes the answer
Grounding matters whenever the right answer depends on internal terminology, approved process, or source-of-truth documentation.
Pricing discussion
A generic model can detect that pricing came up. A grounded Kit can evaluate whether the rep described the current pricing rules accurately.
Product fit
Grounding lets a Kit compare buyer needs to your actual product capabilities, gaps, integrations, and limitations.
Compliance QA
Policy-grounded checks can evaluate whether required language was used and whether restricted claims were avoided.
Sales methodology
A discovery Kit can evaluate against your internal methodology, not a generic definition of good discovery.
Competitor detection
Grounding can keep competitor and alternative names close to the evaluation so mentions are not missed or misclassified.
Support resolution
A support Kit can evaluate whether the answer matched the current help center article or approved escalation policy.
Why it matters
Generic context is not enough
Many mistakes are not language mistakes. They are business-context mistakes. Grounding narrows the evaluation to the context the workflow actually needs.
Without grounding
With knowledge grounding
Governance
Context without losing control
Knowledge grounding should make evaluation more accurate without leaking raw context into downstream systems or searching every document for every check.
Scoped retrieval
Knowledge bases are organised by domain so the Kit can use focused context rather than every document in the workspace.
Attached at the Kit level
KB logic belongs on Kits. That keeps each workflow's grounding explicit and avoids hidden Brick-level context drift.
Structured output remains
Grounded runs still return output.bricks with configured values, reasons, evidence, and confidence for downstream systems.
Ready to build
Use your docs. Return better signals.
Attach the knowledge your workflow depends on, then keep the API output structured and predictable.
No card required. Start testing in minutes.
Questions
Frequently asked questions
Short answers about knowledge bases, grounded Kits, and how context changes conversation evaluation.
From the blog
Further reading
Practical pieces on why conversation evaluation needs product, pricing, policy, and ICP context to produce trustworthy signals.