Key facts
| API surface | OpenAI-compatible /v1/chat/completions with streaming for chat experiences |
| Data grounding | Embeddings and RAG are live for product, policy and FAQ content |
| Order lookup | Function calling connects the assistant to order and returns APIs |
| Structured output | JSON mode returns categorized tickets, attributes and enrichment records |
| Deployment | Cloud, VPC, on-prem or air-gapped |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Free tier | Free plan with 2 free AI models; 14-day full-access trial |
| Endpoint roadmap | Images, audio, moderation, files and batch are coming soon |
TL;DR
- Deflect support with cited answers over policies, shipping rules and product data.
- Use embeddings to power semantic product search and catalog enrichment.
- Connect order systems through function calling so answers use live data.
- Serve multiple languages from one API for cross-border storefronts.
- Measure containment and repeat-contact rate, not just deflection.
How it works, step by step
- Pick one contact reason with high volume, such as where-is-my-order or returns eligibility.
- Index policies, shipping rules and product FAQs with embeddings and keep the index refreshed.
- Expose a read-only order lookup tool with customer authentication handled by your service.
- Draft answers that cite the policy or product source, and refuse when no source is retrieved.
- Run the assistant in a side panel for agents first, then open it to customers.
- Track containment, escalation and repeat-contact rates, and review weekly.
- Expand to catalog enrichment and semantic search once support quality is stable.
Try it yourself
Support deflection that actually resolves
Deflection only counts when the customer does not come back with the same issue. Build the assistant around resolution:
- Order status: fetch live status through a read-only tool rather than reciting a policy page.
- Returns and exchanges: apply the actual eligibility rules to the specific order.
- Shipping questions: answer from current carrier and delivery rules with citations.
- Product compatibility: retrieve specifications and answer fit or usage questions.
Every answer should say where it came from, and offer a human handoff when confidence is low.
Product knowledge and catalogue enrichment
The same embeddings that power support can power search and catalog work. Generate missing attributes, normalize titles and descriptions, and translate product copy into your storefront languages. Use JSON mode so enrichment output lands in a schema your PIM can validate and import.
Human merchandisers review bulk changes; the model proposes, your rules dispose.
Multilingual storefronts and seasonal peaks
Multilingual chat and embedding models in the 30+ model catalogue let one integration serve multiple locales, which helps when support coverage does not scale with traffic. Flat monthly self-serve plans keep costs stable through promotional peaks, unlike per-token billing that spikes with volume.
Load-test before a major sale and watch latency, not only correctness. Streaming responses make long answers feel faster to customers.
Measuring what matters
Track containment rate, escalation rate, repeat contact within seven days, and customer satisfaction on handled conversations. Audit a sample of transcripts weekly for policy accuracy and tone. A model or prompt change should be judged against the same numbers, with a rollback plan if they regress. Review the sample against the same rubric each week so drift is caught early.
Honest comparison
| Capability | Plugsky | Scripted chatbot | Building in-house |
|---|---|---|---|
| Policy answers | RAG with citations over live content | Static decision trees | You build ingestion and evals |
| Order lookup | Function calling into order APIs | Form-based integration | Custom orchestration |
| Product search | Embedding-based semantic retrieval | Keyword search only | Vector stack to operate |
| Languages | Multilingual chat and embedding models | Per-locale scripts | Depends on hosted model |
| Peak load costs | Flat monthly self-serve plans | Licensed seats | Fixed GPU capacity |
| Time to launch | Days | Weeks of flow design | Quarters |
Frequently asked questions
How do we keep answers accurate as policies change?
Refresh the retrieval index whenever policies, shipping rules or products change, and include the source revision in logs so a wrong answer can be traced to stale content.
Can the assistant look up a customer's order?
Yes, through a read-only tool you expose. Keep authentication and authorization in your own service, and pass only the fields required to answer the question.
What about abusive or off-topic messages?
The moderation endpoint is coming soon. Until then, apply your own input filters and hand off to human agents when a conversation leaves support scope.
Should we put the assistant in front of customers immediately?
Start with an agent-assist side panel for a week, review transcripts, then open it to customers on the highest-volume contact reasons.
Can it handle multiple languages?
Multilingual chat and embedding models handle mixed-language catalogs and support traffic. Evaluate answer quality per locale before enabling each storefront.
How do we control costs during sales peaks?
Self-serve plans are flat monthly with fair-use usage, so traffic spikes do not change your bill. Review the live pricing page for current plan details.
What is live today?
Chat, streaming, JSON mode, function calling, embeddings, RAG and agents are live. Images, audio, moderation, files, batch, assistants, responses and fine-tuning are coming soon.