Introduction
This is the eighth post in the series. Here is how the series got here.
- Part 1: I built the base of a file portal on top of Amazon FSx for NetApp ONTAP (FSx for ONTAP below) S3 Access Points.
- Part 2 on: I added management screens for snapshots, SnapLock (which makes files impossible to change or delete), ARP/AI (ransomware detection), FlexCache and more.
- Part 7: I compared three tools for building the portal itself.
This post does not add a feature. It is about replacing the emoji on this file portal's screens with line icons. No product update triggered it.
It covers two web apps. The file portal is built on AWS Amplify Gen 2, and the permission-aware RAG app in the second half is built on Next.js. I ran the same detector on both and replaced the emoji with line icons. At the end I also set out what GenU (Generative AI Use Cases JP), another generative AI web app on AWS, has in common with them and where it differs.
What started it was a skill for building app screens that do not look AI-made, published by minorun365 (unslop-ai-app), and the post on X that introduced it. The post compares a screen left entirely to AI with a screen built after the skill was read.
The repository collects rules for keeping a web app screen from looking AI-made, plus a script that checks the screen source (tools/check-ui.mjs). Its README says that when AI builds a screen, it tends to add parts nobody asked for and to line up cards colored on one side only and emoji icons. My file portal had emoji icons lined up in just that way. So I ran the same checking script on my own screens and replaced the emoji it flagged.
The portal used emoji for navigation, buttons, headings and status (success, warning, locked). That was quick at the start. As the management screens grew, two things started to bother me.
- An emoji's shape comes from the font of the OS or browser. The Unicode emoji specification (UTS #51) also says emoji images can vary significantly between implementations. The same "⚠" is a different picture on a Mac, on Windows and on Android.
- An emoji's color does not follow the text color. UTS #51 says a character shown in emoji presentation does not change color when the text color changes. The portal has three themes (light, dark, follow the system), and the emoji stayed outside all of them.
On a storage management screen, a status such as "locked (for example a volume locked with SnapLock)", "warning" or "failed" feeds straight into the next action. That status looking different from one device to the next is not good for a portal.
This post covers:
- Before / After for 10 screens
- How the 585 emoji were replaced screen area by screen area in this portal's repository
- How stored data and the translation dictionaries were handled
- Why a script that checks the screen text (the detector) went in first
- What happened when the same detector ran on the RAG app (a permission-aware web app): 440 replacements
- Commonalities and differences with GenU (Generative AI Use Cases JP)
What this post does not cover:
A hands-on comparison with GenU (the section is based on its README and documentation; I did not deploy it)
How the icons themselves are drawn (SVG path design; the icons are built like common line icons)
Handling the findings other than emoji (spacing steps, wording and so on; they need a separate visual and wording decision)
Checking with a screen reader (button names were added, but nothing was tested with assistive technology)
Terms used here
- Line icon: a single-color icon drawn as outlines only. Here it is SVG whose color follows the surrounding text color (
stroke="currentColor", see MDN). - Translation dictionary: the files holding the portal's UI text in eight languages (Japanese, English, Korean, Simplified Chinese, Traditional Chinese, French, German, Spanish). The Japanese file is the type source, and the other seven follow it.
- Detector: a script that finds template-like patterns in UI source, such as emoji or
transition: all, and counts them per rule. - Sandbox: an AWS Amplify Gen 2 development environment. One full set of the portal's backend (Lambda, AppSync, Cognito) is deployed in it.
What was replaced, and how many
The first step was to point the detector at the UI source and count. The detector is check-ui.mjs from unslop-ai-app, brought in under the Apache License 2.0.
Across 122 files it raised 658 findings, 590 of them emoji.
Not all 590 were on screen, though. Five sat inside code comments and were false positives, which leaves 585 strings that render. The 191 in the translation dictionaries render through t("key"), so they count.
The 585 broke down as follows.
- Translation dictionaries: 191 (23 to 24 strings per language, across eight languages)
- General components and navigation: 262
- Management screens (under
admin): 132
The most frequent were ⚠ (31), ✅ (28), 🔒 (27) and ✕ (24). All of them mark a status or an action, exactly the places that should follow the theme.
Before / After
Each pair was taken on the same sandbox, the same data and the same viewport (1512 × 805; phone 390 × 844). Before was served from the commit just before the change. Names and addresses on screen were replaced with generic ones just before each capture.
File list and sidebar
All 18 sidebar items were replaced. In Before, every item shows a picture of a different color, and on the selected item (blue background) the emoji keeps its own color. In After, the icons take the text color, which turns white on the selected item.
Dark theme
The dark theme shows the largest difference. An emoji carries its own colors, so on a dark background its brightness does not match the text around it. A line icon uses the text color, so it changes along with the theme.
Snapshots
The filter labels used to live in the translation dictionary with the emoji inside the string, such as "🔐 Tamperproof" and "📅 Scheduled". The emoji came out of the dictionary, leaving only the words, and the icon now sits in the screen component. The reason is in "Translation dictionaries" below.
Lock
Before assigned 🔒, 🪣 and 🔐 to the three mechanisms. In some fonts 🔒 and 🔐 are hard to tell apart, even though they stand for different mechanisms. After uses a lock, an archive box and a lock with a keyhole.
ARP/AI
Incident states (detected, contained, investigating, resolved) used the 🔴 🟠 🟡 🟢 circle emoji. Now a small CSS dot sits before the text label, in the theme's success, warning and error colors. The label carries the meaning; the dot only adds color.
Resource management
This screen shows 23 cards. In Before, each emoji had its own color and visual weight, so which card stood out depended on the emoji's color rather than on the card. In After, stroke width and color match, and the card names and descriptions read first.
FlexCache management
Each row had four emoji: origin (📦), cache (⚡), metrics (📊) and write mode (✍). They are supplementary details, yet their colors made them the loudest part of the row. In After they share the gray of the supplementary text.
AI settings
First-run welcome
The larger an icon is drawn, the more the font differences show. The Mac font draws the folder in 3D, which stood out against the flat screen around it.
Phone menu
Approach: splitting 585 changes by screen area
Replacing everything at once would make one diff across 78 files. That cannot be reviewed, and when something breaks there is no telling where. So the work was split by screen area in the order below, and each step was required to lower the detector's emoji count.
- The icon component and the navigation (#520): a component drawing 20 icons in SVG, used in the sidebar
- 39 general component files (#521): the icon set grows to 79, plus a status-dot component
- 26 management screen files (#522)
- The translation dictionaries in eight languages, and the icons on the screens that show those strings (#523)
The count went 585 → 562 → 323 → 191 → 0.
The icons are hand-written SVG rather than an icon library. The set is 79 icons, and every added dependency is one more thing to keep updated. That said, if the set keeps growing, a library would be easier to maintain.
What needed care
Stored data
The AI agent feature stores an icon the user picks for each agent. Existing agents hold an emoji string in that field, and the backend's default is an emoji too.
Rewriting the stored data would turn this into a data migration. Instead, the display side draws both an icon name and an earlier emoji. A table maps the 12 earlier emoji to line icons, and newly created agents store an icon name. Neither the data nor the backend changed.
Translation dictionaries
The dictionaries held strings like "🔒 Lock" with the emoji inside. To show a line icon, the emoji has to leave the dictionary and the icon has to sit in the screen component, because a dictionary string cannot hold SVG.
The eight dictionaries have a check on their key structure, so no key was added or removed. Only the wording changed.
Some strings, though, described the emoji itself. For example: "No favorites yet. Click the ☆ on any file to pin it here." Removing the ☆ alone leaves a broken sentence, so it became "Click the star icon", and the same fix went into all eight languages.
Some places cannot hold an icon at all, such as a confirmation dialog (window.confirm) or the inside of an <option>. There the emoji was removed and only the words remain.
Names for icon-only buttons
A button that showed only an emoji ("✕" to close, "🗑" to delete) used the emoji as its accessible name. The line icons are decorative (aria-hidden), so those buttons would be left without a name. Each icon-only button got an aria-label from an existing translation key.
Why the detector went in first
The detector was brought into the repository before any replacement (#519). Without a count, finding what is left means searching the screens by eye.
The detector also raises findings other than emoji (transition: all, italics, spacing steps and so on). This post does not cover them.
The same replacement in the RAG app
Everything so far was about the file portal. The detector is not specific to the portal. I ran the same script, unchanged, on my other project, the RAG app (FSx-for-ONTAP-Agentic-Access-Aware-RAG).
That Next.js-based web app takes the search results from Amazon Bedrock Knowledge Bases, keeps only the documents the signed-in user may read, and answers with citations.
How it differed from the portal
The detector checked 686 files and raised 633 findings, 519 of them emoji. Not all 519 render on screen, though. Excluding 59 in test code, 5 in a sample component nothing imports, 12 in comments and 3 in copies of compiled .js files leaves 440 emoji that render.
Three things differed from the portal.
- Where the icons come from: in the portal I drew 79 line icons as SVG myself. This web app already had the icon library lucide-react as a dependency, so I used it. That is the setup I said in the portal's approach section would be easier to maintain if the icon set kept growing.
- Data that holds an icon as a string: card and status definitions held an emoji string, such as
icon: '🔍'. Unlike the AI agents in the portal, this was not stored data but a definition in the source code, so no table mapping earlier emoji was needed. I changed the type fromstringto the icon component type (LucideIcon), and the display side now draws<card.icon />. - Places that cannot hold an icon: Markdown strings in chat responses,
alertand error messages cannot hold an icon component. There I removed the emoji and kept the sentence. Where a leading symbol alone told success from failure, a word now says it (for example "Warning:"). Icon-only buttons got anaria-label, as in the portal.
The replacement went into one pull request (#247), split into four commits by directory: icon data and shared code, the Bedrock settings screens and the app skeleton, agent management and shared components, and chat and the sidebar.
Retaken screenshots
The repository's README uses six screenshots. I retook three of them: the KB-mode card list, a response with citations and the automatic model selection. They were taken on a deployed environment, signed in and asking real questions.
The Agent mode and multi-agent mode screenshots were retaken too. Both show a question asked after signing in as the legal department's demo user.
The Registry tab of the agent directory lists the records of AWS Agent Registry. AWS Agent Registry moved from the bedrock-agentcore namespace to agent-registry on 2026-08-06, and the old namespace shuts down on 2026-10-30 (migration guide). The app uses the SDK for the new namespace. The screen shows three sample records in a test registry.
Commonalities and differences with GenU
GenU (Generative AI Use Cases JP) is a generative AI web app sample that AWS publishes on GitHub (aws-samples/generative-ai-use-cases, MIT-0).I line it up here with the two apps in this series, as another generative AI web app on AWS.
The content comes from the README (read again on 2026-11-19) and the deploy-options document (as of 2026-11-14).I did not deploy GenU and compare it hands-on.
Commonalities
- Built from AWS services: GenU and the RAG app use AWS CDK for the backend, and the file portal uses Amplify Gen 2.
- React-based screens: GenU uses React and Vite, the RAG app uses Next.js (a React framework), and the file portal uses React.
- Amazon Cognito for sign-in: GenU can also federate an external IdP over SAML.
- Amazon Bedrock Knowledge Bases for RAG: S3 Vectors is an option for the vector store in both (
ragKnowledgeBaseStorageTypein GenU; S3 Vectors is the RAG app's default). - Units of work as the entry point: GenU lists use cases by kind of task. The RAG app and the file portal list industry cards and templates.
- Published as source, meant to be customized: all three publish their source on GitHub. They are meant to be adapted, by choosing features and adding screens and processing for your organization, not only run as they are.
Differences
The two apps now have most of what GenU has. The table gives the overall picture, and the sections after it give the details.
Feature table
The numbers are pull request numbers; the links are in the sections below. "Added" means merged into the repository's main branch.
| Feature | GenU | RAG app | File portal |
|---|---|---|---|
| Translation, proofreading, diagram generation, web extraction | Yes | Added (#272) | Added (#530) |
| Video generation, video analysis, meeting minutes, code execution | Yes | Added | Added |
| Image generation (from text) | Yes | Added (off by default) | Added (off by default) |
| Image from image, prompt help | Yes | Added (#290) | Added (#562) |
| Choice of minutes style | 3 styles | 5 styles (#289) | 4 styles (#558) |
| Real-time transcription for minutes | Yes | No (an options document only, #299) | No |
| Saving, sharing and JSON for use cases | Yes | Added (#286) | Added (#560) |
| Saving ratings of answers | Not in the README | Yes (#283) | Yes (#561) |
| Reranking | Yes | Added (#288) | No |
| Query decomposition | Yes | Added (#297) | No |
| Advanced Parsing | Yes | Added (#298) | No |
| Importing Bedrock Flows | Yes | Partly (only flows an administrator registered, #294) | No |
| MCP chat where users add MCP servers | Yes | No (left out on purpose, #293) | No |
| External MCP tools | Added by users in MCP chat | One: AWS documentation (#296) | No |
| Voice chat, web search agent | Yes | Yes (feature flags) | No |
| SAML federation | Yes | Yes (feature flag) | No (an example only) |
| PWA (installable from the home screen) | Yes | Added (#300) | No |
| Browser extension | Yes | No | No |
What the two apps have
- GenU's default use cases are split by kind of work.
- They are chat, text generation, summarization, meeting minutes, proofreading, translation, web extraction, image and video generation, video analysis, diagram generation and voice chat.
- There are a web search agent and a Code Interpreter agent.
- You can import agents built with Bedrock Agents, AgentCore Runtime and Bedrock Flows.
- The RAG app lists work grounded in internal documents.
- It has 14 general KB-mode cards and 21 industries with 3 cards each (find, draft, check).
- It has an agent mode and a multi-agent mode, and the agents run on Amazon Bedrock AgentCore Runtime.
- Agents can be created, edited and deleted from the agent directory on screen. Changes to shared agents, teams and the like are limited to administrators in the Cognito group
admins(#285). - The agents' document search goes through the same permission filter as KB mode.
- In multi-agent mode, a supervisor agent calls the retrieval, analysis and output agents as tools.
- Feature flags enable the following:
- AgentCore Gateway, Memory and Policy
- Step Functions workflows exposed as agent tools
- web search, voice chat and Guardrails
- multimodal search with images
- The file portal is a screen for working with files on FSx for ONTAP.
- You can browse and upload files, and share them with expiring links (5 minutes to 1 hour) or QR codes.
- The AI features are:
- AI processing that starts the industry processing patterns UC1 to UC28 and the operations patterns
- an AI chat that uses tools on files
- an agent directory, multi-agent teams, and 16 industry agent templates
- semantic search over a whole volume
- Management screens for snapshots, SnapLock, ARP/AI and more are in the same portal.
Tools added first
- Translation, proofreading, diagram generation, web extraction, image generation and saved use cases are now in both apps (RAG app PR #272, file portal PR #530). Saved use cases are close to GenU's Use Case Builder.
- Video generation, meeting minutes, video analysis and code execution are in both as well. Code execution is close to GenU's Code Interpreter.
- When internal documents are the input, only the documents the user can read are passed to the model. A test checks that no trace of an unreadable document reaches the model's input.
Meeting minutes
- GenU makes minutes from a recording or from real-time transcription. You choose a transcript close to what was said, a newspaper article, or an FAQ.
- The file portal offers minutes, a transcript close to what was said, an article, and an FAQ (#558).
- The RAG app offers minutes (the default), a cleaned-up transcript, a summary, Q&A, and the user's own instruction of up to 1000 characters (#289). Nothing in it corresponds to GenU's newspaper-article style.
- In both apps the finished transcript stays for a day. Redoing it in another style does not call Transcribe again.
- Real-time transcription is in neither app.
- For the RAG app I only wrote a document that sets the options side by side (#299).
- It compares six: three ways of using Amazon Transcribe streaming, Nova Sonic, the browser's built-in recognition, and keeping the current approach.
- The points compared are authentication, where the audio goes, what to run, limits and cost.
- Which one to choose is not decided.
- Nova 2 Sonic's official language list has no Japanese. The connection limit is 8 minutes.
Image generation
- GenU generates images from text or from an image. An LLM can help write the prompt.
- Both apps now have generating from an image and prompt help (RAG app #290, file portal #562).
- Starting from an image calls SD3.5 Large, not Stable Image Core, which takes text only. Among the Regions I checked, SD3.5 Large is only in us-west-2.
- SD3.5 Large needs its AWS Marketplace subscription enabled once in the account.
- Image generation is off by default in both apps.
- As far as I checked on 2026-11-14, Tokyo has no model that outputs images.
- Of the Regions I checked, text-to-image exists only in us-west-2.
- When it is on, it calls us-west-2, so the prompt and the image are processed outside Japan. The screen says so.
- Luma Ray2, used for video generation, is also only in us-west-2 among the Regions I checked. It is off by default like image generation, and the screen says the data is processed outside Japan.
Sharing use cases
- GenU can share a use case made in the Use Case Builder with every user who can sign in. It can export and import JSON.
- File portal (#560):
- Sharing with every user and JSON export and import are added.
- The export has the same shape as GenU's Use Case Builder. A use case exported from GenU can be imported when it has one input field.
- Only the author can change it. It is not shared with external users.
- Any role, including the viewer role, can create a use case, because it is the user's own template and writes nothing to files.
- Saving a result to a folder needs the writer role.
- RAG app (#286, feature flag
enableUseCaseSharing, off by default):- Sharing with every user and JSON export and import are added.
- A shared use case that uses internal documents searches with the permissions of the user running it.
- The owner's ID is not returned to other users. Only the owner can change it, and an imported use case is private.
- The JSON is the RAG app's own format (
rag-use-cases). Compatibility with GenU was not checked.
- Chat histories cannot be shared in either app.
Stronger Knowledge Base search
GenU lets you choose Advanced Parsing, query decomposition and reranking. The RAG app has all three now, each off by default. The file portal has none of them.
- Reranking (#288,
enableRerank)- It reorders with the Bedrock Rerank API.
- Retrieve's built-in reranking runs before the permission filter and would send fragments of unreadable documents to the reranking model, so it is not used. This one runs after the filter.
- Query decomposition (#297,
enableQueryDecomposition)- A text model splits a question that contains several topics into up to 4 search queries.
- Bedrock's own
QUERY_DECOMPOSITIONanswers before the filter, so it is not used. - The model sees only the question, never a document.
- Advanced Parsing (#298,
kbParsingStrategy=BEDROCK_FOUNDATION_MODEL)- It reads PDFs with a foundation model.
- Every PDF in the data source is read by that model and charged by tokens.
- What else the RAG app lets you choose is the chunking strategy and the search type (hybrid or semantic).
Flows and MCP
- MCP chat, where users add external MCP servers, is in neither app.
- For the RAG app I sorted out the threats first (#293). There are ten items, such as an instruction written in a retrieved document or a tool response making the model call another tool, and a tool changing later.
- Bedrock Flows can be called from the screen only when an administrator registered them in the CDK context (#294).
- Before each call the nodes of the version the alias points to are checked.
- Only flows made of input, output, prompt, condition, loop and similar nodes pass.
- A flow with a knowledge base, agent, S3 or Lambda node is not called.
- The window between the check and the call, in which the alias could be repointed, is recorded as a remaining risk.
- For an external MCP server I added one: the AWS Documentation MCP Server, version pinned (#296).
- It searches and reads AWS documentation, and runs in my own account's AgentCore Runtime.
- The reasons and the comparison with other candidates are in the document of #295.
- It has three fixed tools and refuses any URL outside
docs.aws.amazon.com. - Only an agent whose definition lists
AwsDocumentationcan use it.
- The RAG app connects to AgentCore Gateway over MCP, creates Bedrock Agents and calls AgentCore Runtime.
- The file portal can neither import an agent built elsewhere nor use Flows.
PWA and the browser extension
- I added a PWA to the RAG app (#300). The file portal has none.
- The service worker caches only what is identical for every user: build output, the icons and an offline page.
- It does not store API responses, authentication or pages (HTML). If the next person uses the same browser, the previous user's questions and answers are not left on the device.
- GenU's browser extension is in neither app.
Time limits in the file portal
- The file portal's screens go through AppSync, so one call is cut off at 30 seconds.
- So the slow minutes (Amazon Transcribe) and video generation are split into two calls: start the job, then check its status.
- Code execution is kept to 25 seconds in total, and video analysis to 50 MB, the most it can answer in one call.
- For minutes and video analysis, the S3 Access Points alias is passed straight to Transcribe and Bedrock. The file is not copied to another bucket.
- The model used for video analysis (Amazon Nova 2 Lite) looks only at the video frames and does not process audio, as AWS's documentation states.
- What was said is handled by the minutes.
Collecting user ratings and improving from them
- GenU's REA...


























