Infrastructure

Jeremy runs entirely on Cloudflare's developer platform. This page covers each service, how Jeremy uses it, and relevant configuration details.

Workers

Cloudflare Workers provides the serverless compute layer for Jeremy's API and frontend.

0ms cold start -- Workers run on Cloudflare's edge network with no cold start penalty.
Handles all HTTP requests -- API routes, authentication, static asset serving, and ingestion orchestration.
Compatibility -- requires the nodejs_compat compatibility flag for Node.js API support.

D1 is Cloudflare's distributed SQLite database.

Stores structured data -- libraries, documentation chunks, user accounts, sessions, and API keys.
SQLite semantics -- full SQL support with transactions, indexes, and foreign keys.
Migrations -- schema changes are applied with wrangler d1 migrations apply.

See Database Schema for the complete table definitions.

Vectorize is Cloudflare's vector database for similarity search.

768-dimension vectors -- matches the output of the bge-base-en-v1.5 embedding model.
Cosine similarity -- vectors are compared using cosine distance for semantic matching.
Metadata filtering -- queries are filtered by libraryId so searches are scoped to a single library.
Index configuration -- created with wrangler vectorize create docs-index --dimensions=768 --metric=cosine.

Workers AI runs machine learning models on Cloudflare's infrastructure.

Model -- @cf/baai/bge-base-en-v1.5, a text embedding model that produces 768-dimension vectors.
Batch processing -- supports embedding up to 100 texts per API call, used during bulk ingestion.
Used for both ingestion and query -- the same model embeds documentation chunks during ingestion and search queries at query time, ensuring consistent vector representations.

R2 is Cloudflare's S3-compatible object storage.

Raw chunk backup -- stores the original chunk data from each ingestion run.
Recovery -- enables restoring chunks without re-fetching from the documentation source.
S3-compatible API -- accessible via the Workers R2 binding or standard S3 clients.

Browser Rendering provides headless Chromium instances on Cloudflare.

Puppeteer integration -- accessed via the @cloudflare/puppeteer package using the BROWSER binding.
JavaScript rendering -- loads documentation pages with full JavaScript execution, capturing content that isn't available in the raw HTML.
Used for URL ingestion -- when ingesting from a web URL (as opposed to llms.txt), Browser Rendering ensures dynamically rendered content is captured.