AI Mastery Open-Hours – Knowledge Management, Tana, Mind-Mapping & Automation Q&A

Intro & Opening Remarks

• Session starts with host apologising for late start; informal greetings; participants show venues (e.g. Mikis’ party space in Greece).

• Light banter: haircuts, age comments, positive energy.

Routine House-Keeping & Q-and-A Workflow

• Participants instructed to drop questions in Zoom Q&A panel.

• 1-minute countdown, then 20-second extension to type questions.

• Up-vote period: 2 minutes for prioritisation; highest votes answered first.

• 6–7 questions collected; Tracy’s wins with 11–12 up-votes.

Question 1 – Tracy: OpenAI → Tana Automation Formatting

• Automation: class transcript → OpenAI summary → HTTP module → Tana.

• Problem: summary arrives as a single block, wants paragraphs.

• Advice:
  • Check "Tana Paste" markdown spec (official doc, subtle differences vs. common MD).
  • Copy/paste from ChatGPT creates new nodes on each newline ⇒ API should mimic this.
  • Action items: read spec, test, post results; host will test too.
  • API still young – possible limitation.

Mini-Demo: Copying Content Out of Tana

• Right-click node → Commands → Copy as plain MD.

• Pasted into Gmail/Notepad shows Tana-specific markdown (good for automation).

Tana Deep-Dive

• Category: hybrid of Airtable (structure) & outliners (flexibility).

• Strengths:
  • Create tables (super-tags) from anywhere; no need to navigate bases.
  • Objects are nodes; can morph (text → super-tag, search node, etc.).
  • Dangerous like Notion/Obsidian (rabbit-hole building) yet powerful.

• Comparison vs. Airtable: easier quick-entry, proper note-taking.

• Example workflow:
  • Create super-tag "Event" on the fly; automatically spawns record & table.
  • Search node lists all frameworks records dynamically.

Tana AI Actions & YouTube Workflow

• Custom super-tag "YouTube" with template fields (link, transcript, key-points).

• Two custom buttons:

  1. Get Transcript (pulls captions, stores under node).

  2. Get Key Points (GPT prompt summarises transcript).

• Prompts written in Tana AI configuration (multi-bullet supported).

• Can chain HTTP actions (call any API).

Exporting / Importing Structured Markdown via AI

• Showed custom GPTs for Obsidian & Logseq:
  • Provide tool-specific MD examples inside system prompt.
  • ChatGPT returns ready-to-paste blocks with zero format loss.
  • Same approach planned for Tana.

Question 2 – Yael: PDF → JSON → Airtable via Make.com

• Scenario: Google Drive watch → OpenAI file-upload → Vector store search → Assistant → JSON → Airtable.

• Error 400: "All tools must be type function when response format is json_object".

• Fixes:
  • Remove function/tool mismatch; set assistant response_format to "auto".
  • Add second "Create Completion" step with Parse JSON = Yes, Temperature $\approx 0$ to minimise hallucinations.
  • Clean stray "json:" text via regex or extra GPT step.

• Educational points:
  • Temperature param controls creativity (0=factual).
  • Assistants module lacks auto-parse – hence extra completion step.
  • Debug by inspecting Make execution history.

Mind-Mapping Tools & Tips

• Logseq mind-map plugin: number-keys (1–4) cycle levels; non-editable but great zoom/declutter.

• Obsidian “Enhancing Mind Map” plugin:
  • Settings → Layout (Right, Center, etc.) to avoid split branches.
  • Editable; changes reflect in underlying note.

• Multiple Obsidian plugins exist – trial to match personal aesthetics.

• Canvas/whiteboard options (Obsidian, Logseq, Excalidraw) for flowcharts & diagrams.

• Mermaid, PlantUML etc. can be generated by GPT for visuals.

Canvas / Whiteboard Criteria

• Infinite canvas (nest canvas inside canvas).

• Each visual element = real note (bidirectional links/backlinks).

• Import/export via code/text for automation.

• Examples: Heptabase, Scrintle, Milanote.

Question 3 – Capacities Pros & Cons

• Nickname: “child of Notion + Obsidian”.

• Visual page builder (slash menu, toggles) like Notion.

• Graph view & backlinks like Obsidian.

• Unique strength: Object-based note-taking (definitions, quotes, tweets, stories, etc.).
  • Any highlight can be Turn into → {Object}; object autolives in its list + original context.

• Non-hierarchical; avoids folder rigidity.

• Lacks true outliner indentation ⇒ host prefers Tana.

Question 4 – Flowise Reliability for Commercial Chatbots

• Flowise built on LangChain; majority of SaaS chatbots use LangChain.

• Errors may be due to external modules or compute limits.

• Troubleshoot: share Loom of node-graph, log errors; tag community experts (Juan, Rui).

• Alternatives: LangFlow, other drag-and-drop LangChain GUIs.

Question 5 – Fine-Tuning vs. Assistants (Tejwant)

• Fine-tuned models can be selected in Assistant dashboard (dropdown shows ft:org/…).

• Workflow:

  1. Create fine-tune job (no Make needed).

  2. Build Assistant using that model; upload large knowledge (10 k files supported).

  3. Call via standard Assistants API or through Make’s "Message Assistant".

• Distinction:
  • Custom GPT = consumer UI only, limited file KB.
  • Assistant = API-first, huge file store, function/tool calling.

Question 6 – Voice Capture to PKM (Tejwant)

• Tana Capture mobile app:
  • Quick voice memo, photo, or text; syncs to desktop.
  • Defaults to Inbox; later move/retag (supports task/person detection).

• Desktop hotkey $\text{Cmd+Shift+E}$ records microphone instantly.

• Future: refine regex/tag parsing to auto-route.

System-Mapping Query (Olya)

• Desired complex stakeholder & causal loop maps (Kumu style).

• Suggested avenues:
  • Use GPT to output Mermaid / PlantUML / SysML code; import into tool supporting that dialect (e.g. Excalidraw, Obsidian Canvas with Mermaid render).
  • Key is finding mapping platform that accepts text import.
  • Specialised graph tools may still require manual refinement.

Web Scraping & No-API Workflows (Anya / Olya)

• If insurer sites lack API: use headless bots.
  • Advanced scrapers (BrowseAI, Apify) simulate login, click, submit, extract.
  • Workflow: variable 🡒 robot 🡒 Make webhook 🡒 downstream app.

• BrowseAI perks: schedule cadence, stores cookies for authenticated areas, unofficial API layer.

• Legal note: check TOS; cat-and-mouse vs. anti-bot systems.

Miscellaneous Tips & Resources

• Jim Kwik 4×4 quadrant note-taking: recursively divide topic into four; replicate digitally with nested nodes.

• Hotkeys summary:
  • Tana voice: $\text{Cmd+Shift+E}$.
  • Logseq mind-map view cycle: keys 1–4.
  • Obsidian mind-map layout toggles via Settings → Enhancing Mind Map.

• Recommended video: “100 Tana Tips” (46 min, x2 speed ok).

• Excalidraw open-source whiteboard; Obsidian plugin embeds drawings; AI can generate diagram via Mermaid code.

File & Prompt Libraries Shared

• Airtable templates (linked on AI Mastery portal) for:
  • Image generation (MidJourney variables: chaos, fast, raw, stylize, weird).
  • Text→Video & Image→Video prompts.

• Prompts leverage variable substitution; accessible under respective session guides.

Closing & Logistics

• Session ran ~1 h 57 min; host ends quickly to put son to bed (22:00 local).

• Reminder for attendees to explore tools, post Looms/questions in community; peer experts monitor.

Tana is described as a unique category of tool, operating as a hybrid that masterfully combines the structured data management capabilities of Airtable with the flexible, hierarchical organization of outliners.

Strengths

• Dynamic Table Creation: Users can create tables (referred to as "super-tags") from any location within the note-taking environment, eliminating the need to navigate to separate databases or bases, which significantly streamlines workflow.

• Node Flexibility: All elements within Tana are inherently nodes, granting them the ability to seamlessly transform from one type to another (e.g., plain text can become a super-tag, a search node, or various other entity types) without loss of context.

• Powerful yet Potentially Complex: Similar to Notion or Obsidian, Tana offers immense power for building interconnected knowledge systems, but this can also lead to a "rabbit-hole" where users spend considerable time building rather than just noting. Despite this, its capabilities are profound.

Comparison vs. Airtable

• Tana offers a significantly easier and more intuitive experience for quick data entry and authentic note-taking, making it superior for fluid information capture compared to Airtable's more rigid structure.

Example Workflow

• Users can dynamically create a super-tag, for instance, an "Event" super-tag, directly on the fly. This action automatically generates a corresponding record and table, instantly structuring the new information.

• A search node can be established to dynamically list all records associated with specific frameworks, providing an always up-to-date overview of related information across the system.

Tana AI Actions & YouTube Workflow

• A custom super-tag named "YouTube" can be created, which includes pre-defined template fields such as "link," "transcript," and "key-points," streamlining the organization of YouTube-related content.

• Two custom buttons are integrated into this super-tag:

  1. Get Transcript: This button is configured to automatically pull captions from a YouTube video and store the full transcript directly under the associated node within Tana.

  2. Get Key Points: This button leverages a GPT prompt to generate a concise summary of the retrieved transcript, extracting the most crucial information into a digestible format.

• The prompts for these AI actions are configured directly within Tana's AI configuration settings, supporting multi-bullet point outputs for clarity.

• Tana allows for chaining HTTP actions, enabling users to call almost any external API to extend its functionality.

Exporting / Importing Structured Markdown via AI

• Custom GPTs designed for Obsidian and Logseq were demonstrated, showcasing their ability to handle structured markdown.

• The approach involves providing specific tool-dependent Markdown examples within the system prompt of the GPT. This trains the AI to understand and replicate the desired formatting.

• As a result, ChatGPT can return blocks of text that are ready-to-paste into Obsidian or Logseq, ensuring zero loss of formatting or structural integrity.

• The same methodology is planned for implementation with Tana, aiming to achieve similar seamless import/export capabilities.

Question 2 – Yael: PDF → JSON → Airtable via Make.com

Scenario

• The workflow begins with a Google Drive watcher, which triggers an OpenAI file-upload. This is followed by a vector store search, interaction with an Assistant, generation of a JSON output, and finally, integration into Airtable.

Error Encountered

• An "Error 400" occurs with the message: "All tools must be type function when response format is json_object," indicating a mismatch in how tools are defined or used when expecting a JSON output.

Fixes & Solutions

• Remove Function/Tool Mismatch: Ensure that there are no conflicting definitions or usages of functions/tools that contradict the json_object response format. The assistant's response_format should be set to "auto" to allow the system to infer the appropriate format.

• Add Second "Create Completion" Step: Implement an additional "Create Completion" step. In this step, explicitly set Parse JSON = Yes to ensure the incoming text is correctly interpreted as JSON. Crucially, set the Temperature parameter to approximately $0$ (e.g., $Temperature \approx 0$) to minimize the model's creativity and ensure a factual, consistent JSON output, reducing "hallucinations" or irrelevant text.

• Clean Stray "json:" Text: Sometimes, the model might output extraneous text like "json:" before the actual JSON object. This can be cleaned using regular expressions or by adding an extra GPT step specifically for text purification.

Educational Points

• Temperature Parameter: The Temperature parameter in AI models controls the level of randomness or creativity in the output. A value of $0$ makes the output deterministic and factual, ideal for structured data like JSON.

• Assistants Module Limitation: The Assistants module in OpenAI currently lacks an inherent auto-parsing feature for JSON. This necessitates the additional completion step to ensure proper JSON interpretation.

• Debugging Strategy: For effective troubleshooting, thoroughly inspect the execution history within Make.com (or similar integration platforms). This provides detailed logs of each step, helping pinpoint where errors originate.

Mind-Mapping Tools & Tips

• Logseq Mind-Map Plugin: This plugin utilizes number keys ($1-4$) to cycle through different levels of the mind map, allowing for dynamic zooming and decluttering. While not directly editable in this view, it's excellent for visual navigation.

• Obsidian “Enhancing Mind Map” Plugin:

  • Layout Settings: Within the plugin settings, users can adjust the layout (e.g., Right, Center) to prevent branches from splitting awkwardly, ensuring a cleaner visual presentation.

  • Editability: This plugin allows for direct editing within the mind map view, and any changes made are immediately reflected and saved in the underlying note document.

• Multiple Obsidian Plugins: Given the thriving Obsidian plugin ecosystem, numerous mind-mapping plugins exist. Users are encouraged to trial various options to find one that best matches their personal aesthetics and workflow preferences.

• Canvas/Whiteboard Options: Tools like Obsidian Canvas, Logseq's whiteboard, and Excalidraw are highly recommended for creating flowcharts and diagrams, offering a flexible visual workspace.

• AI-Generated Visuals: Diagramming languages like Mermaid and PlantUML can be generated by GPT and then rendered in compatible tools, enabling automated creation of complex visuals.

Canvas / Whiteboard Criteria

• Infinite Canvas: The ideal tool should offer an infinite canvas, with the ability to nest canvases within other canvases, providing limitless space for organizing information hierarchically.

• Real Note Integration: Each visual element on the canvas should correspond to a real note, supporting bidirectional links and backlinks. This ensures that visual representations are deeply integrated with the knowledge base.

• Code/Text Import/Export: The ability to import and export canvas data via code or text is crucial for automation and programmatic manipulation of diagrams.

• Examples: Leading examples of tools that embody these criteria include Heptabase, Scrintle, and Milanote.

Question 3 – Capacities Pros & Cons

• Nickname: Capacities is often referred to as the “child of Notion + Obsidian,” suggesting it combines the strengths of both popular note-taking platforms.

• Features Borrowed from Notion: It offers a visual page builder with a user-friendly slash menu and toggle features, similar to Notion's intuitive content creation and organization.

• Features Borrowed from Obsidian: Capacities includes a graph view and robust backlink functionality, allowing users to visualize connections between notes and easily navigate related content, much like Obsidian.

• Unique Strength: Object-Based Note-Taking: This is a core differentiator. Capacities allows users to define and manage various types of "objects" such as definitions, quotes, tweets, stories, and more.

  • Any highlighted text can be transformed into an object using the "Turn into → {Object}" command. Once created, the object automatically lives in its designated list (e.g., a list of all quotes) while also maintaining its original context within the note, providing powerful contextual organization.

• Non-Hierarchical Structure: Capacities avoids rigid folder structures, promoting a more fluid and less constrained organization of information.

• Limitation: It lacks true outliner indentation capabilities, which is a key reason why the host prefers Tana for its superior outlining features.

Question 4 – Flowise Reliability for Commercial Chatbots

• Foundation: Flowise is built on LangChain, a highly popular framework for developing language model applications. The vast majority of SaaS chatbots currently in commercial use are also underpinned by LangChain.

• Error Causes: Reliability issues in Flowise are typically not inherent to the platform itself but might stem from external modules being used or from reaching compute limits of the underlying infrastructure.

• Troubleshooting: To diagnose problems effectively, users are advised to share a Loom video demonstrating their node-graph, log any errors encountered, and tag community experts (such as Juan or Rui) for specialized assistance.

• Alternatives: Other drag-and-drop LangChain GUIs, such as LangFlow, serve as viable alternatives if Flowise proves unsuitable for specific commercial chatbot requirements.

Question 5 – Fine-Tuning vs. Assistants (Tejwant)

• Fine-Tuning Integration: Fine-tuned models can be seamlessly selected and utilized within the Assistant dashboard, typically appearing as an option in a dropdown menu with a format like ft:org/....

• Workflow for Using Fine-Tuned Models with Assistants:

  1. Create Fine-Tune Job: Initiate a fine-tuning job for a custom model. This step does not require an external platform like Make.com.

  2. Build Assistant with Model: Construct an Assistant, specifying the newly fine-tuned model for its operations. This Assistant can be uploaded with a substantial knowledge base, supporting up to $10,000$ files, making it suitable for large-scale data.

  3. Call Assistant: Interact with the Assistant either via the standard Assistants API or through specific integration modules like Make's "Message Assistant."

• Distinction Between Custom GPT and Assistant:

  • Custom GPT: Primarily a consumer-facing UI with limited file knowledge base capacity.

  • Assistant: An API-first service designed for programmatic interaction, offering a massive file store for document retrieval and robust capabilities for function/tool calling, making it more suitable for complex, integrated applications.

Question 6 – Voice Capture to PKM (Tejwant)

• Tana Capture Mobile App: This application allows users to quickly record voice memos, capture photos, or input text notes on the go. These captures automatically sync to the desktop version of Tana.

  • Captured items default to an "Inbox" for initial processing. Users can later move or retag these items, with support for automatic detection of tasks or mentions of people.

• Desktop Hotkey: On desktop, the hotkey $ext{Cmd+Shift+E}$ (Command + Shift + E) provides instant microphone recording, allowing for rapid voice note capture directly within Tana.

• Future Enhancements: There are plans to refine regex/tag parsing capabilities to enable automatic routing of captured voice notes or text snippets to specific locations or categories within the PKM system.

System-Mapping Query (Olya)

• Desired Capabilities: The user (Olya) is seeking tools capable of generating complex stakeholder maps and causal loop maps, similar to those created with Kumu.

• Suggested Avenues:

  • AI Code Generation: Utilize GPT to generate Mermaid, PlantUML, or SysML code. This code can then be imported into a mapping tool that supports these diagramming dialects (e.g., Excalidraw, or Obsidian Canvas with a Mermaid renderer).

  • Key Requirement: The crucial factor in tool selection is finding a mapping platform that inherently accepts text-based imports for diagram generation.

  • Manual Refinement: While AI can provide a strong starting point, highly specialized graph tools may still necessitate some degree of manual refinement to achieve the desired level of detail and accuracy.

Web Scraping & No-API Workflows (Anya / Olya)

• No-API Scenarios: For websites that do not offer a public API, such as many insurer sites, the recommended approach is to use headless bots for web scraping.

• Advanced Scrapers: Tools like BrowseAI and Apify are designed to simulate human interactions, including logging in, clicking buttons, submitting forms, and extracting data.

• Workflow: The general workflow involves defining a variable, deploying a robot (scraper), which then sends extracted data to a Make.com webhook, and from there, the data flows into a downstream application.

• BrowseAI Perks: BrowseAI offers several advantages, including the ability to schedule scraping cadences, store cookies for accessing authenticated areas (e.g., after login), and provides an unofficial API layer for integration.

• Legal Note: It's crucial to review the website's Terms of Service (TOS) before engaging in web scraping, as it can escalate into a "cat-and-mouse" game with anti-bot systems if not permitted.

Miscellaneous Tips & Resources

• Jim Kwik's 4x4 Quadrant Note-Taking: This method involves recursively dividing a topic into four main quadrants. In a digital environment, this can be replicated effectively using nested nodes within a note-taking application to create a structured hierarchy.

• Hotkeys Summary:

  • Tana Voice Capture: Utilize $ext{Cmd+Shift+E}$ (Command + Shift + E) for instant voice recording.

  • Logseq Mind-Map View Cycle: Use keys $1-4$ to cycle through different levels of the mind map view.

  • Obsidian Mind-Map Layout Toggles: Layout adjustments for the mind map can be accessed and configured via the Settings menu, specifically within the "Enhancing Mind Map" plugin options.

• Recommended Video: A video titled “100 Tana Tips” (approximately $46$ minutes long) is highly recommended, and it can be watched at $2 ext{x}$ speed for efficiency.

• Excalidraw: This is an open-source whiteboard tool. The Obsidian plugin allows for embedding Excalidraw drawings directly into notes, and AI can generate diagrams for Excalidraw via Mermaid code.

File & Prompt Libraries Shared

• Airtable Templates: Access to various Airtable templates is provided through the AI Mastery portal, designed for specific creative tasks:

  • Image Generation: Templates include variables for MidJourney commands such as chaos, fast, raw, stylize, and weird, allowing for nuanced control over image outputs.

  • Text-to-Video & Image-to-Video Prompts: Specialized prompts are available for generating videos from text or images.

• Prompt Leverage: These prompts are structured to leverage variable substitution, making them highly adaptable. They are easily accessible under their respective session guides, ensuring users can quickly find and utilize them.

Closing & Logistics

• Session Duration: The session concluded after approximately $1$ hour and $57$ minutes, with the host ending promptly to attend to a personal commitment (putting son to bed at $22:00$ local time).

• Call to Action: Attendees are reminded and encouraged to actively explore the discussed tools, post Loom videos showcasing their workflows or issues, and submit questions within the community forum, where peer experts are available to monitor and provide assistance.