Describe the solution you'd like
The chat bubble, in combination with the new "Read more..." button, should have a "Hide" or "Hide content..." button to collapse the text after it was expanded by the "Read more...".

Describe the bug
Memories count in memory page is not updated after wipe.

To Reproduce
Steps to reproduce the behavior:

1. Go to the memory page
2. Choose a collection to wipe
3. Click on WIPE

Expected behavior
After the wipe the memories count should be 0.

At the moment when the admin connects to core, it shows the API_KEY modal both if an API_KEY is requested to access the endpoints, but also when there are connection issues (most of the time noobs that misconfigured or not configured at all CORE_HOST in .env file).
Can we give out a proper error message for wrong address?

Another option is to follow @sambarza's suggestion and instead of injecting CORE_HOST to the admin, just make it connect to the same address it has been served (window.location.something...)

Describe the bug
The microphone button doesn't appear in some browsers (like Firefox) and in others does but without working. Actually, I tested it only on Edge.

To Reproduce
Steps to reproduce the behavior:

1. Go to 'Home' page
2. Click on the microphone button (if you see it)
3. Nothing happens when talking

Expected behavior
I expect the microphone to work on every browser.

• Admin detail color is purple
• dark mode does not work (remains light, cat logo goes away, why button is a black circle)

cat.js complains in home page about a missing path attribute

LLM and embedder settings:

• LLM starts on chatGPT instead of DefaultLLM
• There is an error from cat.js about a label field

As suggested by @Mte90,
we can use github workflows/actions to build the admin when a release is created.
In this way:

• this repo remains clean, sources only
• the core does not need nodejs to build the admin

Describe the bug
The full response output looks a bit like a big blob, and harder to digest. See the screenshot for the current output.

To Reproduce
Enable Ollama and LLama2 as the LLM provider

Expected behavior
The output should be well formatted with distinct paragraphs.

Screenshots

As discussed by @zAlweNy26 and @nicola-corbellini , we could have an easy way to customize the prompt from the admin by expanding Chat Settings

1 - chat settings open up in the side panel adn not in the modal
2 - there is also a textarea with prompt instructions (who the cat is)
3 - the prompt is sent via websocket under prompt_settings.prefix

Describe the reason of this issue
Can also totally replace the current library used for dynamic form

The issue is present here:

import { merge } from "lodash"
import { Form, Field } from "vee-validate"

// FEATURE: Improve form to allow any type of input based on json schema
// Can also totally replace the current library used for dynamic form

const props = withDefaults(
	defineProps<{
		fields: SchemaField[]

How to replicate

Under https all endpoints work fine in the admin, with the exception of the endpoint to PUT the LLM configuration. When hitting "Save" on the config:

• browser complains Mixed Content: The page at 'https://DOMAIN/admin/settings' was loaded over HTTPS, but requested an insecure XMLHttpRequest endpoint 'http://DOMAIN/llm/settings/LLMOpenAIChatConfig'. This request has been blocked; the content must be served over HTTPS., By checkin the browser network console, looks like not the request header, but the response header misses https.
• Also a strange redirect happens in the logs serverside
  

Observations:

• no problems with all the other https and wss endpoints
• using the same endpoint from the swagger /docs, gives no problem and the config is correctly saved; also no redirect server side.
• maybe the server side redirect is simply due to the / at the end of the endpoint call

Describe the bug
You can save an empty configuration from LLM/Embedder settings side panel
(look at the Save button)

To Reproduce

1. Select an LLM configuration requiring a field (example openai key)
2. leave the field empty
3. hit the Save button

Expected behavior
Save button is only enabled when required fields are set

It would be nice to have another Import Memories button close to the Export Memories one, in the "Memory" page.

Here is an example proposal (the button color is random):

P.S.: an issue like this, with a wireframe like this può accompagnare solo

Is your feature request related to a problem? Please describe.
When an user copy/pasters a long text in the chat, the backend gives an error (will be fixed) and in any case the message bubble would become huge.

Describe the solution you'd like
Backend should be able to deal with long texts, and the admin just make the message box small so it does not occupy visual space

People is lamenting that the plugin on/off switch "does not work"

Of course we know it is just deactivated, but since it creates confusion let's hyde it until the backend supports activating/deactivating plugins

Describe the reason of this issue
The current code doesn't work, it just imports them and use the latest imported

The issue is present here:

const { isDark } = storeToRefs(store)

watchEffect(() => {
	// BUG: Fix css styling for code snippets in markdown
	// The current code doesn't work, it just imports them and use the latest imported
	if (isDark.value) import("highlight.js/styles/github-dark.css")
	else import("highlight.js/styles/github.css")
})

Describe the bug
plugins page scrolls to the bottom repeatedly if the cat is talking.

To Reproduce
1.) talk to the cat
2.) while the cat is thinking get bored and click on plugins.
3.) after the cat is done thinking he'll start talking and the plugins page will repeatedly scroll to the bottom until the cat is finished.

Expected behavior
I'm not sure I guess... a normally functioning plugins page since the message window isn't visible.

Screenshots
Here's a video instead.
https://drive.google.com/file/d/1CyHpf6qOMuRuinrCoBqkjG0RSHyOkqUW/view?usp=drive_link

Additional context
Add any other context about the problem here.

The input box in home page should autofocus when a message is received

Same thing at admin startup and when there is a change of page to the home page

Note
I was sure this was a backend issue about tools, but I just verified that the websocket message in fact arrives to the admin. Maybe there is some check about content being a string

To Reproduce

1. create a tool with @tool(return_direct=True)
2. tool returns a number return 2
3. message does not end up in the chat

If you return "2", the message is correctly picked

Describe the bug
If a plugin fails to activate due to an error, the admin correctly shows an error notification, but the toggle is set to on anyway.

To Reproduce
Steps to reproduce the behavior:

1. Develop a plugin with a bug
2. Activate the plugin from the admin
3. See the toggle activating but the notification and the terminal raising an error

Expected behavior
The toggle should stay on deactivated

Screenshots

prompt_settings functionality can be fully delegated and expanded from plugins, there is no reason to keep them in core and admin

See cheshire-cat-ai/core#453

To Reproduce
Steps to reproduce the behavior:

1. Go into Memory page
2. Do a search and wait for the points to show up in the plot
3. Click on a memory point
4. Click again

Expected behavior
Should work on the first click

I think should be useful to have the app version info directly on the UI (at least in the admin panel).
The app version could be useful to check if there are updated releases.

EDIT: Solved

Fix: close your fucking tabs!

Describe the bug
Having plugins such the summarization, I see that the GUI freeze whenever an alert arrives, so is basically impossible do chat during the ingestion. I'm using the latest debug branch!

To Reproduce
1 - Install Summarization plugin
2 - Ingest a PDF
3 - Try to chat

Expected behavior
Maybe you should be able to chat during ingestion.

Screenshots
I see a loop of the following screenshots

Admin keeps a cached version of LLM/embedder settings that has side effects on other settings visualization and save/load.

In this case model is not visible while having a clear default in factory.
IF any setting is saved, old config and new one gets mixed (breaking installation)

IMPORTANT: Do not cache settings, as it happens in plugins, make a request every time the panel opens up

Is your feature request related to a problem? Please describe.
Since we have timestamps for memories, it would be nice to have a couple of time filters in the memory page (from / to).

Steps:

1. recall endpoint ignores time filters and the time filtering is done in the admin
2. recall endpoint is able to filter by time interval directly in the backend

Filters (from/to) could be shown as dd/mm/yyyy and handled internally as unix timestamps (which is waaay better)

I suspect this is due to window.location.port is empty if window.location.protocol == 443

See file: https://github.com/cheshire-cat-ai/admin-vue/blob/main/src/api.ts#L12

export const apiClient = new CatClient({
	baseUrl: window.location.hostname,
	port: parseInt((DEV ? '1865' : window.location.port) ?? '80'),
	secure: window.location.protocol === 'https:',
	timeout: 15000,
	ws: {

This is in fact a misuse of window.location.port, see the example below to understand:

Can we find a way to link from the Plugins page the docs about plugin building and publishing? Maybe the template?

Core now gives what is the selected embedder, even if it was automatically selected
In this case:

"selected_configuration": "EmbedderOpenAIConfig"

The chosen Embedder is chosen rightfully, but only on second open

Admin is served from core only to authenticated users.
When the admin loads up, there is a Cookie header with ccat_user_token=xxx

Flow:

• at startup, get the cookie value. If there is no cookie, do everything as usual.
• for http requests, use the header:
  • Authorization: Bearer xxx
• for ws connection, use the url query param:
  • ws://<cat-url-and-port>/ws/?token=xxx

User info:

• user info can be obtained by decoding the JWT
• (show username)

Notes:

• it may be hard to pass the JWT to the admin in dev version

Later issues:

• hide selectively pages according to permissions
• for another issue, when core supports refresh token, refresh it

Is your feature request related to a problem? Please describe.
It is better if the settings use a JSON Schema form generator for better validation and readability.
This can also easily fix #27 .

Describe the solution you'd like
I found this one which seems to be good also for customization (like custom styling).

UPDATE
I tried the library but it doesn't seem to work properly:

Is your feature request related to a problem? Please describe.

As a plus, would be nice to have drag and drop into the chat to ingest documents and URLs instead of buttons only.

Refer to this bug in core:
cheshire-cat-ai/core#589

The core container implements API_KEY to protect from unauthorized requests.
https://github.com/pieroit/cheshire-cat/blob/5795d6bb888d8ed8edc092f87e22ed5336543f61/core/cat/routes/openapi.py#L6

Running the core container without API_KEY is a huge risk because it can leak the OpenAI keys stored in the configuration.

The admin container does not support connecting to the core container specifying the API_KEY. This should be implemented as soon as possible to make the usage of API_KEY mandatory by default.

I think it would be really convenient if the Mathematical Formulas were rendered and displayed correctly, as is the case with ChatGPT. In my case I tested Markdown, KaTeX and LaTeX, and none of the three are displayed correctly.

Additional context

Because EU > US

Is your feature request related to a problem? Please describe.
At the moment the lists are not supported in plugin settings, making it impossible to ask for an array of strings (for example).

Describe the solution you'd like
Support the lists json schema so that it can be used in plugin settings.

Describe the bug
The LLM is returning a numbered list but it is not rendered correctly in the chat window.

To Reproduce

This is the text returned by the LLM:

"\nSure! Here are the steps to prepare Tiramisu:\n\n1. Brew espresso and let it cool.\n2. Beat egg yolks with sugar until pale and creamy.\n3. Add mascarpone cheese and beat until smooth.\n4. In a separate bowl, beat egg whites until stiff peaks form.\n5. Fold egg whites into the mascarpone mixture.\n6. Dip ladyfingers in espresso and arrange them in a dish.\n7. Pour half of the mascarpone mixture over the ladyfingers.\n8. Repeat steps 6 and 7.\n9. Dust with cocoa powder and refrigerate for at least 4 hours.\n10. Enjoy!"

Expected behavior
I expect the text to be rendered as a numbered list.

Screenshots

This is the returned text from the browser developer tools:

This is the screenshot of how it appears, the numbers are missing:

Describe the bug
When trying to get something written in markdown, the ui starts vibrating and shaking in a strange way.

Screenshots

Additional context
In the "Prompt prefix" of the "C.A.T. - Cat Advanced Tools" Plugin I wrote "Always write math formulas in markdown language."

There are endpoints doing possibly long and delicate jobs, for example:

• change of llm
• change of embedder
• plugin install/uninstall
• plugin toggle

Admin should:

• trigger the endpoint
• endpoint will respond immediately with status code 200 (job started)
• admin will poll endpoint with setInterval to check for data change (i.e. plugin is active/inactive or installed/uninstalled)
• admin refreshes accordingly

Summoning @zAlweNy26 & @kodaline with moral support by @pieroit

There are core endpoints doing possibly long jobs, for example:

• change of llm
• change of embedder
• plugin install/uninstall
• plugin toggle
• rabbithole file/url/memories

At the moment their behaviour is inconsistent, as some of them are sync (client gets response when job is finished) and some are async (client gets response immediately and gets a websocket notification when job is finished).

We should devise a common behaviour so clients (and first of all the admin) are offered a consistent interface.
After that we can align the admin so it auto refreshes its stores when such endpoints are finished with their job (e.g. no need to manually refresh after uploading a plugin or a memory)

This issue should be tackled after those endpoints are aligned in core, see here

Currently the memory page shows query results after projecting them in 2D with tSNE.
UMAP could be a faster, more accurate dimensionality reduction algo.

Can you @zAlweNy26 and @nickprock collab on this when possible?
Have fun ;)

Is your feature request related to a problem? Please describe.
Users may be interested in activating or deactivating one or both memories.

Describe the solution you'd like
Send via websocket a json containing not only the text but also prompt settings

{
    "text": "....",
    "prompt_settings": {
        "use_declatarive_memory": false,
        "use_episodic_memory": true,
    }
}

Describe alternatives you've considered
After a conversation in the Discord we reached agreement on these as the most useful options at the moment.

@zAlweNy26 @cristianorevil feel free to redesign wireframe and datastructure as you see fit or involve the community in case you want support

At the moment, the project doesn't have enough E2E and Unit tests to ensure that everything works fine.

Describe the solution you'd like
There should be unit tests for each component and store present.
Also, there should be other e2e tests, that work fine even if the core is needed.

When a user wipes a collection or all of them, both the plot and the counts near the Wipe button are not updated.
If you refresh the page, they are updated

Describe the bug
On Safari in MacOS Sonoma, the "Upload File" button is unresponsive and does not bring up the expected dialog. This is in contrast to the "Upload Memories" document, which appears to function correctly.

To Reproduce
Steps to reproduce the behavior:

On MacOS, using the latest version of Safari:

1. Open the Admin UI.
2. Click the "Lightning Menu"
3. Click "Upload File"
4. Observe nothing occurs.
5. Click "Upload memories"
6. Observe the safari file upload dialog is displayed.

Expected behavior
I'd expect a similar file upload dialog to be displayed for "Upload File", but with a different file filter applied..

Screenshots

Nothing happens

"Upload memories" brings up the expected dialog.

Additional context
The functionality does appear to work on MSEdge on MacOS Sonoma, just not Safari.

Is your feature request related to a problem? Please describe.
There are some TODOs present in the code that should be fixed.

Describe the solution you'd like
There should be a fix for each TODO, also replacing the current workaround code if needed.

In Plugins page, an admin should see active plugins first, then the others

core has been updated to show whether a plugin has a new version from registry

• info is shown for the GET /plugins/ endpoint
• inside an installed plugin data, there is a key upgrade:
  • null if there are no new versions
  • string (e.g. "1.3.0") if there is a new version

Example of not upgradable plugin:

{
      "id": "purrcast_podcast_finder",
      "name": "PurrCast - Podcast Finder",
      "description": "Search for public podcast shows and episodes on Spreaker.",
      "author_name": "Eugenio Petullà",
      "author_url": "https://github.com/EugenioPetulla",
      "plugin_url": "https://github.com/EugenioPetulla/PurrCast_Podcast_Finder",
      "tags": "podcast, multimedia, entertainment",
      "thumb": "https://raw.githubusercontent.com/EugenioPetulla/PurrCast_Podcast_Finder/main/logo.png",
      "version": "1.1.0",
      "active": true,
      "upgrade": null,
      "hooks": [],
      "tools": []
},

Example of upgradable plugin:

{
      "id": "purrcast_podcast_finder",
      "name": "PurrCast - Podcast Finder",
      "description": "Search for public podcast shows and episodes on Spreaker.",
      "author_name": "Eugenio Petullà",
      "author_url": "https://github.com/EugenioPetulla",
      "plugin_url": "https://github.com/EugenioPetulla/PurrCast_Podcast_Finder",
      "tags": "podcast, multimedia, entertainment",
      "thumb": "https://raw.githubusercontent.com/EugenioPetulla/PurrCast_Podcast_Finder/main/logo.png",
      "version": "1.1.0",
      "active": true,
      "upgrade": "1.2.0",
      "hooks": [],
      "tools": []
},

Endpoint to update and checks to delete the older version and upload the new are not yet there.
We can just use the already present endpoint to install plugins from registry and core can do the checks, so admin side you can just get the plugin URL and use the endpoint POST /plugins/upload/registry

Describe the reason of this issue

The issue is present here:

// BUG: Fix `resetAllStores` and remove the workaround
router.beforeEach(async to => {
	if (!isReadyAndAuth.value && to.name !== 'settings') {
		return { name: 'settings' }
	}
})

Describe the bug
Using the web UI there is a problem to with the link to the documentation

To Reproduce

1. docker-compose up
2. Go to local host
3. Click on docs

Expected behavior
The link should be https://cheshire-cat-ai.github.io/docs/ instead is https://cheshire-cat-ai.github.io/core/