These automations and macros enhance mindmaps created by MindManager on macOS and Windows.
More animated examples are in the doc folder.
- Azure OpenAI w/
GPT-4o(use your key or log in withAzure EntraID) -> best tested - OpenAI w/
GPT-4o,o1-preview,o1-mini(use your key) -> best results - Anthropic w/
Claude 3/3.5(use your key) - xAI w/
grok-beta(use your key) - Google
Geminiw/Pro,Pro-ExpandFlash(use your key) - Google Vertex AI w/
Gemini ProandGemini Flash(use your access token / OAuth2) - DeepSeek w/
DeepSeek V2.5(use your key) - Alibaba Cloud w/
Qwen-Max,Qwen-Plus,Qwen-Turbo(use your key) - Mistral AI w/
Mistral-Large,Pixtral,Ministral - AWS Bedrock w/
Titan Text G1 Premier(native),Anthropic,Mistralmodels (use your key + secret) - Azure AI Foundry (platform) w/
LLaMA,Mistralmodels (use your key) - GitHub Models (platform) w/
GPT-4o,LLaMA(use your key) - Perplexity (platform) w/
LLaMA,Mixtral(use your key) - Groq (platform) w/
LLaMA,Mixtral,Gemma2(use your key) - Hugging Face (platform) w/
LLaMAand more models (use your token) - Fireworks AI (platform) w/
Qwen 2.5,QwQ(serverless) and more models (use your key) - OpenRouter (platform) w/
o1-preview,o1-miniand many more models (use your key) - Ollama (local) w/ any LLM (use
LLaMA,ZephyrorMixtralmodel for best results) - LMStudio (local) w/ any
llama.cpporMLXmodel - MLX (local w/ Apple Silicon) w/ any LLM (use
LLaMA,Qwen2.5model for best results)
- Azure OpenAI w/
DALL-E 3(use your key or log in withAzure EntraID) -> best tested - OpenAI w/
DALL-E 3(use your key) -> best results - Stability AI w/
Stable Diffusion 3SD3.5/SD3/Ultra/Core(use your key) - Google Vertex AI w/
Imagen3(use your access token / OAuth2, GCP approval required!) - Ideogram AI w/
V1/V2(use your key) - Black Forrest Labs w/
Flux Pro 1.1 Ultra,Flux Pro 1.1,Flux.1 Pro,Flux.1 Dev(use your key) - Recraft AI w/
RecraftV3,Recraft20B(use your token) - MLX (local w/ Apple Silicon) w/
Flux.1,SD3
- DeepL (use your key)
- Windows compatible (run macro/context menu or call the Python script directly)
- macOS compatible (run Automator workflow (Quick Action) or call the Python script directly)
- Only native API requests to AI systems - no middleware needed
- Map format can be radial map or orgchart
- Using map templates on macOS
- Map styles on Windows are persistent, automatic collapsing of nodes
- Refinement of the map or topic.
- Refinement of the map or topic from a development perspective.
- Create examples for one, more (selected) or all topics.
- Clustering topics from scratch.
- Clustering by one or more criterias eg. Organization/Process/Project/Expertise, Capex-Opex perspective.
- Complex cases (multiple calls): eg. refinement + clustering + examples.
- Image generation from topics
- Professional translation of all topics by DeepL
- Export
Mermaidmindmap HTML document - Export
Markmapmindmap HTML document - PDF to mindmap (multiple files / batch processing)
- Generate a working paper (argumentation) HTML document for a detailed mindmap
- Generate a glossary HTML document of all terms
- Export mindmap to any other text format
- Change map layout by using a template (macOS)
- Reorder topics by business value or importance
- Misspelling or syntax correction
- Create a map based on external text data
First install the Windows package manager Chocolatey from an administration shell or choose any other way following https://chocolatey.org/install.
This is one line:
Set-ExecutionPolicy Bypass -Scope Process -Force; [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072; iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
Change to folder %localappdata%\Mindjet\MindManager\23\macros:
cd %localappdata%\Mindjet\MindManager\23\macros
Copy all files from the GitHub repository to this location.
Change to windows folder:
cd windows
Run install.bat or the following commands (requirements_auth.txt only if you want to use Azure Entra ID or GCP OAuth2):
choco install python3
pip install -r .\..\requirements.txt
pip install -r .\..\requirements_win.txt
pip install -r .\..\requirements_auth.txt
powershell -ExecutionPolicy Bypass -File .\macro_registration.ps1
Check in registry and MindManager, if the macros are available (right click on topic).
Hint: The macro list is ordered according to the GUID-string, not the macro name.
Macros can also be executed by the macro editor. The macros are similar but the action parameter.
You can also check here if the path to the python file is correct.
Python has to be installed first. Go to https://www.python.org/downloads/macos/ and download the desired installer.
Install required python libraries (requirements_auth.txt only if you want to use Azure Entra ID or GCP OAuth2, requirements_mac_mlx.txt is only needed for local image generation using MLX):
pip install -r requirements.txt
pip install -r requirements_mac.txt
pip install -r requirements_mac_mlx.txt
pip install -r requirements_auth.txt
Create the directory structure ~/git/mindmanager_ai with Terminal:
cd ~/
mkdir git
cd git
mkdir mindmanager_ai
cd mindmanager_ai
Copy all repository files to this location as the Automator workflows contain this path.
Alternatively you can clone the repository in Terminal:
cd ~/
mkdir git
cd git
git clone https://github.com/robertZaufall/mindmanager_ai.git
cd mindmanager_ai
Change to folder macos and copy the Automator workflows to the ~/Library/Services (hidden) folder:
cd macos/automator
chmod +x ./copy_to_services.sh
./copy_to_services.sh
If you need elevated privileges for copying the files use this command:
sudo sh ./copy_to_services.sh
All Automator workflow settings are similar but the action parameter:
The workflows are then available at the "MindManager" main menu -> Services
I prefer to execute the python script directly from VSCode. Here you can easily adjust the settings, try different LLMs on the fly and even debug, if problems occur (external systems are sometimes not available).
There are some actions already predefined for quick execution.
First of all you have to open the config.py in a text editor of your choice.
Don't be overwhelmed by the huge list of variables. Just pick the lines you need and set its values.
Use the apropriate LLM system for which you have an API key. These keys are available on the developer platforms of the AI vendors.
If you want to run local models with Ollama or GPT4All you have to have either a newer Apple Mac model with M1, M2, M3 processor or a desktop or notebook with NVidia graphic card with at least 8GB graphic ram.
You can have more than one open document in MindManager. The document which should be processed must be the active document. For every processing a new document with the new topics will be created.
To process the whole map, select the central topic (for right-clicking) or no topic at all (call macro manually (Windows), choose Automator Workflow from Mindmanager Menu -> Services or call the python script from VSCode or commandline python3 process.py <action> <format>.
Select the central topic or deselect all topics and call the automation.
You can also select one or more topics and start the automation for just these topics, e.g. to generate examples for these topics, refine just these topics etc.
Just select the topics for which you want to generate an image and choose the action "Generate Image" (macro on Windows or Automator Workflow on macOS) or call the Python script with parameter image or image_n.
After a while, the image will be opened and also stored in the MindManager-Library Images-folder.
Unfortunately, on macOS the image cannot automatically be inserted into the map or added to a topic due to insufficient library support.
On Windows the image can be automatically set as the background image of the map.
The results from the generation process are best with FLUX.1, good with DALL-E 3 and SD 3.5. Prompt crafting/engineering is still in progress.
The filename is enriched with the generation seed where this feature is supoported. This seed is useful if you want to generate similar images (e.g. with different prompt). DALL-E 3 does not support a seed value anymore (by the time of writing).
The prompt for image generation can optionally be optimized using a LLM call.
Images can also be generated locally on macOS with Apple Silicon using the native Apple MLX framework.
Recently there are more image generation plattforms trending. Black Forrest Labs, Ideogram AI and Recraft AI image generation from mindmaps is already implemented and the results are amazing.
Put the files into the input-folder and use the action pdf_mindmap. The PDF files are first converted to markdown (MD) format. 'Reference' sections are removed as these contain no information but take a lot of tokens (e.g. arXiv papers). No OCR takes place by now. Tables are removed and the content will be highly sanitized by removing irrelevant characters, code blocks, href-links, whitespace etc.
Still lots of input-tokens are needed in order to summarize the text by the LLM. These models have been tested and are working by now: GPT-4o, o1-preview, o1-mini, Gemini 1.5, Claude 3.5 Haiku and Claude 3.5 Sonnet.
There is no local LLM model using Ollama working for me by now.
Sonnet 3.5 lately supports native PDF processing, which is also implemented (action pdfsimple_mindmap).
Generation of larger text outputs needs a model with an higher max-token value like GPT-4o, Gemini Flash, Sonnet 3.5. Results are very good, most of the time.
The solution is best tested with Azure OpenAI. Results are perfect for every use case. Execution time is quite fast using the newest o1, GPT-4o models. Azure EntraID authentication can be used in enterprise scenarios.
Gemini Pro results are good. Gemini Flash does (most of the time) only generate up to 3 levels at max, so a refinement does currently not work.
Vertex AI needs an access token which has a default expiration time of just 1 hour (authentication flow is partly implemented).
The newest Anthropic Claude 3.5 Sonnet model seems to be best in class. Anthropic Claude 3 Opus ist also very good but little bit expensive. Anthropic Claude 3.5 Haiku is good and cheap.
Grok is very good and is able to refine mindmaps for several levels. By today (2024-10-21) only the grok-beta model is available in public preview.
DeepSeek created an extraordinary open source model DeepSeek V2.5 which seems to be as good as GPT-4o.
Alibaba Cloud models cannot generate large amounts of tokens (Qwen-Max: 2000, Qwen-Plus + Qwen-Turbo: 1500) but the results are good. Qwen-Turbo is very fast. Qwen 2.5 model is still not available outside China by now (2024-11-22).
Mistral AI is hosting their commercial flagship models Mixtral-Large and Pixtral-Large. Mixtral-Large is a 'best in class' model. The maximum numer of possible output tokens is a little bit unclear (max_tokens may meant to be the sum of input and output tokens).
Groq is sure the fastest LLM platform by now. LLaMA3, Mixtral and Gemma2 are proven models.
Perplexity works perfect as an universal LLM platform.
To access better models a pro-subscription is needed. LLama-3-8B still can be used.
On the Open Router platform there are a variety of models and systems available. Also fallback scenarios are supported. Furthermore you get access here to the newest OpenAI models like o1-preview.
Results are dependent on the used model. LLaMA3, Zephyr and Mixtral are working well.
MLX results are dependent on the used model. LLaMA3 works well.
The solution is best tested with Azure OpenAI. Results are very good. There is a problem with texts generated within images. Azure EntraID authentication can be used in enterprise scenarios.
Image generation with SD3.5 and SD3 is the most flexible, as you can use a seed value, negative prompt, etc. Prompt engineering is most important here, as the results are far from being perfect by now.
Image generation results are too simple by now as prompt engineering is also most important here. Imagen3 has the highest image resolution (1:1 with 1536x1536). Imagen3 is GA (globally available) but there is an approval process to get access to the API.
Image generation is quite good using the V_2 model. When activating API access, keep in mind that generating an API key immediately results in a $40 bill.
Image generation is extraordinary. Flagship model is Flux Pro 1.1 / Ultra. As usual token have to be prepaid and you need accepted for accessing the platform.
Image genration is very good. There are many pre-defined styles which can be activated as needed. Available Models are RecraftV3 and Recraft20B.
This local image generation alternative is only available on macOS with Apple Silicon processors like M1 and higher. The results are above average using Flux.1 model and under average using SD3 mostly because the prompt is optimized for Flux.1.
There is a new action defined (image_n eg. image=10) to generate a bunch of images in a row. A pre-executing step can be added to optimize the prompt using a LLM call. If there is only one topic selected there is a different prompt used as when more topics are selected. Only the first level of topics together with the central topic should be selected for better results.
When using this image generation way, the desired model and embeddings tokenizer will be downloaded automatically. The expected data amount to be downloaded is about 50GB using Flux.1 and 6GB using SD3. If you are using SD3 for the first time you have to login at huggingface with your token first as you have to agree to the terms of Stability AI and the usage of their model: huggingface-cli login --token <xyz>. Downloaded models are cached at ~/.cache/huggingface.
Translation works for these languages:
# supported languages as source
# BG,CS,DA,DE,EL,EN,ES,ET,FI,FR,HU,ID,IT,JA,KO,LT,LV,NB,NL,PL,PT,RO,RU,SK,SL,SV,TR,UK,ZH
# supported languages as target
# BG,CS,DA,DE,EL,EN-GB,EN-US,ES,ET,FI,FR,HU,ID,IT,JA,KO,LT,LV,NB,NL,PL,PT-BR,PT-PT,RO,RU,SK,SL,SV,TR,UK,ZH
Source language will be detected automatically. Formality parameter is not supported for all languages, so it is disabled by now. Context parameter was not used as DeepL states it's deprecated.
API requests point to the free tier. If you have a paid subscription change the URL in the config.py.
Prompt crafting is lightly implemented using the following strategy:
MindManager COM objects are addressed by using the PyWin32 library:
MindManager objects are addressed by using the AppScript library:
The Mermaid mindmap syntax is used when talking to the OpenAI LLM as an intermediate "language". Log file contents for input, output, prompt can be used in other use cases eg. mindmap visualizations in GitHub markdown files.
Log files content:
Example using a Mermaid mindmap in a GitHub markdown file.
Code:
```mermaid
mindmap
Creating an AI Startup
Market Research
Identify Target Audience
Analyze Competitors
Understand Market Trends
Assess Market Needs
Evaluate Market Size
Business Model
Define Value Proposition
Choose Revenue Streams
Plan Monetization Strategy
Identify Cost Structure
Determine Key Partnerships```
Github rendering of the map:
mindmap
Creating an AI Startup
Market Research
Identify Target Audience
Analyze Competitors
Understand Market Trends
Assess Market Needs
Evaluate Market Size
Business Model
Define Value Proposition
Choose Revenue Streams
Plan Monetization Strategy
Identify Cost Structure
Determine Key Partnerships
You can also use the content inside the Mermaid online editor (https://mermaid.live/edit):
The API execution time depends heavily on the used LLM model or system and token count.
Currently, this project is in the early development phase, and generated outputs may include errors. Automated testing has not yet been implemented.