vs-product-qa

name	vs-product-qa
description	Answer Viking AI Search product questions, CLI usage questions, API/auth questions, configuration questions, and troubleshooting questions by grounding every claim in either the installed `vs` CLI's own output or official Volcengine documentation. Never fabricate.
category	shared
applies_to	codex, agents, external-agent
requires_cli	>=0.2.0
keywords	viking ai search, product question, product concept, concept, how to, usage, api, authentication, ak sk, configuration, error, troubleshooting, docs, official docs, help, faq, data count, effective data, valid data, data volume, item-data-count
commands	doctor, auth status, llm status, skill list, skill search, skill show

vs-product-qa

When to Use

Use this skill when the user asks grounded questions about Viking AI Search, including:

• product concepts such as scene, dataset, hybrid search, ANN, rerank, or agentic search
• capability or configuration questions such as recommendation setup, boost/bury, or rerank enablement
• API and authentication questions such as AK/SK usage, field semantics, or error meaning
• CLI usage and local troubleshooting such as vs item apply, vs auth status, or local stack traces
• console path, billing, pricing, or quota questions

Do not use this skill for workflow execution. Delegate instead:

• sign-up, purchase, payment, or first AK/SK setup -> vs-user-onboarding
• multi-modal dataset creation and ingestion (records with image/video URLs + text) -> vs-item-onboarding
• web crawling for content ingestion -> vs-crawler
• search tuning suggestions or execution -> vs-search-tuning

If another skill is active and the user asks a product question outside that skill's scripted scope, answer with vs-product-qa, then return to the original workflow after the answer is complete.

Preconditions

• The installed vs CLI is available when answering CLI behavior, local auth, LLM configuration, or local error questions.
• This skill includes a bundled internal documentation helper under references/volcengine-documentation/SKILL.md.
• Official Volcengine documentation must be fetched through that bundled helper when answering product concepts, API field semantics, purchase, billing, quota, or console UI path questions.
• The agent must classify the question before choosing a source.
• If no selected source can answer, explicitly say the point is not covered by the CLI output or official documentation checked in this turn, then suggest support-ticket / oncall escalation.

Core Principle

Treat the user's installed vs CLI as authoritative for how vs behaves on that machine. For CLI behavior, always check the real CLI first. If CLI output conflicts with documentation, trust the CLI and state that the documentation may be stale.

This skill must not rely on repository source code, generated snapshots, or model memory as the source of truth for customer-facing product answers.

Question Routing

Classify the question before choosing a source.

Question type	Primary source	Example
CLI command / flag / usage	installed CLI output: vs <cmd> --help; use the API Reference router only when the user asks how that command maps to an API category	"How do I use vs item apply"
Local authentication / credential	vs auth status / vs doctor	"Why does vs auth status say invalid"
Local CLI error / stack trace	the error's own recovery output	"I see ERR_AUTH_REQUIRED; what now"
Product concept	official docs via the bundled documentation helper	"What is a scene in Viking AI Search"
API contract / request-response shape / field semantics / enum / validation / error code	GitHub main API Reference, starting from the root README URL in the API Reference source section	"What does RelevanceCutoffConfig accept"
Application / dataset effective (valid) data volume	first resolve each bound dataset's Type (via app get / app dataset-config list). For item/video datasets, call app item-data-count per dataset and sum ValidCnt (DataNum is not the answer). For document datasets, do NOT call app item-data-count (it returns InvalidParameter: DatasetType) — use app dataset-config list --full (the --full flag is required; the default compact output omits DocumentStats) and read the document count from Config[].Dataset.DocumentStats (DocumentNum + DocumentFromHomepageNum), not DataNum. Always report the document count, including when it is 0. For user_event (behavior) datasets only, exclude them entirely because they have no effective-data-volume metric at all — do not query a count and do not list them (this "omit even when 0" rule applies to user_event only, never to item/video/document, which are always reported even when their count is 0).	"How much effective data does application X have"

Do not use vs --help to answer product-concept questions. Do not use docs to explain command flags when the installed CLI can answer directly.

Knowledge Sources

CLI source

Allowed CLI checks include:

• vs --help
• vs <cmd> --help
• vs doctor
• vs auth status
• vs llm status
• vs skill list
• vs skill search <query>
• vs skill show <name>

For command-related questions, use vs <domain> --help or vs <cmd> --help as the primary source for command existence, flags, examples, and currently installed behavior. Use the online API Reference router only when the user asks how a command maps to an API category or contract.

Before recommending any command or flag, verify that it exists through vs skill list, vs <domain> --help, or vs <cmd> --help.

API Reference source

API contracts are not answered from local installed snapshots. When the user asks about OpenAPI request/response shape, field semantics, enums, validation rules, or error codes, fetch the latest API Reference from GitHub main:

https://raw.githubusercontent.com/volcengine/SearchCLI/main/skills/vs-product-qa/references/api-references/README.md

Read it progressively:

api-references/README.md
  -> command/resource routing or control-plane / data-plane
  -> second-level category README, such as control-plane/scene/README.md
  -> one OpenAPI Markdown file, such as PublishSearchSceneV2.md

Resolve every relative Markdown link in the API Reference tree against the GitHub raw URL of the current document, not against the local workspace. For example, ./control-plane/README.md from the root README means:

https://raw.githubusercontent.com/volcengine/SearchCLI/main/skills/vs-product-qa/references/api-references/control-plane/README.md

Rules:

• CLI flags and command availability still come from the installed vs --help; the GitHub API Reference must not override observed CLI behavior.
• API contract details come from GitHub main API Reference.
• If GitHub fetching fails or the target API row is marked missing, say the online API Reference is unavailable or incomplete for that API. Do not fall back to old local snapshots and do not guess.

Documentation source

This skill includes a private bundled documentation helper under references/volcengine-documentation/SKILL.md. Use it only as an internal sub-workflow of vs-product-qa. Do not expose it as a separate skill, and do not ask the user to install, trigger, or switch to it.

Use the helper only for official Volcengine documentation, with these fixed rules:

• base root URL: https://www.volcengine.com/docs/85296/1544972
• Chinese questions use https://www.volcengine.com/docs/85296/1544972?lang=cn
• non-Chinese questions use https://www.volcengine.com/docs/85296/1544972?lang=en
• stay within that root page and its child pages only
• do not access sibling product pages, other documentation roots, site-wide search, homepage navigation, or external sites
• do not switch between ?lang=cn and ?lang=en in the same answer unless the user's language changes in a later turn

For Viking AI Search documentation lookup, the data-source restriction is hard:

• product code is Universal AI Search
• the helper script's search action must use ServiceCodes="Universal AI Search"
• the helper script's fetch action must use URLs under https://www.volcengine.com/docs/85296 only

Documentation lookup protocol:

1. Choose the root URL by the user's language.
2. If the page URL is already known, or the user provides a page URL under the allowed root, use the helper script's fetch action first.
3. Otherwise, use the helper script's search action with ServiceCodes="Universal AI Search".
4. After identifying the correct page under the same subtree, use fetch to retrieve the full page content and extract the relevant section.
5. If exact sub-page lookup fails because of selector change, network error, or timeout, return the selected root URL and tell the user to browse manually. Do not guess sub-page URLs.

Fetch budget:

• at most 3 fetches per answer
• at most 5 seconds per fetch
• total budget 15 seconds or less
• on timeout or overrun, degrade honestly and cite the root URL

Not used in MVP:

• https://www.volcengine.com/llms.txt
• https://www.volcengine.com/sitemap.xml
• undocumented per-page markdown endpoints

These may be used later only if they become publicly available and are actually fetched during the current turn.

Commands

• doctor: inspect local CLI environment and service readiness
• auth status: inspect local Viking AK/SK auth state
• llm status: inspect local LLM auth state for LLM-backed features
• skill list: list installed skills before recommending handoff
• skill search: find a relevant installed skill before recommending handoff
• skill show: inspect another skill's workflow before delegating
• vs <cmd> --help: inspect command usage and flags; this is a source rule, not a frontmatter command entry

Workflow

1. Classify the question using Question Routing.
2. Pick the source:
   • CLI usage -> run the relevant vs ... --help; use the online API Reference router only for command-to-API-category questions
   • local auth / environment -> use vs auth status, vs doctor, or vs llm status
   • local CLI error -> use the recovery output from this turn first; only run more CLI checks if needed
   • API contract -> fetch GitHub main API Reference from the root README, then follow category and endpoint links
   • product docs -> use the bundled documentation helper privately, with the hard Universal AI Search + docs/85296 restriction
3. Extract only the lines or sections needed to answer.
4. Respond using the required output format.
5. If the topic involves credentials, environment variables, or vs auth import-env, append the AK/SK security notice from references/aksk-notice.md.
6. If the request implies write operations such as apply, update, create, or bind, explain only the safe draft or dry-run path unless the user explicitly switches to the correct execution workflow.
7. If no grounded source can answer, return unknown, cite the checked CLI command or documentation root URL, and suggest support-ticket / oncall escalation.

Output Format

Use this structure and omit fields that have no grounded content:

**Conclusion**: <one-line direct answer>
**Source**: <doc URL + section heading> OR <CLI command + relevant output>
**Steps**: <ordered steps, only if actionable>
**CLI entry**: <`vs ...` command, or "see `vs <x> --help`">

CLI entry is only for questions that ask how to use a command. For a data/result question (e.g. "how much effective data does app X have"), omit CLI entry entirely — do not list the commands you ran to obtain the number.

For a data/result question, output only Conclusion. Omit Source, Steps, and CLI entry entirely — do not name the command or backend action, do not add a provenance line, and do not list the commands you ran to obtain the number.

Rules:

• for CLI-usage questions, Source cites the command output used in this turn
• for doc-grounded answers, Source must be a real URL retrieved in this turn
• do not fill missing fields with speculation
• do not paste large documentation blocks; summarize in your own words and link the source
• do not narrate internal method selection or source-routing decisions in the answer (e.g. which field/API/skill was chosen, or why another field such as DataNum was rejected); report the grounded result directly
• do not list the commands used to fetch a result (no CLI entry command dump) unless the user explicitly asked how to run it
• for a data/result answer, output only the Conclusion line; do not emit a Source line at all (no command name, no backend action, no field references)

Constraints

1. Grounded only: every factual claim must be backed by CLI output from this turn or an official documentation URL retrieved in this turn. Citing that provenance in a Source line is required for CLI-usage and doc-grounded answers, but for a data/result answer (e.g. an effective-data-count number) do not print a Source line — stay grounded internally and output only the Conclusion.
2. No memory answers: do not answer Viking AI Search product questions from training memory.
3. No fabricated URLs: sub-page URLs must come from routing actually performed in this turn. Do not guess paths.
4. No CLI hallucination: do not recommend commands or flags that do not exist.
5. CLI overrides docs: when CLI help and documentation conflict, trust the installed CLI and say the docs may be stale. 5a. Command questions check installed CLI first: when the user asks about a concrete command, its parameters, or flags, first confirm the installed CLI surface with vs ... --help. Use the online API Reference router only for command-to-API-category questions. 5b. API contract questions use GitHub main API Reference: when the user asks about OpenAPI request/response, fields, enums, validation, or error codes, fetch the root API Reference README from GitHub main and follow links progressively to the target endpoint. Do not answer API contracts from local installed snapshots.
6. No silent execution: do not run write commands such as apply, update, create, or bind on the user's behalf.
7. AK/SK notice required: whenever credentials or vs auth import-env are involved, append the AK/SK security notice.
8. Honest unknowns: if available sources cannot answer, say unknown, explain what source was checked, and suggest escalation.
9. Honest fallback: if exact sub-page lookup fails, explicitly say: "exact sub-page lookup failed; here is the chapter root."
10. No large doc dumps: summarize rather than pasting long documentation blocks.
11. Bundled helper only: use the documentation helper only inside vs-product-qa; do not expose it as a standalone skill.
12. Language-aware docs: use ?lang=cn for Chinese questions and ?lang=en for non-Chinese questions.

Delegation

Use references/delegation.md. When delegating, briefly say which skill takes over and why, then stop.

Examples

CLI usage question

User: "What does the --scene-id flag of vs search run do?"

1. Classify as CLI usage.
2. Run vs search run --help.
3. Answer using only that command output.

Expected shape:

**Conclusion**: Per `vs search run --help`, `--scene-id` selects the configured search scene used for the request.
**Source**: `vs search run --help` output from this turn.
**CLI entry**: `vs search run --scene-id <id> ...`

Product concept question

User: "What is a scene in Viking AI Search?"

1. Classify as product concept.
2. Use the bundled documentation helper privately inside vs-product-qa.
3. Choose the documentation root by user language: ?lang=cn for Chinese, otherwise ?lang=en.
4. Run the helper script's search action with ServiceCodes="Universal AI Search".
5. Use the helper script's fetch action only if the chosen page URL stays under https://www.volcengine.com/docs/85296.
6. Answer using only retrieved documentation.

Delegation question

User: "How do I buy Viking AI Search and create my first AK/SK?"

Answer: "This is covered by vs-user-onboarding because it is a sign-up, purchase, and first AK/SK workflow. Handing off."

References

• AK/SK security notice: references/aksk-notice.md
• Delegation matrix: references/delegation.md
• Bundled documentation helper: references/volcengine-documentation/SKILL.md
• Online API Reference entry: https://raw.githubusercontent.com/volcengine/SearchCLI/main/skills/vs-product-qa/references/api-references/README.md