VIZION AI
Back to Systems
Knowledge & Media System

Knowledge Pipeline

Companies rarely have a shortage of recorded material. They have a retrieval problem. The Knowledge Pipeline turns unorganized video, audio, and meeting archives into structured, source-grounded reference libraries that teams can actually query and verify.

System Architecture & Flow

Vizion Knowledge Pipeline System Flow — Deterministic Media Ingestion & Private Workspace Grounding
Click to expand full diagram

Starting Baseline Setup & Maintenance

Starting at 10,000 CZK setup · 10,000 CZK / mo

(approx. €420 setup · €420 / mo baseline)

The 10,000 CZK baseline setup covers reviewing your existing audio and video archives, cleaning and structuring transcripts, grouping them by topic, and configuring your first private Google NotebookLM workspace with grounded search and citation badges.

Ongoing monthly maintenance (from 10,000 CZK / mo) covers reviewing new recordings your team produces, cleaning transcripts of technical terms, and adding them to your notebooks so search stays up to date.

For teams that want to turn archives into training materials, expanded tiers cover producing custom NotebookLM Studio assets: two-host Audio Overview podcasts, employee onboarding study guides, and briefing sheets.

Your data never leaves your control: everything is loaded into your own Google accounts. We clarify the exact scope during discovery before any commitment.

See Riffle, our public mobile showcase built from over 500 technical videos, organized into searchable reference cards with direct source links.

The Real Problem

Most company video and audio sits in folders completely untouched. Team meetings, client calls, webinars, and training sessions accumulate constantly, but nobody has the time to watch an hour of video to find a five-minute practical answer.

The Grounding Standard

Generic AI tools summarize meetings by guessing what mattered. We anchor every note to the original recording and transcript. You see the exact discussion in its original context, instead of trusting an unverified chatbot summary.

Your Data Stays Yours

We do not host your files or store your recordings on proprietary servers. Everything goes directly into your own Google Workspace and NotebookLM accounts. We only see what you share with us during the project, and you keep all the files.

Challenging Our Own Claims

What we hoped for, what we tested, and what actually happened.

It is easy to make an AI demo look impressive in a five-minute video. It is much harder to make it work reliably across 500 hours of real recordings. Here are four times our own assumptions broke down, and what we changed to fix them.

The test said "Passed." The answers were completely wrong.

Why our automated test reported a 90% success rate while returning the exact same answer across eight questions.

What we assumedWe wrote automated test scripts to check whether our tool was pulling answers from the screen correctly. When the script reported a 90% success rate with green checkmarks, we assumed the system was reading and saving answers accurately.
What actually happenedWhen we opened the saved files to check the actual words, eight out of the ten questions had the exact same paragraph copied into them. The script had counted the text boxes on the webpage and grabbed the last one. But in Google NotebookLM, the last text box on the screen is a static help container that never changes. The automated test gave us a green checkmark, but the system was quietly copying the wrong box over and over.
What we changedWe stopped trusting simple "success" checkmarks. We rewrote the tool to read backwards through the screen, ignoring anything it had already seen until it found brand new text. More importantly, every test run now saves its own text file and checks that the answer actually mentions the specific topic we asked about.

The uncomfortable question: Do you even need our tool, or can ChatGPT do this alone?

Testing whether a leading AI model with no access to our recordings could answer just as well.

What we assumedWhen you spend months building a system to structure company recordings, you have to ask the most uncomfortable question: what if a client doesn't need this at all? What if they can just open a normal AI chat window and get the same result?
What actually happenedWe ran a test where we gave a leading AI model (with no access to our video transcripts) the exact same technical questions. Without our clean source material, the model either made up convincing-sounding steps that did not work, or admitted it didn't know. But when we connected the model to our cleaned, verified transcript archive, it gave exact, accurate steps with links back to the original video.
What we changedWe realized the AI itself isn't the valuable asset—the clean, organized video archive is. As AI models get smarter, they don't replace your company's recordings; they make having clean, verified source material even more important.

The browser reload that secretly corrupted new questions.

Why refreshing the page failed to clear workspace memory, and how old chats leaked into new questions.

What we assumedWe assumed that refreshing the browser between questions would wipe the slate clean so the AI could start fresh with no leftover baggage.
What actually happenedIt didn't start fresh. Even after a full page refresh, Google quietly reloaded the previous conversation in the background. The tool thought it was asking an isolated question, but the AI was still reading the old conversation, letting earlier topics bleed into brand new questions without warning.
What we changedWe had to program the tool to perform the full three-step deletion: opening the chat menu, clicking delete history, and confirming the prompt. When working with AI interfaces, you can never assume a session resets itself. Without deliberate cleanup, answers from one project will quietly bleed into the next.

What happened when we let an AI organize video tags on its own.

Why smaller AI models should only propose draft ideas, and why automation should never approve its own work.

What we assumedWe thought we could save time by letting a smaller, faster AI model automatically categorize and tag hundreds of video transcripts into neat topics.
What actually happenedIt created a mess. Across hundreds of videos, the AI started inventing slight variations of existing tags, creating duplicate categories, and drifting away from our topic structure. It gave the appearance of being organized, but searching through it became frustrating.
What we changedWe took away the AI's permission to edit our records directly. Now, the model only gets to pick from a frozen list of approved tags and writes its choices as a draft. A separate check confirms the tags are valid, and a person on our team reviews the batch before anything is saved. An AI can propose ideas quickly, but it should never be allowed to approve its own work.

Why This Exists

Most business teams accumulate hours of video and audio every month: onboarding sessions, customer interviews, weekly strategy meetings, and recorded software walkthroughs. Once recorded, those files sit in cloud folders because nobody has the hours required to listen through them again.

Typical AI transcription tools generate a wall of text, but reading through a 50-page transcript is almost as slow as watching the video. Generic chatbot summarizers gloss over technical nuances and smooth out the specific steps you actually needed.

The Knowledge Pipeline is built to make that accumulated knowledge reachable. By organizing recordings into clear thematic collections and loading them into Google NotebookLM, team members can ask natural questions and get grounded answers with clickable citation badges pointing to the exact transcript passage.

For teams that want more than basic search, it opens up NotebookLM's studio tools: you can turn a five-hour training playlist into a 10-minute conversational audio overview (podcast), an onboarding study guide, or an executive briefing document for your team.

Best of all, you own the infrastructure. We do not store your recordings on our servers or charge you a per-seat software subscription. Everything sits inside your own Google account, with access strictly controlled by you.

The Reality

This is not about generating generic video recaps or automated social clips.

It is a structured way to turn your existing video and audio assets into a verified, searchable internal reference tool that saves your team time.

What it actually does

  • Pulling transcripts from your video, audio, or playlist archives and stripping conversational filler
  • Labeling each recording with session dates, titles, and direct links back to the original source files
  • Grouping transcripts by topic so related team discussions stay together instead of getting lost in folders
  • Configuring your private Google NotebookLM workspace so team queries return answers with clickable citation badges
  • For larger setups, generating NotebookLM Studio assets: two-host Audio Overview podcasts, onboarding study guides, and briefing sheets
  • Zero vendor lock-in: everything runs inside your own Google accounts without recurring third-party software subscriptions

What a client project can include

  • Reviewing your existing recording archives and assessing audio clarity for transcription
  • Extracting and cleaning transcripts to strip verbal filler and correct technical terms
  • Organizing recordings into clear topic collections suited to your team workflows
  • Setting up your Google NotebookLM workspace and testing queries against real team questions
  • Optional studio tier: producing custom Audio Overviews (two-host podcast discussions) and employee onboarding study guides
  • Full handoff and documentation so your team can add future recordings independently

Live Demonstration

See the pipeline in action.

We tested this pipeline against a public corpus of over 500 technical videos, compiling the results into Riffle: a mobile card feed with over 580 curated reference cards.

Working Proof

  • 588+ real cards across 6 months of technical releases
  • Direct links back to original videos and source transcripts
  • Mobile-first snap feed designed for intent capture and prompt creation
View the Riffle overview