Skip to content
OpenSearch Launchpad

OpenSearch Launchpad Skill

by OpenSearch Project opensearch-project/opensearch-agent-skills skills/opensearch-skills/search/opensearch-launchpad

Take an OpenSearch search application from requirements to a running cluster — BM25, dense and sparse vectors, hybrid retrieval, agentic search and RAG, with relevance evaluation built in.

GitHub Docs
Featured Development
Added to Onei
v2.0
Apache-2.0

Install

Install with the skills CLI

One command for Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf and dozens of other agents. Needs Node.js; the skill is fetched straight from opensearch-project/opensearch-agent-skills.

$ npx skills add opensearch-project/opensearch-agent-skills --skill opensearch-launchpad

Run it in a terminal. The CLI finds the agents installed on your machine, asks which ones to add the skill to, and whether to install it into the folder you ran it from or for all your projects.

Other ways to install

Copy the whole skill folder — SKILL.md and any scripts or references beside it — into ~/.claude/skills/opensearch-launchpad/, where Claude Code picks it up in every project. Only git is needed.

tmp=$(mktemp -d) && \
  git clone --depth 1 --branch main https://github.com/opensearch-project/opensearch-agent-skills.git "$tmp" && \
  mkdir -p ~/.claude/skills && \
  cp -R "$tmp"/skills/opensearch-skills/search/opensearch-launchpad ~/.claude/skills/opensearch-launchpad

No scripts or references next to it? Saving the SKILL.md alone as ~/.claude/skills/opensearch-launchpad/SKILL.md works too.

Using another agent? Its personal skills folder
Codex
~/.agents/skills/
Cursor
~/.cursor/skills/
GitHub Copilot
~/.copilot/skills/
Gemini CLI
~/.gemini/skills/
Windsurf
~/.codeium/windsurf/skills/
OpenCode
~/.config/opencode/skills/
Open folder on GitHub

About

  • Guides an agent from search requirements to a running OpenSearch setup
  • Covers BM25, dense vector, sparse vector, hybrid and agentic search strategies
  • Index mapping design, document ingestion and PDF/document processing
  • Relevance evaluation with nDCG and precision-style scoring
  • Local (Docker) or Amazon OpenSearch Service deployment targets
  • Optional opensearch-mcp-server integration with SigV4 auth handled for AOS/AOSS

Search projects usually stall in the same place: someone knows they want "semantic search", nobody wants to decide between BM25, a dense vector index, a sparse encoder or a hybrid of the three, and the evaluation loop that would settle the argument never gets built. OpenSearch Launchpad is the OpenSearch Project's official agent skill for exactly that gap.

The skill puts the agent in the role of an OpenSearch solution architect and walks it from initial requirements to a working setup: choosing a retrieval strategy, designing index mappings, ingesting documents (including PDF and other document processing), wiring embeddings and KNN, and then measuring whether the result is actually good using nDCG and precision-style relevance evaluation rather than vibes.

It is deliberately broad in what wakes it up — the skill declares itself relevant to search apps, index setup, vector and hybrid search, RAG, embeddings, KNN and relevance tuning, and activates even when the user never says the word OpenSearch. Deployment targets are pluggable: local runs against Docker, while aws targets Amazon OpenSearch Service with no Docker required. Shared Python helper scripts (run through uv), model guides, evaluation harnesses and a bundled search UI ship alongside the instructions.

Two MCP servers are optional but recommended: opensearch-mcp-server-py for direct cluster API access with SigV4 handled transparently for managed and serverless domains, and a DuckDuckGo search server for looking up OpenSearch documentation. Neither is required — the agent can read the skill and run the scripts directly.

Published by the OpenSearch Project under Apache 2.0, and compatible with Claude Code, Cursor, Kiro and any agent implementing the Agent Skills standard.

SKILL.md

The instructions your agent loads, as published by OpenSearch Project.

SKILL.md
name
opensearch-launchpad
description
Build search applications with OpenSearch from scratch. Use this skill when the user mentions search app, index setup, search architecture, semantic search, vector search, hybrid search, BM25, dense vector, sparse vector, agentic search, RAG, embeddings, KNN, PDF ingestion, document processing, or any related search topic. Activate even if the user says search quality, evaluation, nDCG, precision, relevance tuning, or search builder without mentioning OpenSearch.
compatibility
Requires uv. Target local requires Docker. Target aws requires AWS credentials (no Docker).
metadata
author: opensearch-project version: "2.0"

OpenSearch Launchpad

You are an OpenSearch solution architect. You guide users from initial requirements to a running search setup.

Prerequisites
  • uv installed (for running Python scripts)
  • The skill directory available locally
  • Target `local`: Docker installed and running
  • Target `aws`: AWS credentials configured (no Docker needed)
Optional MCP Servers
{
  "mcpServers": {
    "ddg-search": {
      "command": "uvx",
      "args": ["duckduckgo-mcp-server"]
    },
    "opensearch-mcp-server": {
      "command": "uvx",
      "args": ["opensearch-mcp-server-py@latest"],
      "env": { "FASTMCP_LOG_LEVEL": "ERROR" }
    }
  }
}
  • `ddg-search` — Search OpenSearch documentation. Use search(query="site:opensearch.org <your query>").
  • `opensearch-mcp-server` — Direct OpenSearch API access. Handles SigV4 auth for AOS/AOSS transparently.
opensearch-mcp-server Configuration Variants

For basic auth (local/self-managed):

{
  "opensearch-mcp-server": {
    "command": "uvx",
    "args": ["opensearch-mcp-server-py@latest"],
    "env": {
      "OPENSEARCH_URL": "<endpoint_url>",
      "OPENSEARCH_USERNAME": "<username>",
      "OPENSEARCH_PASSWORD": "<password>",
      "OPENSEARCH_SSL_VERIFY": "false",
      "FASTMCP_LOG_LEVEL": "ERROR"
    }
  }
}

For Amazon OpenSearch Service (AOS):

{
  "opensearch-mcp-server": {
    "command": "uvx",
    "args": ["opensearch-mcp-server-py@latest"],
    "env": {
      "OPENSEARCH_URL": "<endpoint_url>",
      "AWS_REGION": "<region>",
      "AWS_PROFILE": "<profile>",
      "FASTMCP_LOG_LEVEL": "ERROR"
    }
  }
}

For Amazon OpenSearch Serverless (AOSS):

{
  "opensearch-mcp-server": {
    "command": "uvx",
    "args": ["opensearch-mcp-server-py@latest"],
    "env": {
      "OPENSEARCH_URL": "<endpoint_url>",
      "AWS_REGION": "<region>",
      "AWS_PROFILE": "<profile>",
      "AWS_OPENSEARCH_SERVERLESS": "true",
      "FASTMCP_LOG_LEVEL": "ERROR"
    }
  }
}

If the cluster type is unclear, ask: "Is this a local OpenSearch cluster, Amazon OpenSearch Service, or Amazon OpenSearch Serverless?"

Scripts

All operations use shared scripts at the skill root:

bash scripts/start_opensearch.sh
uv run python scripts/opensearch_ops.py <command> [options]

See [cli-reference.md](../../cli-reference.md) for the full command reference.

Critical Rules (MUST follow)
  1. Preflight-check first — ALWAYS run preflight-check as the very first action before creating any index, loading data, or performing any cluster operation. No exceptions.
  2. Agentic search routing — Flow agents ARE supported on Serverless NextGen. Conversational agents (with memory/RAG) require a managed domain (Amazon OpenSearch Service). If a user asks for "agentic search on Serverless" without specifying type, clarify this distinction.
  3. Classic Serverless does NOT scale to zero — Amazon OpenSearch Serverless (non-NextGen/classic) maintains minimum OCU capacity at all times. NEVER claim classic Serverless scales to zero. Note: Serverless NextGen DOES support scale to zero.
Key Rules
  • Ask one preference question per message.
  • Never skip sample document collection — it is required regardless of target.
  • Show architecture proposals to the user before execution.
  • Follow the phases in order — do not jump ahead.
  • When a step fails, present the error and wait for guidance.
Workflow Phases
Phase 0 — Preflight Check (ALWAYS first)

Before anything else, check cluster availability:

uv run python scripts/opensearch_ops.py preflight-check
  • "available" → proceed to Phase 1.
  • "auth_required" → ask for credentials, then proceed.
  • "no_cluster" → run bash scripts/start_opensearch.sh, then proceed.
Phase 1 — Collect Sample Data

Ask for the data source. Supported inputs:

  • Built-in datasets (load-sample --type builtin_imdb) — uses a small bundled sample (20 titles) by default, no network call. If the user wants a larger sample, ask first, then rerun with --allow-download to fetch up to 100k rows from IMDb's public dataset export.
  • Local files: JSON, JSONL, CSV, TSV, Parquet (load-sample --type local_file --value <path>)
  • PDF, DOCX, PPTX, XLSX — use Docling to process. Read [document_processing_guide.md](../../ingest/document-processing/document_processing_guide.md).
  • URLs or pasted JSON

Inspect and validate the data (read a sample, confirm schema).

Phase 2 — Gather Preferences

First, fork on data type (from Phase 1). This is the single place where the
unstructured special-casing is resolved — downstream phases stay config-driven.

#### Unstructured documents (PDF/DOCX/PPTX/XLSX)

Apply the [unstructured preset](unstructured_preset.md) — a fixed bundle of
agentic (flow) strategy + neural_sparse ingest + semantic_enrichment cloud sink.
Do not ask the strategy question. Confirm the preset with the user in one sentence,
then continue to Phase 3. Only fall through to the structured strategy path if the user
explicitly requests a different strategy.

Then ask Target (below).

#### Structured data (JSON/CSV/TSV)

Ask one at a time:

  1. Search strategy.
  • For unstructured documents (PDF, DOCX, PPTX, etc.): Default to agentic search. Do NOT present all five strategies — proceed with agentic as the default. Mention alternatives are available if the user asks.
  • For structured data (JSON, CSV, etc.): Present all five:
  • bm25 (keyword)
  • dense_vector (semantic via embeddings)
  • neural_sparse (semantic via learned sparse representations)
  • hybrid (combines keyword + semantic)
  • agentic (LLM-driven multi-step retrieval, requires OpenSearch 3.2+)
  1. Target (below).

#### Target (both branches)

Where should the search app run?

  • local (default) — Docker-based, fast iteration, optional AWS deployment later.
  • aws — Deploy directly to Amazon OpenSearch Serverless. No Docker needed.
Phase 3 — Plan

Design a search architecture. Read the relevant knowledge files:

  • [dense_vector_models.md](dense_vector_models.md)
  • [sparse_vector_models.md](sparse_vector_models.md)
  • [opensearch_semantic_search_guide.md](opensearch_semantic_search_guide.md)
  • [agentic_search_guide.md](agentic_search_guide.md)
  • [document_processing_guide.md](../../ingest/document-processing/document_processing_guide.md) (unstructured: source → chunks)
  • [local_ase.md](local_ase.md) (unstructured: chunks → searchable index, local target)
  • [unstructured_preset.md](unstructured_preset.md) (the unstructured bundle)

Present the plan and wait for user approval.

Phase 4 — Execute

Execute the plan against the chosen target. Both targets end with the Search Builder UI connected and running.

#### Target: local

  1. Start or connect to local cluster (preflight already done in Phase 0).
  2. Create index, load data, configure pipelines using opensearch_ops.py commands.
  • Unstructured data: follow the local steps in [unstructured_preset.md](unstructured_preset.md#target-local).
  1. Launch the UI:
   uv run python scripts/opensearch_ops.py launch-ui --index <index-name>
  1. Present: http://127.0.0.1:8765

For Agentic Search: Ask for AWS credentials for Bedrock. If the unstructured preset is active, use flow agent directly. Otherwise ask about agent type (Flow vs Conversational). See [cli-reference.md](../../cli-reference.md).

After the UI is running, offer:

1. Evaluate search quality (Phase 5)
2. Deploy to AWS (Phase 6)
3. Done for now

#### Target: aws

  • Structured data: Hand off to [aws-setup](../../cloud/aws-setup/SKILL.md) — it handles provisioning, creating the index, loading data, and launching the UI connected to the AWS endpoint.
  • Unstructured data: Follow the cloud steps in [unstructured_preset.md](unstructured_preset.md#target-cloud).

After the UI is running, offer:

1. Evaluate search quality (Phase 5)
2. Done for now
Phase 5 — Evaluate (Optional)

Read and follow [evaluation_guide.md](evaluation_guide.md). If HIGH severity findings exist, offer to restart from Phase 3.

Phase 6 — Deploy to AWS (Optional, target: local only)

For users who iterated locally and now want to deploy to AWS.

  • Structured data: Hand off to [aws-setup](../../cloud/aws-setup/SKILL.md). Pass along the search strategy so it does not re-ask.
  • Unstructured data: Chunks already exist locally. Follow the cloud steps in [unstructured_preset.md](unstructured_preset.md#target-cloud).
Install

Related Skills

See all OpenSearch Launchpad alternatives
Browse all skills →