Docker Configuration#

Blinko uses a three-stage multi-stage Dockerfile to produce a lean production image. The image is built from dockerfile at the repo root and is published to Docker Hub as blinkospace/blinko.

Build Stages#

Builder stage 1: runs bun install, bunx prisma generate, bun run build:web, and bun run build:seed. It also generates a start.sh entrypoint that runs Prisma migrations and seeds before starting the server.

Runner stage 2: copies only what is needed from builder:

• dist/ → ./server (compiled server bundle)
• server/lute.min.js (Markdown renderer)
• prisma/ and .prisma/client (schema + generated client)
• start.sh and dumb-init

node_modules from the builder stage is not copied. Instead, a smaller, explicit dependency set is installed at runtime via npm (see Runtime Dependencies below).

ARM Architecture Support#

Both the builder and runner stages detect ARM (aarch64 / arm64) and pre-install the platform-specific sharp@0.34.1 / @img/sharp-linux-arm64 binary before the main dependency installation 3 4. The SHARP_IGNORE_GLOBAL_LIBVIPS=1 environment variable and custom mirror URLs ensure the native binding is fetched reliably during the build 5.

Mirror / Registry Override#

A build argument USE_MIRROR (default false) switches npm/bun to use the Taobao mirror (https://registry.npmmirror.com). This is useful in China or other environments with poor access to the default npm registry 6 7.

Runtime Dependencies#

The runner stage installs the following packages from scratch via npm 8:

@node-rs/crc32 lightningcss sharp@0.34.1 prisma@5.21.1
sqlite3@5.1.7
llamaindex @langchain/community@0.3.40 (via --legacy-peer-deps)
@libsql/client @libsql/core

The --legacy-peer-deps flag is required because llamaindex and @langchain/community declare optional peer dependencies that are not installed in the runtime stage. The Dockerfile comment explains this directly:

"legacy peer deps avoids optional peer conflicts from LangChain." 9

Known omission: pdf-parse#

@langchain/community's PDFLoader lists pdf-parse as an optional peer dependency. Because it is not installed in the runtime stage, PDF parsing will fail at runtime with:

ERR_MODULE_NOT_FOUND
Cannot find package 'pdf-parse' imported from
/app/node_modules/@langchain/community/dist/document_loaders/fs/pdf.cjs

This is a known open issue 10. The workaround is to install it manually inside the running container:

cd /app && npm install pdf-parse

Until this is included in the Dockerfile, PDF attachment indexing for RAG is broken in the default Docker image 10.

Startup Sequence#

The container entrypoint is dumb-init wrapping start.sh 11. At container start, start.sh runs:

1. npx prisma migrate deploy — applies pending DB migrations
2. node server/seed.js — seeds initial data
3. node server/index.js — starts the application on port 1111

12

Compose Files#

The dev compose file uses USE_MIRROR: "true" by default 13. The prod compose file uses host network mode and includes log rotation (max-size: 10m, max-file: 3) 14.

.dockerignore#

.dockerignore excludes node_modules, build artifacts (dist/, .next/, .turbo/), test directories, logs, documentation, and macOS/IDE metadata from the build context. The one exception is **/dist-js/*, which is explicitly allowed 15.

Key Environment Variables (Runtime)#

16 17