AI FORWARD

Knowledge Systems

Transcribe Video to a Living Wiki

Turn videos from feeds, chats, demos, interviews, and training libraries into source-backed Markdown that can be reviewed, searched, and folded into a maintained wiki.

Created June 5, 2026 · Last updated June 5, 2026 · Research checked June 5, 2026 · By Forward IC

TLDR

Video-to-wiki breaks into two distinct problems: transcription (largely solved) and wiki maintenance (mostly unsolved). The underused shortcut: YouTube auto-captions via yt-dlp --write-auto-subs cover most public video at zero cost and zero privacy exposure — no audio download, no API key, no GPU required. For private or offline content, local Whisper plus Ollama via synto is the only fully private path. The hard problem is entity reconciliation: 50 videos that mention the same researcher should converge to one continuously updated wiki page, not 50 disconnected notes. Of current tools, synto alone handles all four properties of a living wiki: incremental updates, entity reconciliation, contradiction flagging, and backlink maintenance. steipete/summarize is the community-preferred single-command path for URL-to-Markdown summaries (88 Hacker News upvotes, May 2026). Entity extraction — converting raw transcripts into named entities with relationships — is the step that unlocks the jump from a folder of notes to a searchable, interlinked knowledge base.

The workflow below turns a personal habit into a general capture-to-knowledge system. A person sends a video link into a private channel. A worker resolves the source, extracts audio, transcribes it, writes source-backed Markdown, and hands that Markdown to an LLM-maintained wiki.

Data Flow
Human saves or sends video link The user stays inside the capture channel they already use. Message trigger A prefix, tag, channel, or command creates an ingestion job. URL extraction and platform detection The worker identifies the source and chooses an adapter. Video sourcing API Official API, unified API, downloader, or managed scraper. Media URL and metadata Title, source URL, platform, publish time, and media reference. Audio extraction Normalize source media into a transcription-ready audio file. Speech-to-text model Hosted or local transcription, with provider and model recorded. Transcript Markdown with frontmatter Immutable evidence, not final synthesis. Raw source folder Append-only storage for review, search, and re-ingestion. LLM wiki ingest Agent reads the raw source and applies wiki-maintenance rules. Topic, entity, and synthesis pages Durable knowledge is updated separately from raw evidence. Obsidian graph, search, and review Humans browse, audit, correct, and ask new questions. Short summary A fast status result can be generated from transcription. Notification back to capture channel The loop closes where the source was captured.
The critical split happens after transcription: one path creates durable raw Markdown, while the other gives the user a fast acknowledgement.

The Core Idea

The inspiration is close to Andrej Karpathy's LLM Wiki pattern: raw sources remain immutable, generated wiki pages become the structured middle layer, and an agent follows explicit instructions for ingesting, linking, revising, and linting the knowledge base.

Video-to-Markdown is a missing ingestion lane for that pattern. Many knowledge bases already handle web pages, PDFs, notes, and books. Video is harder because the source is transient, platform-specific, and often locked behind feed products instead of clean APIs.

Reader Map
What readers need to understand first Capture is a reliability problem. A saved link is not knowledge until the source, transcript, metadata, and review state are preserved. What the architecture optimizes for Keep raw markdown immutable, update synthesized wiki pages separately, and review generated claims before they become durable knowledge.
This before/after frame helps readers see why the pipeline exists before they read implementation details.

What Makes a Wiki "Living": The Four Properties

The difference between a folder of timestamped notes and a living wiki is maintenance intelligence. A living wiki must handle four properties continuously — not just at initial ingest:

Four Properties That Distinguish a Living Wiki from a Notes Folder
Incremental updates New content enriches existing pages instead of creating duplicate notes. Adding a second video about the same researcher updates their page, not appends a new file. Entity reconciliation Names, organizations, and concepts across multiple sources resolve to canonical pages. "OpenAI", "Open AI", and "OAI" become one entity. Without this, the graph fragments. Contradiction flagging When new content contradicts existing facts (a valuation changed, a paper was retracted), the wiki surfaces the conflict rather than silently overwriting the prior version. Backlink maintenance When a page is updated or renamed, all pages that reference it update automatically. The knowledge graph stays consistent without manual link repair after each ingest.
Of current open-source tools, only synto implements all four properties for local content. Entity extraction — converting raw transcripts into named entities with typed relationships — is the step that unlocks the jump from a folder of notes to a searchable, interlinked knowledge base.

Business Use Cases

  • A venture team captures founder interviews, demo videos, and market commentary into a firm wiki.
  • A product team turns customer-call clips into maintained product-insight pages.
  • A sales team captures public competitor demos and updates battlecards with source-backed observations.
  • A research team captures conference talks and maintains topic summaries and disagreement logs.
  • A support team turns training videos into searchable internal runbooks.

Separate Capture From Synthesis

Sequence
Send tagged video link Receive message
Expose new tagged item Poll, extract URL, detect platform
Resolve metadata, extract audio, transcribe
Receive status and summary Reply to capture channel Store raw transcript Markdown
LLM wiki agent reads new raw source Update pages, indexes, links, and review surfaces
The transcript file is an auditable input with stable metadata. The wiki can summarize, merge, tag, and link it later without destroying the source trail.

Architecture Layers

Architecture
Capture iMessage, Slack, Discord, email, webhook, browser clipper. Ingestion URL parser, platform resolver, media retrieval, audio extraction, transcription. Storage Immutable transcripts, metadata frontmatter, append-only ingest log. Wiki LLM-maintained pages, backlink hygiene, stale-claim checks, graph/search review.
For a company version, the capture and storage products change; the interface between raw evidence and generated synthesis should not.

Video Sourcing Options

Video sourcing is the least stable layer because APIs, anti-bot systems, rate limits, legal terms, and media URL formats change frequently. Pricing below was checked on June 5, 2026 and should be rechecked before procurement.

Option Pricing signal Strengths Tradeoffs Best fit
Official platform APIs Usually free or quota-based for approved use; engineering and review cost dominate. Defensible compliance posture and stable contracts where supported. Narrow scopes, app review, limited media access, and missing platforms. Enterprise workflows where policy compliance matters more than coverage.
TikHub-style unified APIs TikHub lists pay-as-you-go at $0.001-$0.01/request, with daily volume discounts. One integration for TikTok, Douyin, Xiaohongshu/RedNote, and other social platforms. Third-party dependency and vendor-risk review. Pragmatic ingestion for mixed social-video sources.
yt-dlp Open-source tool cost is $0, but you own compute, retries, monitoring, and compliance review. Broad community-maintained extractor coverage and easy local testing. Extractors can break; some sites need cookies or extra runtime support. Personal use, internal tooling, or permitted public-source fallback.
Apify Actors Apify platform plans start free; paid plans list $0.13-$0.20/compute unit plus actor-specific charges. Managed scraping jobs, datasets, API access, and platform-specific actors. Actor quality, pricing, and scraping risk vary. Batch research and market monitoring.
Browser/scraping APIs Usually priced by request, page, bandwidth, browser time, or plan; verify per vendor. Useful for web-page context, screenshots, Markdown extraction, and structured page data. Usually not a primary answer for video media extraction. Companion context around the video source.

Community Transcription Toolkit

The most underused path: YouTube auto-captions via yt-dlp --write-auto-subs --skip-download require no audio download, no API key, and no GPU — and cover most public YouTube content. Caption-first is free and instant. Only fall back to audio transcription when captions are absent, inaccurate, or the content is private.

Tool Input Privacy Community signal
yt-dlp --write-auto-subs YouTube URL No audio sent anywhere De facto standard; widely adopted in automation pipelines
steipete/summarize URL, file, or text Sends to chosen API 88 HN upvotes (May 2026); preferred single-command URL-to-Markdown path
macparakeet Audio file 100% local (Apple Silicon CPU) Emerging favorite for M-series Macs; no GPU required
OpenBrief Video URL API-dependent One-click summary with source links; non-developer audience
Whisper (local) Audio / video file 100% local Community-standard for offline transcription; integrates with synto
KrillinAI Video URL API-dependent Multilingual transcription and translation; subtitle generation pipeline

Speech-to-Text Model Choices

Transcription is more stable than sourcing, but model choice still depends on language mix, cost, latency, diarization, privacy, and whether the output needs raw transcript only or semantic cleanup. Pricing below was checked on June 5, 2026.

Provider/model family Pricing signal Strengths Best fit
OpenAI transcription models OpenAI lists gpt-4o-transcribe at $0.006/minute, plus token pricing for text. Hosted transcription with modern multilingual quality and prompt support for context. High-quality default for multilingual knowledge ingestion.
Groq Whisper Large v3 / Turbo Groq lists Whisper V3 Large at $0.111/hour and Whisper Large v3 Turbo at $0.04/hour. Very fast Whisper-family inference for low-latency workflows. Fast personal ingestion and near-realtime processing.
Deepgram Nova-3 Deepgram lists Nova-3 pay-as-you-go around $0.0048-$0.0092/minute depending on model and mode. Enterprise speech stack with real-time and batch options. Call-center, meeting, and conversational workloads.
AssemblyAI Universal models AssemblyAI advertises pay-as-you-go starting as low as $0.15/hour. Transcription plus speech-intelligence product features. Teams needing chapters, entities, summaries, or speech analytics.
Local Whisper variants No per-minute API fee; cost moves to GPUs, operations, queues, and throughput constraints. Privacy, offline operation, and predictable hardware-based costs. Sensitive sources, offline workflows, and cost-controlled backfills.

The Knowledge-Base Layer

Markdown Knowledge Loop
Immutable raw transcript Metadata frontmatter LLM-maintained wiki page Human graph/search review
The final system is not just "save transcript files." Markdown remains the stable source format, Obsidian or a graph viewer becomes the human workspace, and local search/MCP tools become agent affordances.

Design Principles

  1. Preserve raw sources. A summary should never replace the transcript.
  2. Separate ingestion from synthesis. A transcript is evidence; the wiki is interpretation.
  3. Make capture habitual. The best input channel is the one the user already uses.
  4. Keep providers swappable. Video sourcing and transcription vendors will change.
  5. Record metadata. Platform, URL, title, capture time, publish time, model, and provider should survive.
  6. Use Markdown as the stable substrate for Obsidian, Git, static sites, search tools, and agents.
  7. Build review into business workflows. Shared organizational knowledge usually needs approval.
  8. Lint the wiki periodically for stale claims, orphan pages, duplicates, and contradictions.

Production Architecture

AWS Reference Architecture
API Gateway / EventBridge Capture webhooks from Slack, Teams, browser extensions, or internal tools. SQS ingestion queue Buffer retries, isolate provider failures, and protect downstream workers. Lambda or ECS workers Resolve URLs, call source adapters, extract audio with FFmpeg, and emit status events. S3 raw evidence lake Store media references, transcripts, metadata JSON, and immutable Markdown. Amazon Transcribe or external STT Use AWS-native speech when compliance and account consolidation matter. DynamoDB / Aurora Track jobs, source provenance, provider costs, retries, and review state. Bedrock / containerized wiki agent Generate proposed wiki diffs, summaries, links, and stale-claim checks. CloudFront static wiki Publish approved Markdown-derived pages, search indexes, and audit views.
Google Cloud Reference Architecture
Cloud Run / API Gateway Receive capture requests from chat tools, extensions, CRM notes, and webhooks. Pub/Sub job topics Decouple capture, media sourcing, transcription, and wiki synthesis stages. Cloud Run Jobs Run URL resolvers, source adapters, FFmpeg extraction, and provider connectors. Cloud Storage evidence bucket Keep immutable transcripts, frontmatter, normalized audio, and ingest artifacts. Speech-to-Text v2 or external STT Use Google-native speech when regional controls and Cloud billing matter. Firestore / Cloud SQL Store job state, source registry, cost records, access policy, and review status. Vertex AI / Cloud Run wiki agent Create source-backed page updates and route risky changes to review. Cloud CDN static wiki Serve approved pages, indexes, graph data, and reviewer-facing static assets.

Risks and Open Questions

Review Gates
Capture Policy What can be downloaded, stored, transformed, and redistributed? Provider Drift APIs, extractors, rate limits, and media URL formats change. Transcript Quality Language mix, audio quality, timestamps, and diarization affect trust. Wiki Approval Personal systems can auto-merge; company systems usually need staged diffs.

References