G-Drive Ingestion Pipeline

Clean start, dedicated to Zuzu. Jonathan creates and shares with Amy, Vince, Rachael.

Use a subfolder of an existing HHE shared drive. Less setup, more permission complexity.

Organization-wide drive. Most formal, but requires admin setup.

Runs 24/7 without human login. Amy shares folder with the service account email. No token expiration.

Requires periodic re-authentication. More familiar, but needs human intervention when token expires.

Simplest but limited to public files only. Not suitable for private HHE content.

Example: cell-talk_2026-02-15_transcript.pdf. Pipeline extracts metadata from filename. LLM validates.

Upload anything. LLM extracts all metadata from content. Most flexible, but slower and less reliable.

Organize by course slug subfolders (e.g., IN/cell-talk/). Filename is freeform within.

Highest member impact. These three represent >70% of all chunks. Quality improvement felt immediately.

Amy and team know which content members ask about most. May differ from enrollment numbers.

Most recent recordings from Rachael and Vince. Tests the pipeline with freshest material.

Simplest and most predictable. Three tiers: current semester full weight, previous slightly reduced, older still searchable.

Smooth decline over time. More granular but harder to reason about.

Fast initial drop, long tail. Most aggressive for favoring new content.

Content authority matters. A 2025 lecture loaded yesterday is still 2025 content. Reflects actual knowledge currency.

Simpler to track. But a bulk load of old content would appear “new.”

Easy to extract but unreliable — file copies and edits change this date.

Reduce old version’s weight from 1.0 to 0.3. Zuzu prefers current but can still reference history if asked.

All content treated equally forever. Simplest but risks outdated guidance ranking alongside current.

Clean but lossy. If a user asks “what was the previous recommendation?” Zuzu has no answer.

Balanced: files appear in Zuzu within 15 min of upload. Not burning Google API quota.

Near real-time. Higher API usage. Best if content producers need immediate availability.

Minimal API usage. Files may sit for up to an hour before processing.

Jonathan triggers ingestion when notified. Maximum control, zero automation.