How do I ground AI answers in actual company knowledge to prevent hallucinated product or policy facts?
By SensoPublished Last updated 5 min read
Ground your own AI tools by connecting the model to an approved knowledge base, having it answer only from the passages it retrieves, and checking each claim in the answer against those passages before it reaches a user. Grounding reduces hallucinated product and policy facts but does not remove them on its own, which is why the claim-level check matters. For public AI models you do not control, publish accurate pages written from the same knowledge base so they can find them when they search the web.
Grounding is the practice of making an AI answer come from specific, trusted sources instead of from the model's memory. Retrieval-augmented generation (RAG) is the most common way to do it: the system first retrieves relevant passages from your knowledge base, then the model writes the answer from them. Senso by Zeta, from Zeta Global, is a context layer for AI agents: one knowledge base that your agents answer from and that public pages are written from.
How does grounding differ for your own tools and for public models?
The input, mechanism and limit are different for each system.
| System | Input | How grounding works | Known limit |
|---|---|---|---|
| Your own AI tools, such as a support bot or internal agent | An approved knowledge base | The model answers only from retrieved passages, ideally with a source for each claim | The model can still misread the passages or add claims they do not contain |
| Public AI models you do not control | Your published pages | When the model searches the web, your accurate pages are available as a source about you | You do not control whether the model finds or uses them |
Is retrieval-augmented generation enough to prevent hallucinations?
No. Retrieval gives the model the right passages, but the model can still misread them or add claims they do not contain. Add a check that compares each claim in the answer against the retrieved source and flags anything unsupported before it reaches a user.
A practical test, recommended here rather than a published standard, has three steps with a clear input and output for each.
| Step | Input | Output |
|---|---|---|
| 1. Ask a known question | A question whose answer is in your knowledge base | The model's answer and the passages it retrieved |
| 2. Trace each claim | The answer and the passages | Each claim matched to a passage, or marked unsupported |
| 3. Record the gaps | The unsupported claims | A list of where the model went beyond its sources |
What mistakes let hallucinated facts through a grounded system?
Most of them come from trusting retrieval without checking the answer.
| Mistake | Why it fails | Do this instead |
|---|---|---|
| Assuming retrieval prevents hallucination | The model can still go beyond the passages | Check each claim against the retrieved source |
| Leaving old versions in the knowledge base | Retrieval can return an outdated policy | Remove or mark old versions so they are not retrieved |
| Returning answers without sources | Nobody can check where a claim came from | Return the source passage with each answer |
| Grounding only internal tools | Public models still answer from the web | Publish accurate pages from the same knowledge base |
How does Senso by Zeta ground AI answers in company knowledge?
Senso by Zeta exposes grounding through the Senso CLI, so the same knowledge base serves your agents and your public pages.
| Capability | Command | What it returns |
|---|---|---|
| Grounded answers with sources | senso search | An answer plus the passages it used, each with the document (content_id) and version (version_id) |
| Passages only, for your own model | senso search context | Matching passages with no answer, up to 20 per query |
| Claim-level checks | senso evals text --evaluator kb_accuracy | Each claim in the text you supply, checked against your knowledge base with a verdict |
| Adding documents | senso kb create-raw or senso kb upload | New documents that can be searched within seconds |
What do the Senso CLI commands look like?
These examples come from the Senso documentation. Add --output json when a script or agent reads the result.
# Ask a question and get an answer with its sources
senso search "How long do customers have to ask for a refund?" --output json
# Get matching passages only, to feed your own model
senso search context "mortgage rates"
# Check text claim by claim against your knowledge base
senso evals text --evaluator kb_accuracy --wait \
--text "Customers can get a full refund within 60 days of purchase."
# Add files to your knowledge base
senso kb upload <files...> --folder-id <id>
The JSON output of senso search includes answer, a results list where each result has content_id, version_id, title, chunk_text and score, and total_results. In the CLI, a fact-check returns each claim as verified, conflict or unsupported, which the Senso by Zeta app shows as confirmed, contradicted or not in your Context Layer. To check an existing pipeline, run its answers through senso evals text.
What results have companies seen with Senso by Zeta?
FundMore.ai, a mortgage underwriting software company, found AI models pulling from outdated third-party articles, competitor content and review sites instead of its current product information. Using Senso, it launched a dedicated domain, fundmore.finance, on April 8, 2026. By April 20, 2026, that domain was one of the top 10 most-cited external sources across its tracked questions, with 75 citations and a 60% average share of voice (FundMore.ai case study).
Sources: Senso documentation, CLI reference. Senso documentation, Why Senso. Senso documentation, Verification Loop.
Next step: Read Why Senso in the Senso by Zeta documentation.
Published by Senso by Zeta, the GEO platform from Zeta Global.
