[Feature] Conversational Documentation Discovery Powered by OpenSearch

[owaiskazi19]

Is your feature request related to a problem?

Finding the right OpenSearch documentation for a specific use case is difficult, especially for newcomers. The documentation covers a broad surface area — search, observability, security analytics, vector search, ML, ingest pipelines, index management — and users often don't know the right terminology to find what they need.

Users know what they want to build but not which docs to read.

Examples:

• "I want to build a product search with typo tolerance and faceted filtering"
• "How do I set up log monitoring with alerts for my Kubernetes cluster?"
• "I need to implement RAG with my own PDF documents"

The current keyword-based search requires users to already know OpenSearch-specific terms and concepts, creating a significant barrier to entry.

Describe the solution you'd like

An AI-powered conversational search widget embedded on the OpenSearch documentation website — built using OpenSearch's MCP Server, agentic search or opensearch-launchpad.

| Capability | How it helps |
|---|---|
| **Multi-step reasoning** | Decomposes "I want semantic search with filtering" into sub-queries for neural search docs, query DSL docs, and pipeline docs |
| **Answer synthesis** | Returns a coherent step-by-step guide instead of a ranked list of links |
| **Conversation memory** | Users ask follow-ups ("What about adding personalization?") and the agent retains context via `memory_id` |
| **Automatic retrieval selection** | Agent chooses BM25 for keyword lookups vs. dense vector for conceptual questions |
| **Query planning** | `QueryPlanningTool` generates optimal OpenSearch DSL against the documentation index |

User experience

Input: A text box asking "What are you trying to build with OpenSearch?"

Output: A synthesized response containing:

• Step-by-step guidance tailored to the user's problem
• Direct links to relevant documentation pages
• Suggested OpenSearch features and configurations
• Starter code snippets (index mappings, queries, pipelines) where applicable

Follow-ups: Users can ask clarifying questions in the same conversation, with the agent retaining context.

Example prompts (shown as suggestions or via a "Surprise me" button):

• "Build a product catalog search with autocomplete"
• "Monitor application logs and detect anomalies"
• "Implement RAG over my company's knowledge base"
• "Migrate from Elasticsearch to OpenSearch"

[owaiskazi19]

@kolchfa-aws @epugh can you take a look and let me know your thoughts?

[epugh]

Something better than what we have would be amazing... I am bummed that we never quite got the alignment to feed searches and clicks in as part of UBI and move beyond basic Google Analytics.. The project OpenSearch server setup is rather opaque and hard to understand. I really wish that it was a public repo with nice docs and explainer how to use OpenSearch itself. --> There have been mulltiple "calls to action" or "I want to improve it", but until the keys to how it's defined and run are made more common/easier, it's just tough.

In terms of your prompts, I think anything is better than what we have.

One thought though, if we are putting energy out there into "a tool for building a search experience", we may be better off directing folks to the dedicated OpenSearch Launchpad tool. Which would be a great repsonse to "I want to implement RAG over my company's knowledge base" --> Here are some resources and you should use this tool!

[owaiskazi19]

One thought though, if we are putting energy out there into "a tool for building a search experience", we may be better off directing folks to the dedicated OpenSearch Launchpad tool. Which would be a great repsonse to "I want to implement RAG over my company's knowledge base" --> Here are some resources and you should use this tool!

That's a good call. We can use the opensearch launchpad https://github.com/opensearch-project/opensearch-launchpad to power the documentation as well.

[epugh]

One thought though, if we are putting energy out there into "a tool for building a search experience", we may be better off directing folks to the dedicated OpenSearch Launchpad tool. Which would be a great repsonse to "I want to implement RAG over my company's knowledge base" --> Here are some resources and you should use this tool!

That's a good call. We can use the opensearch launchpad https://github.com/opensearch-project/opensearch-launchpad to power the documentation as well.

Interesting... As in, "use the tool for what it's meant for"? I like it. I mean, if that worked, that would demonstrate the value of it!

[kolchfa-aws]

I think that while having a docs assistant can be beneficial, we first need to fix the search so it works better. We can now use semantic search on the search backend in addition to keyword search, and we need to tune it so users can find related documentation easily. I agree with Eric in that downloading skills is probably an easier path for LLM-based search. Also, I would rename the opensearch-launchpad repo to opensearch-skills or opensearch-agent-skills because "launchpad" is not standard naming, and people searching for LLM skills would not think that this repo contains skills.

[owaiskazi19]

I think that while having a docs assistant can be beneficial, we first need to fix the search so it works better. We can now use semantic search on the search backend in addition to keyword search, and we need to tune it so users can find related documentation easily.

Agreed. Switching to semantic/hybrid or agentic could be the first step for this longer vision for conversation chat bot for the documentation.

Also, I would rename the opensearch-launchpad repo to opensearch-skills or opensearch-agent-skills because "launchpad" is not standard naming, and people searching for LLM skills would not think that this repo contains skills.

Launchpad repo serves a different purpose. We would have an agent-skills repo to be a unified place for all skills coming soon opensearch-project/.github#482