// Turn a sample document into a running Milvus / Zilliz Cloud search app in minutes. Guide the user through Collect → Configure → Plan → Execute and start a local search UI.
| name | zilliz-launchpad |
| description | Turn a sample document into a running Milvus / Zilliz Cloud search app in minutes. Guide the user through Collect → Configure → Plan → Execute and start a local search UI. |
You are running the zilliz-launchpad skill. Your job is to take a user from "I have some documents I want to search" to "a working search app on Milvus / Zilliz Cloud, scored for quality, running in production" with as little friction as possible.
The skill has six phases:
Every imperative action goes through the CLI scripts/zilliz_ops.py. Phases write outputs into scripts/runs/<utc-iso>/.
Do not start a phase before the previous one has produced its artifact. If a phase exits non-zero, surface the error and stop. Each phase can be rerun — the CLI is idempotent.
| Phase | CLI subcommand | Required input | Output file |
|---|---|---|---|
| 1 Collect | collect | --sample <name> or --input <path> | collect.json |
| 2 Configure | configure | --from-json <file> or flags | configure.json |
| 3 Plan | plan | run dir with collect + configure | plan.json, plan.md |
| 4 Execute | execute | run dir with plan | execute.json (+ live Milvus + running sidecar) |
| 5 Evaluate | evaluate | run dir with execute | eval_report.json, eval_report.md |
| 6 Deploy | deploy | run dir with execute + Cloud credentials | deploy.json (+ observability.json) |
Check the basics:
If the user is going local for the first time, bring Milvus up:
cd skills/zilliz-launchpad/scripts
./start_milvus.sh up
Resolve secrets from env first. Only if a variable is missing and needed by the current phase, prompt the user in dialogue and re-invoke with it exported. Typical variables:
OPENAI_API_KEY — OpenAI embeddings / rerankerVOYAGE_API_KEY — Voyage embeddingsCOHERE_API_KEY — Cohere embeddings / rerankZILLIZ_TOKEN — Zilliz Cloud authenticationZILLIZ_BYOM_URL, ZILLIZ_BYOM_KEY — Zilliz BYOM endpointWhen the CLI fails with {"code":"missing_credential",...}, the payload includes the variable name and an export_hint. Use that hint verbatim when asking the user.
zilliz CLIInstall the zilliz CLI (≥ 0.3.0) to unlock three Cloud-only enhancements. The CLI is never required for local Milvus.
| Phase | With zilliz on PATH | Without |
|---|---|---|
| 2 Configure | zilliz cluster list → pick a cluster; writes cluster_id to configure.json | prompt for URI + token |
| 4 Execute pre-flight | zilliz cluster describe gates on RUNNING; fails fast with cluster_not_ready | skipped |
| 4 Execute bulk | zilliz import create for corpora above bulk_import_threshold (default 100k) | client-side upsert |
| 2/4 Token fallback | zilliz auth whoami returns a scoped token when ZILLIZ_TOKEN is unset | missing_credential |
| 6 Deploy create | zilliz cluster create provisions a new cluster when --create --confirm is passed | prompt for --cluster-id of an existing cluster |
| 6 Deploy bulk | zilliz import create for corpora above the plan's threshold | client-side upsert |
Load only the references you need for the current phase. Do not prefetch all of references/ — that wastes context.
| Phase | Load from references/ |
|---|---|
| 1 Collect | knowledge/document_processing.md |
| 2 Configure | knowledge/rag_templates.md, deploy-*.md (match the user's target) |
| 3 Plan | knowledge/dense_embedding_models.md, knowledge/sparse_embedding_models.md, knowledge/index_tuning.md, knowledge/schema_design.md, knowledge/hybrid_search_guide.md, knowledge/reranker_guide.md, knowledge/image_embedding_models.md (image-search only) |
| 4 Execute | cli-reference.md |
| 5 Evaluate | knowledge/evaluation_guide.md |
| 6 Deploy | observability/metrics.md, observability/query-analysis.md, deploy-*.md (match the target) |
| Ops (any) | ops-attu.md — when the user wants to inspect, debug, or administer a Milvus cluster beyond what the CLI covers |
Attu is an opt-in admin UI for developers and ops, not a replacement for the Next.js demo UI. Bring it up with ./start_milvus.sh attu up (serves on http://localhost:8000, bound to loopback) when the user needs to verify ingest, drill into eval bad-cases, or run Cloud operations Attu covers but the CLI does not. See references/ops-attu.md for playbooks and Cloud connection instructions.
Assume the working directory is the repo root.
# Phase 1 — sample data (or --input your.jsonl / .csv / .txt / .md / .pdf / image dir)
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py collect --sample movies
# Phase 2 — take defaults with overrides
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py configure \
--use-case rag --dataset-size 20 --deployment local-standalone
# Phase 3
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py plan
# Phase 4
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py execute --sample movies
# Phase 5 — derived smoke eval, or with --qrels / --compare for a real eval
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py evaluate
# Phase 6 — promote to an existing Cloud cluster, or --create --confirm to provision
uv run python skills/zilliz-launchpad/scripts/zilliz_ops.py deploy --cluster-id <id>
On success of Phase 4, start the demo UI:
cd skills/zilliz-launchpad/scripts/ui
pnpm install
pnpm dev
# open http://localhost:3000
For text collections the demo UI exposes a Milvus filter-expression input (e.g. year >= 2023) and a reranker toggle (Off / Default, where Default follows the plan's recorded reranker). Result cards render every non-vector scalar field returned by the collection, so ingested metadata like title / year / genre shows up next to the chunk text without further work.
Every CLI error arrives on stderr as a single-line JSON object:
{"code": "missing_credential", "message": "...", "env_var": "OPENAI_API_KEY", "export_hint": "export OPENAI_API_KEY=..."}
Known codes and how to react:
| code | action |
|---|---|
missing_credential | prompt the user for the value; ask them to export it; retry |
missing_dependency | optional extra not installed (e.g. image-search needs [multimodal]); surface install_hint verbatim and ask the user to install |
schema_conflict | a collection exists with a different schema; offer to drop or to rename the plan's collection_name |
sparse_unavailable | the user picked hybrid/sparse but the collection was built without sparse; offer to rebuild |
invalid_profile | Phase 3 got a malformed input; go back to Configure |
backend_unsupported | requested index (e.g. DiskANN) isn't available on the target; pick a fallback from index_tuning.md |
zilliz_cli_missing | a Cloud-only feature needs zilliz; point at install_url in the payload |
zilliz_cli_auth | CLI present but not logged in; tell the user to run zilliz auth login |
cluster_not_ready | pre-flight found a non-RUNNING cluster; surface state + remediation verbatim |
qrels_missing | comparison mode needs labels; ask the user for --qrels <path> |
judge_unavailable | --judge-llm requested but the provider's API key env var is unset |
cluster_create_failed | surface stderr and exit_code verbatim; often a quota or region issue |
bulk_import_failed | include job_id when telling the user to check their Cloud console |
destructive_without_confirm | deploy --create needs --confirm; summarise projected cost before re-running |
lib/)--use-case video-search, UI inline <video> deep-link playback); audio / speech-to-text, subtitle / OCR extraction, and live video streams are out of scopeIf the user asks for any of the above, politely say it's on the roadmap and keep scope tight.
Be concrete. When the user asks an open question (e.g., "should I use HNSW or DiskANN?"), consult the relevant reference in knowledge/, state a recommendation with one-line reasoning, and proceed. Don't pile up options without a pick.