How ajndkr/lanarky works
Project type and purpose
• Lanarky is a server-side web framework built on FastAPI, tailored to help developers spin up LLM microservices with zero vendor lock-in.
• It provides a core application class, routing abstractions, streaming support (HTTP & WebSockets), adapter modules for LLM providers (OpenAI, LangChain), plus logging, events, and utility code.
File structure for architectural clues
• Top-level package lanarky/ contains:
• applications.py – core Lanarky app class wrapping FastAPI
• adapters/ – subpackages for each provider: openai/, langchain/ with modules for dependencies, routing, callbacks, responses, utils
• clients.py – HTTP clients for invoking external LLM APIs
• websockets.py – WebSocket streaming support
• events.py & logging.py – event bus & logging hooks
• responses.py, utils.py – response normalization & utilities
• There are also examples/ (demo microservices), docs/, tests/, configuration files (GitHub workflows, MkDocs, pyproject).
README for key insights
• Emphasizes LLM-first, streaming, FastAPI compatibility
• Shows the user flow: instantiate Lanarky(), include an adapter router (e.g. OpenAIAPIRouter), define POST endpoint returning a Resource object.
Main components
a. Client (Browser / HTTP/WebSocket client)
b. Lanarky Application
• FastAPI server underneath
• Core modules:
• Application core (Lanarky)
• Routing layer
• Dependency injection (from FastAPI + adapter-provided deps)
• Event bus & logging hooks
c. Adapters
• OpenAI adapter package
• LangChain adapter package
d. Clients
• HTTP / SDK clients to external LLM providers
e. External LLM Services
• OpenAI API
• LangChain-managed backends
f. Streaming Components
• HTTP streaming
• WebSockets (via lanarky.websockets)
g. User’s Microservice Code
• Custom router functions returning Resource objects
h. Optional
• Persistence / FAISS index (as shown in examples)
• Middleware or other integrations
Relationships and data flows
• Client → HTTP/WebSocket → FastAPI Server (Lanarky)
• FastAPI routes invoke adapter router
• Adapter router uses dependencies to build a Resource (system, user, stream flag)
• Resource handed off to clients.py which calls external LLM API
• Response streaming flows back through adapter → Lanarky → client
• Event bus captures events, logging hooks record metrics
Architectural patterns & design principles
• Plugin / Adapter pattern for provider-agnostic LLM calls
• Dependency Injection via FastAPI’s DI system
• Streaming pipeline abstracted over HTTP & WebSockets
• Unopinionated core: minimal footprint, zero vendor lock-in
• Layer separation: Core app vs. adapter vs. external system