Skip to content
CoS Knowledge
Esc
↑↓navigate↵open⌘Jpreview
On this page

Media pipeline ingest

Description

Bring media into the knowledge base through the Cloudflare media pipeline: upload originals, get derived text (transcripts, converted Markdown), then extract notes and tasks into cos.hodgederrick.com sections.

Purpose

One ingest path for everything non-text: meeting recordings, voice memos, PDFs, images. Originals stay immutable in R2; the KB keeps derived, typed, frontmatter-compliant pages with links back to the media.

Pipeline surface (verified)

Surface Value
API/UI https://ingest.hodgeluke.com (UI: “Library — Hodge Luke”)
Public files https://media.hodgeluke.com (+ cdn-cgi/image/… zone transforms)
Agent docs /llms.txt, /openapi.json
Upload token pass show cloudflare/media-pipeline/upload-token (Bearer, uploads only; reads public)
Refine (S1-mini) GET /assets/{id}/refined — S1-mini cleaned transcript (derived/refined/{id}.md); requires HF_TOKEN worker secret + REFINE_MODEL var
Notes extraction GET /assets/{id}/notes — JSON {summary, notes, tasks, action_items} (derived/notes/{id}.json), LLM via EXTRACT_MODEL
Source github.com/DSamuelHodge/media-pipeline (R2 + D1 + Workflows + Whisper v3 Turbo + Firecrawl + S1-mini refine)

Kinds: image, video, audio, pdf. Images/video → 201 ready on arrival; audio/PDF → 202, poll /assets/{id}/status until ready|failed.

Steps

1. Upload

TOKEN=$(pass show cloudflare/media-pipeline/upload-token | head -1)
curl -X POST https://ingest.hodgeluke.com/upload \
  -H "Authorization: Bearer $TOKEN" \
  -F "file=@./meeting.m4a" -F "title=2026-09-05 staff sync"

2. Wait, then pull the derived text

curl https://ingest.hodgeluke.com/assets/<id>/status
curl https://ingest.hodgeluke.com/assets/<id>/transcript   # audio
curl https://ingest.hodgeluke.com/assets/<id>/refined      # audio, S1-mini cleaned
curl https://ingest.hodgeluke.com/assets/<id>/markdown      # pdf

Refine runs best-effort inside the audio pipeline (absent HF_TOKEN → the stage is skipped; failures never fail the asset). refined may legitimately be empty for filler-only audio; 409 until present.

3. Extract notes and tasks

From transcript/Markdown, produce frontmatter-compliant pages:

  • Meeting/voice notes → professional/briefings/YYYY/MM/DD-slug.md (decision section from the recording, link the source URL + transcript)
  • Action items → professional/commitments/ list lines (or Kaneo)
  • Videos/images → embed the media.hodgeluke.com link in the appropеlia section
  • Research PDFs → professional/ideas/ or the section the topic belongs to

4. Close the loop

  • Append a ## [date] entry (see repo log.md convention)
  • Run npm run graph (emits knowledge-graph.json + log.md in every build)
  • Any dangling/typo breaks CI before it ships

Lane notes

  • Whisper v3 Turbo (audio): transcript + VTT in derived/transcripts/.
  • S1-mini refine (shipped in media-pipeline, see S1 integration plan): the raw ASR text is cleaned (fillers out, numbers/dates/emails normalized) by superwhisper/s1-mini via HF Inference before note extraction.
  • Firecrawl → toMarkdown (PDF): research docs become clean Markdown; run the extraction step above on the result.
  • Shared folder: treat the pipeline as the shared inbox — anyone (Derrick or the CoS) with the upload token can drop media; both parties pull from /assets and file into the KB. No local synced folder required.
  • Secrets never in git: token comes from pass; keys live in Worker secrets/Secrets Store.

media-pipeline (ingest), macro (workspace context), pass (token), whisper/firecrawl (derived text)

Was this page helpful?