Customer-language memory
Embeddings and the index are tuned around how shoppers ask questions, not around how docs are written.
Knowledge
The RAG pipeline built for production
Zowie owns the retrieval stack end to end. Embeddings, vector search, ranking, generation - all tuned for customer service content. 98% answer accuracy, every answer sourced from your content.
knowledge · 247 docs
Search knowledge…⌘K
returns / sale-items / 30-day-window.mdsourced · indexed
Sale items · 30-day window
Sale items remain returnable for 30 days from purchase if defective. Defective items qualify for replacement, exchange, or store credit.
## Eligibility
- Within 30 days of purchase
- Item must be defective
- Original packaging not required
## Resolution
- 01Replacement - same item, same shade
- 02Exchange - equal value, different size
- 03Store credit - issued same day
> Sourced from policies/returns/sale-items/30-day-window.md
> Synced to retrieval · 5 min ago
edited by Sarah · 2h ago
synced 5 min ago
98%
Answer accuracy in production, measured by Supervisor.
70+
Languages with consistent retrieval accuracy across all of them.
0
RAG infrastructure to build or maintain - Zowie owns the stack end to end.
Managed RAG
A retrieval stack we built, tuned, and run
Knowledge is a managed RAG pipeline - not a library plugged into an LLM. Every layer of the stack is built and tuned for customer service content, and the whole thing runs as a service.
Policy match before prose
Ranking narrows the answer to the right source before generation touches the customer response.
Grounded answer
Generation responds from retrieved policy and refuses to fill gaps when the source is weak.
Logged beside every layer
Each answer keeps the source, confidence, generated response, and Supervisor score attached.
Pillar 01
Maps to layers 01–03Domain-tuned retrieval
The embedding models are trained on how customers ask questions, not how engineers write docs. Ranking is tuned for short-form CS queries against policy content.
Pillar 02
Maps to layer 04Grounded generation
The agent responds from retrieved content. When retrieval confidence is low, the agent says so. No gap-filling. No plausible invention.
Pillar 03
Runs alongside the stackSource-level traceability
Every answer is logged with the retrieved policy, the generated response, and a Supervisor quality score. Full audit trail per response.
The example
One question. One policy. One sourced answer.
Knowledge retrieves the policy first, writes from that source, and keeps the evidence attached to the answer. No unsupported claim gets through.
98%Answer accuracy measured in production by Supervisor
1Approved policy used to ground the response
0Unsourced claims allowed into the answer
knowledge · retrieval · conv #71208
Hi - I bought a hoodie three weeks ago in your Berlin sale. The seam came apart yesterday. Can I exchange it?
policy retrievedscore 0.94
policies/returns/sale-items.md
Sale items remain returnable for 30 days from purchase if defective. Defective items qualify for replacement, exchange, or store credit.
You’re inside the 30-day window for defective sale items, so I can sort that for you. You can choose between a replacement, an exchange, or store credit - which works best?
Audit trailsupervisor · pass
sourcedEvery answer cites the policy it came from
loggedQuestion, retrieval, response, and score retained
scoredSupervisor evaluates every interaction automatically
Content sources
Every source your content already lives in
Knowledge ingests from the systems you already maintain. One retrieval index across all of them, with auto-sync where the source supports it - no custom ETL pipeline per system.
Help center integrations
Zendesk · Salesforce · Kustomer · Freshdesk · IntercomAuto-sync from the help center your team already maintains. Articles, categories, and revisions stream into the index.
Website content
Crawl public pagesCrawl product pages, returns pages, FAQs. Re-crawl on schedule or on demand.
API ingestion
Custom systemsPush from internal CMS, knowledge bases, or product databases through one endpoint.
Manual policies
Authored in-platformWrite policies directly in Zowie when there's nowhere upstream they belong.
Targeting
Retrieval, scoped at query time
Different customers see different answers from the same source. Knowledge filters at query time on region and segment. The retrieval index stays unified - the policy that reaches the customer is the policy that applies to them.
Segment
EU · DE
Germany
NA · EN
United States
Standard customersDefault segment
14-day return window. EU consumer rights apply. Refund to original payment within 14 days.
30-day return window. Free returns on orders over $75. Refund to original payment within 7 days.
VIP membersLoyalty tier ≥ Gold
60-day return window. Concierge handling. Full refund or replacement at customer's choice.
90-day return window. Concierge handling. Full refund or replacement at customer's choice.
Final Sale itemsItem flag · final_sale
No returns. Defective items remain returnable for 30 days under EU statutory warranty.
No returns or refunds. All sales final, except where required by state law.
Regions and Segments work together. Add a customer property, define what it filters, and every retrieval that follows scopes against it.
For CX leaders
You write the policy. The agent reads it.
The AI says what you've approved. Not what it generated. You manage the knowledge, not the agent.
Editorial control
Write a policy in plain language. Publish it. The agent reflects it on the next conversation. No retraining, no engineering ticket.
Regions and Segments
Different customers, different answers, same source of truth. VIPs in Germany see VIP DE policy; standard customers in the US see standard US policy. The matrix is yours to define.
Quality oversight
When an answer is off, Supervisor shows you which policy was retrieved and how the agent responded. Edit the policy. The next conversation runs on the new version.
In the platform
Knowledge is one of three
Knowledge handles questions. Flows handle processes that change a system of record. Playbooks handle interpretive guidance. The agent moves between all three inside one conversation.
01
Knowledge
You're here
Grounded, sourced answers from your approved content.
02
Flows
Deterministic execution
Compiled deterministic execution for refunds, claims, and identity verification.
03
Playbooks
Interpreted guidance
LLM-interpreted guidance for retention, troubleshooting, and complaints.
Connect your help center. Watch it answer.
See Knowledge grounded in your content. Connect a help center, ask the questions your team gets every day, and read the answers it gives back - with the policy each one came from.