Workflow first
Choose a repeated task with clear review criteria before choosing a default model.
Use cases
Practical patterns for shipping Chinese model APIs through SmarToken
This page is written as an implementation guide, not a list of invented customer wins. Use it to decide which workflow to test, which model route to start with and what quality evidence to collect before sending real production traffic.
One key per surface
Separate keys for bots, batch jobs, IDE tools and staging so spend is easier to audit.
Evidence before scale
Compare real examples, not only demos, and keep failed outputs in the evaluation set.
Implementation patterns
Four places teams usually start
The safest first project is usually a narrow workflow with a human review loop and a clear definition of a useful answer.
Support volume is rising and the team needs a lower-cost first answer before escalation.
Customer support assistant
SaaS teams with repeated tickets, help-center articles and multilingual users.
Start with a retrieval step that sends the model only the article, product state and customer message needed for one answer. Use a budget-limited API key for the bot service, keep the key server-side and log request IDs with ticket IDs.
Route starting points
- - Qwen or Kimi for multilingual answers
- - DeepSeek for reasoning-heavy troubleshooting
- - A smaller route for classification and intent labels
Quality check
Measure answer helpfulness, citation accuracy, escalation rate and cases where the assistant should have refused to answer. Review failed conversations weekly before increasing automation.
Guardrail
Do not let the model guess account-specific billing, refunds or security status. Route those questions to human support.
The team wants model choice beyond a single vendor while keeping OpenAI-compatible client code.
Developer coding assistant
Engineering teams testing Chinese coding and reasoning models inside IDEs, CLIs or internal review tools.
Give each tool a separate API key, set a daily budget, and record model, latency and token cost beside repository or issue metadata. Keep prompts short enough that diffs and logs stay reviewable.
Route starting points
- - DeepSeek for coding value
- - Kimi for long repository context
- - Qwen for structured explanations and multilingual docs
Quality check
Track accepted suggestions, reverted suggestions, test pass rate after generated patches and the number of follow-up prompts needed for a useful answer.
Guardrail
Never paste private credentials or production logs into prompts. Use redaction before sending stack traces or customer data.
Human graders need a first pass that is consistent enough to edit, not a final automated grade.
Education feedback workflow
Edtech products that need draft feedback, rubric checks or language-learning comments at scale.
Represent the rubric as structured prompt context, ask for short reasons for each score, and keep the final grade decision in the product workflow. Use one model for feedback and another route for moderation or policy checks when needed.
Route starting points
- - Qwen for multilingual classroom feedback
- - Kimi for long essays
- - DeepSeek for reasoning through rubric criteria
Quality check
Sample graded work by course, language and difficulty. Compare consistency against the rubric and audit for overly confident or overly generic feedback.
Guardrail
Avoid presenting model feedback as final academic judgment without review, appeal or teacher override.
Teams have many small documents and lose time asking the same internal questions.
Internal knowledge assistant
Operations, finance or product teams searching policies, release notes and runbooks.
Use retrieval with document permissions, send source snippets with clear titles, and ask the model to return answer, source, confidence and next action. Separate read-only question answering from any workflow that changes records.
Route starting points
- - Kimi for longer source packets
- - Qwen for structured answers
- - DeepSeek for analysis of incident timelines
Quality check
Review no-answer rate, source coverage, hallucinated citations and whether the assistant correctly says when a document is missing.
Guardrail
Permission boundaries matter more than model choice. Do not retrieve documents the requesting user should not see.
Decision guide
Pick a first route by workload
A model route should be chosen because it helps a specific job, not because it is the newest name in the catalog.
| Situation | Starting point | Validation note |
|---|---|---|
| Need lowest-cost reasoning | Start with DeepSeek | Run the same prompt set against Qwen and Kimi before production. |
| Need long document reading | Start with Kimi | Chunk very long files and measure token spend per useful answer. |
| Need broad multilingual product behavior | Start with Qwen | Check tone, structured output and language switching with real user examples. |
| Need predictable spend | Use budget-limited keys | Set separate keys for staging, batch jobs and customer-facing traffic. |
Questions to answer before production
Are these verified customer case studies?
No. This page is a practical implementation guide for common SmarToken use cases. It avoids fabricated customer names or performance claims and focuses on architecture, quality checks and rollout steps.
Which use case should a new team test first?
Start with a narrow workflow that already has examples and review criteria: one support intent, one coding task class, one rubric or one internal document collection.
How should teams compare model routes?
Use the same prompts, expected outputs and scoring rubric across models. Compare quality, latency, cost and failure modes before changing the default route.
What should be measured before scaling traffic?
Measure useful answer rate, cost per accepted output, latency, refusal behavior, escalation rate and incidents where the model used missing or outdated context.