Volcano Engine Milvus

Manage Milvus instances on Volcano Engine — cluster lifecycle and vector data operations.

Quick Start

Use the bundled CLIs (always run via the skill venv):

{baseDir}/venv/bin/python {baseDir}/scripts/control.py <command>
{baseDir}/venv/bin/python {baseDir}/scripts/data.py <command>

If {baseDir}/venv does not exist:

python3 -m venv {baseDir}/venv
{baseDir}/venv/bin/pip install -r {baseDir}/requirements.txt

See CONTROL_PLANE.md and DATA_PLANE.md for workflows and examples.

Low-level control-plane fallback (use only when goal-based commands do not cover the task): {baseDir}/venv/bin/python {baseDir}/scripts/control_tools.py <command>

Available operations

Control Plane (cluster management): Use goal-based workflows to provision, inspect, scale, delete, and expose Milvus instances. → See CONTROL_PLANE.md for goal-based commands and workflows. → See CONTROL_TOOLS.md for low-level control_tools.py fallback commands (use only when CONTROL_PLANE.md does not cover the task).

Data Plane (collections & data): Create/drop collections, insert/upsert/delete data, vector search, scalar query, and get-by-ID. → See DATA_PLANE.md for commands and use cases.

Out of scope

• Deploying or operating Milvus outside Volcano Engine (self-hosted, other clouds).
• Deep Milvus performance tuning or schema design beyond basic collection creation and queries.
• Application-level embedding strategy decisions (chunking, RAG design) unless needed to run the provided data plane commands.

Rules

Common

• Execution environment: Always use {baseDir}/venv/bin/python to run scripts.
• Authentication: VOLCENGINE_ACCESS_KEY and VOLCENGINE_SECRET_KEY are required for all control-plane operations. Data-plane commands also require a reachable Milvus --endpoint plus any needed Milvus auth flags. See DATA_PLANE.md.
• Script usage: Prioritize scripts/control.py and scripts/data.py. Use scripts/control_tools.py only as a last resort when goal-based commands do not cover the task. Do not write ad-hoc Python scripts or use the SDK directly unless existing CLIs cannot satisfy a specific requirement.
• Language (strict): Always reply in the user's language. Use a deterministic heuristic:
  • If the user's message contains any Chinese characters, reply in Chinese.
  • Otherwise, reply in the user's language as inferred from their message.
  • Keep commands/flags/code in English; only the explanation and prompts should be localized.
  • If the user mixes languages and preference is unclear, ask which language they prefer.

Control plane

• Destructive actions (strict): Never run delete operations until:
  1. You first fetch and show what will be deleted (preview/detail/describe; see CONTROL_PLANE.md).
  2. The user replies with an explicit confirmation phrase that includes the exact target identifier.
  3. You pass that exact target identifier into the CLI --confirm argument.
• Missing parameters: Never fail silently. Fetch available options (VPCs, Subnets, Specs) and present them to the user interactively.
• EIP Auto-Reuse: When enabling public endpoints with enable-public or ms-enable-public, prefer using --eip-auto-reuse true to leverage existing unbinded EIPs and avoid quota failures.

Data plane

• Network access: Before any data-plane operation, validate the endpoint is reachable. If the instance has no public address or is unreachable, use the enable-public or ms-enable-public control workflows to expose the instance and set IP allow-groups (whitelists), then re-fetch the endpoint via status.
• Embedding/auto-embedding: Only use the built-in embedding flags supported by scripts/data.py (no custom embedding scripts). Follow the embedding config and input-type rules in DATA_PLANE.md.
• Vector inputs (strict): Do not accept or generate raw vectors for insert/upsert/search. This skill is auto-embedding-only for data writes and semantic search.