Portfolio patch 1.1

Admin for CMS

• GitHub authentication: Uses GitHub OAuth through nuxt-auth-utils, with approved administrator identities matched by username or primary email.
• Persisted access roles: Adds user and admin roles to the D1 user records through Drizzle ORM migrations. The authenticated session carries the assigned role.
• Server-enforced authorization: Requires administrator access for management APIs and checks that browser mutations originate from the same site.
• Document management: Provides PDF upload, preview, indexing, and deletion controls, including upload-progress feedback.
• Résumé publishing: Allows administrators to replace the downloadable résumé stored in Cloudflare R2 while retaining a stable public URL.
• Task inspection: Provides access to indexing history and individual task details through protected Nuxt pages.

Durable indexing

• Asynchronous submission: Creates an indexing task in D1, starts a Workflow instance, and returns HTTP 202 Accepted with the task details.
• Workflow execution: Introduces LibraryIndexingWorkflow, which invokes the ingestion pipeline through internal Nuxt handlers restricted to Workflow execution.
• Stage-by-stage progress: Tracks PDF reading, text extraction, chunking, embedding generation, D1 writes, vector publication, verification, activation, and cleanup.
• Automatic retries: Configures up to two retries for the indexing step, using exponential backoff with an initial five-second delay.
• Persistent task history: Records queued, processing, ready, and failed states, together with attempt counts, timestamps, page and chunk counts, extraction metadata, and error messages.
• Live administration feedback: Refreshes active task details every two seconds and displays progress, the current processing stage, and failures.
• Publication visibility: Reports vector-verification progress before activating the new document generation and cleaning up the previous revision.

Embedding Visualization

• Two-dimensional projection: Uses principal component analysis (PCA) to project 1,024-dimensional document embeddings onto two axes. Vectors are L2-normalized before projection to align with Vectorize’s cosine-similarity metric.
• Visual source identification: Displays document chunks as labeled points, colored by their source PDF page.
• Model and index context: Shows the embedding model, original vector dimensions, distance metric, and number of plotted vectors alongside the chart.
• Verified vector data: Retrieves embeddings from Cloudflare Vectorize and checks their document and ingestion-generation metadata before calculating coordinates.
• Responsive rendering: Built with Vue and Unovis, with loading placeholders and styling for light and dark themes.