Universal Search and Expand Bundle

Overview

Universal Search and Expand is a Gemma Edge Gallery skill bundle that retrieves fresh information from external APIs and turns it into a readable, grounded answer.

The bundle is designed to work well on phones and other resource-constrained devices. It keeps retrieval and expansion separate so the model can first gather context and then rewrite that context into a clearer response.

Use search_and_expand as the main entry point. It is the recommended skill for most users because it chains the retrieval and expansion steps for you.

What the bundle does

The bundle is built around three skills:

1. search_and_expand
2. universal_search
3. expand_answer

Recommended flow:

• search_and_expand calls universal_search
• universal_search calls one or more search APIs and returns dense plain text
• expand_answer rewrites that text into a fuller answer

This design is useful when the model alone produces short, generic answers. By separating search from expansion, the bundle gives the model a better starting point for a grounded response.

Why not just 1 skill? A walk down the development process

I orignallly tried just 1 skill. I separted the concerns into various files. So much more robust in structure, but not in function. Slimmed down to the flattest possible format. Then the problem was Gemma completely ignored the skill instructions. You can try it yourself and send searches to universal_search if you are a glutton for narrated meta summaries.

Next, I tried expand_answer to elaborate further. This worked with a few tweaks, but the narrated voice persisted. Gemma has a way of inferring intent despite what your explicit instructions may be.

Finally, a 3rd level orchestrator. This got to the best answers to date. The drawback is potential accumulated latency along with a much more laborious deployment process.

I'd like to make this more elegant. Still, a good learning process.

Skills in this repository

1. search_and_expand

This is the recommended skill to use.

It orchestrates the workflow and is the user-facing entry point. In normal use, the user only needs to ask one question, and this skill handles the rest.

2. universal_search

This skill performs the search step.

It supports multiple providers and returns the result as a single dense text block. That output is intentionally simple so the model can use it safely on-device.

3. expand_answer

This skill performs the rewrite step.

It turns the dense search text into a more readable response, usually with more detail and better flow than the raw search output.

Architecture overview

graph TD
  U[User Query] --> S[search_and_expand]
  S --> US[universal_search]
  US --> R[Provider Router]

  R --> WF[Wolfram Alpha]
  R --> SP[Serper]
  R --> TV[Tavily]
  R --> BR[Brave]
  R --> GM[Gemini]

  WF --> M[Merge and Clean]
  SP --> M
  TV --> M
  BR --> M
  GM --> M

  M --> EA[expand_answer]
  EA --> O[Final Answer]

Loading

Execution flow

1. The user asks a question.
2. search_and_expand is triggered.
3. search_and_expand calls universal_search.
4. universal_search selects a provider or tries providers in order.
5. The provider returns search or knowledge text.
6. The text is cleaned and merged into a dense plain-text block.
7. expand_answer rewrites that block into a fuller answer.
8. The final answer is returned.

Deployment structure

The repository is organized as three skill directories.

skills/
└── universal_search_bundle/
    ├── search_and_expand/
    │   └── SKILL.md
    ├── universal_search/
    │   ├── SKILL.md
    │   └── scripts/
    │       └── index.html
    └── expand_answer/
        ├── SKILL.md
        └── scripts/
            └── index.html

Prerequisites

Required

• Google AI Edge Gallery
• A Gemma model installed in the app
• At least one supported API key for search or knowledge retrieval

Recommended

• A Serper API key
• One backup provider key, such as Tavily or Brave
• A Wolfram Alpha API key for math, computation, and factual lookups

Installation on phone

Android phone

1. Install Google AI Edge Gallery on your Android device.
2. Open the app and load the Gemma model you want to use.
3. Add the skills from this repository.
4. Make sure the bundle is installed with the skills in this order:
   • search_and_expand
   • universal_search
   • expand_answer
5. Configure API keys in universal_search/scripts/index.html.
6. Save the files and refresh or reload the skills in the app.

Phone-specific notes

• Use search_and_expand as the default skill.
• Keep API keys in the configuration block near the top of universal_search/scripts/index.html.
• Leave debug mode on only when you are testing.
• Keep the output plain and compact in universal_search; let expand_answer do the rewriting.

Installation on desktop or laptop

The bundle is phone-safe, but it can also be used on desktop or laptop through an Android emulator or any supported local build of the Gallery app.

Option 1: Android emulator

1. Install an Android emulator on your computer.
2. Install Google AI Edge Gallery inside the emulator.
3. Load the Gemma model.
4. Copy the skill bundle into the Gallery skill directory used by the emulator.
5. Refresh the skills list.
6. Open search_and_expand and test with a simple query such as OpenAI.

Option 2: Supported desktop build

If you are using a desktop build of AI Edge Gallery or a compatible runtime:

1. Open the skill directory used by that runtime.
2. Copy the three skill folders into the correct location.
3. Confirm that universal_search/scripts/index.html is present.
4. Reload the app or restart the runtime.
5. Run a test query.

Desktop and laptop notes

• Use the same skill order as on phone.
• Verify that API keys are valid before testing.
• Start with a simple query such as debug if the debug toggle is enabled.

Supported search and knowledge providers

The bundle supports multiple providers. In auto mode, universal_search can try them in sequence and fall back if one is missing, disabled, or unavailable.

Provider	Description	Free tier limits	API URL
Serper	Google Search API wrapper with strong general-purpose web search coverage.	2,500 free queries; no credit card required.	https://serper.dev/
Tavily	Search API designed for AI retrieval and RAG workflows.	Free for students is advertised; the public pricing page also shows a free plan with 1,000 API credits per month.	https://www.tavily.com/pricing
Brave Search	Web search API with web, news, image, and other search results.	Includes $5 in free credits every month.	https://brave.com/search/api/
Wolfram Alpha	Computational knowledge engine for math, science, and factual queries.	Up to 2,000 non-commercial API calls per month.	https://products.wolframalpha.com/api
Gemini API	Google developer API for model-based retrieval, grounding, and fallback text generation.	Free tier available; quotas vary by model and tier.	https://ai.google.dev/gemini-api/docs/pricing

Recommended provider order

The default order in universal_search is designed to balance speed, accuracy, and fallback behavior.

Recommended order:

1. Wolfram Alpha
2. Serper
3. Tavily
4. Brave
5. Gemini

This order works well for mixed queries because Wolfram Alpha is strong for numeric and factual lookups, Serper is fast and broad, Tavily is useful for AI-oriented retrieval, Brave is a good fallback search source, and Gemini can act as a final fallback when needed.

Configuration

Most user-configurable values are grouped near the top of universal_search/scripts/index.html.

Typical configuration values include:

• API keys
• default provider order
• debug toggle
• result limit
• output length limit
• timeout settings

How to use the bundle

Recommended usage

Use search_and_expand directly.

Example prompts:

• Search and expand for OpenAI
• Search and expand for OpenAI revenue
• Search and expand for
• Explain quantum computing
• What is the current status of the project?
• Compare Serper and Brave Search APIs

Advanced usage

You can also test the lower-level skills directly:

• universal_search if you want to inspect the raw retrieval output
• expand_answer if you want to test the rewrite step on existing text

Troubleshooting

The response is too short

• Make sure you are using search_and_expand
• Confirm that expand_answer is installed
• Verify that the model is actually loading the expansion skill

No search results appear

• Check that at least one API key is valid
• Make sure the key is not left as the default placeholder value
• Confirm that the selected provider is enabled

The skill does not run

• Check folder names
• Confirm that SKILL.md is in the correct directory
• Confirm that scripts/index.html is present for the tool-backed skills
• Reload the app after changing files

Debug mode is still active

• Set the debug toggle to off in universal_search/scripts/index.html
• Save the file and reload the skill
• Personally, I leave it on, just to be able to test to make sure the most basic case is working (i.e., "search for debug").

Universal Search responses instead of Search and Expand

• Preface your query with Search and Expand

Responses to the same query may vary

• It's a real time search. Life moves fast.

Known Issues

• Searches may sometimes hang. I think that may be related to phone compute / memory. Research shows this may be inherent latency from sucessful calls and orchestrating multiple calls. Could consider firing silmulataneous API calls or setting a global timeout.

Design principles

• Keep retrieval and expansion separate
• Return a single dense text block from retrieval
• Avoid labels, bullets, and heavy structure in the tool output
• Keep the bundle phone-safe
• Prefer search_and_expand for normal use

Configuration Reference

Field	Description	Required	Possible Values / Format
SERPER_API_KEY	API key for Serper (Google search wrapper). Primary default provider for general web results.	Optional (but recommended)	String (e.g., "abc123"). Leave "" or "key" to disable
TAVILY_API_KEY	API key for Tavily search. Provides AI-optimized summarized results.	Optional	String. Leave empty or "key" to disable
BRAVE_API_KEY	API key for Brave Search API. Independent web index fallback.	Optional	String. Leave empty or "key" to disable
WOLFRAM_API_KEY	API key for Wolfram Alpha. Best for math, calculations, and factual queries.	Optional	String. Leave empty or "key" to disable
GEMINI_API_KEY	API key for Gemini API. Used as a fallback LLM-based search provider.	Optional	String. Leave empty or "key" to disable
PROVIDER_ORDER	Defines the order of search providers used in auto mode. First valid result wins.	Yes	Array of strings: ["wolfram", "serper", "tavily", "brave", "gemini"]
ENABLE_DEBUG	Enables debug mode. Allows "debug" query to test skill wiring without API calls.	Optional	true or false
MAX_RESULTS	Maximum number of results fetched per provider (when applicable).	Optional	Integer (recommended: 3–5)
MAX_CHARS	Maximum length of returned text sent to the model. Controls token usage.	Optional	Integer (recommended: 800–1500)
TIMEOUT_MS	Timeout for each API request in milliseconds. Prevents hanging calls.	Optional	Integer (recommended: 3000–5000)

Configuration Notes

• Leaving an API key as "" or "key" automatically disables that provider.
• At least one provider should be configured for meaningful results. You know, so you can actually search for someting.
• Lower MAX_CHARS improves performance on-device but may reduce answer quality.
• PROVIDER_ORDER strongly impacts result quality and latency

Repository layout

This repository contains three skill directories:

• search_and_expand
• universal_search
• expand_answer

The repository also includes support files such as .gitignore, .nojekyll, and the license file.

License

MIT License