For a very long time, one of many widespread methods to start out new Node.js tasks was utilizing boilerplate templates. These templates assist builders reuse acquainted code constructions and implement normal options, reminiscent of entry to cloud file storage. With the most recent developments in LLM, venture boilerplates seem like extra helpful than ever.
Constructing on this progress, I’ve prolonged my current Node.js API boilerplate with a brand new software LLM Codegen. This standalone function allows the boilerplate to mechanically generate module code for any goal based mostly on textual content descriptions. The generated module comes full with E2E checks, database migrations, seed knowledge, and essential enterprise logic.
Historical past
I initially created a GitHub repository for a Node.js API boilerplate to consolidate the very best practices I’ve developed through the years. A lot of the implementation relies on code from an actual Node.js API operating in manufacturing on AWS.
I’m enthusiastic about vertical slicing structure and Clear Code ideas to maintain the codebase maintainable and clear. With current developments in LLM, significantly its assist for giant contexts and its potential to generate high-quality code, I made a decision to experiment with producing clear TypeScript code based mostly on my boilerplate. This boilerplate follows particular constructions and patterns that I imagine are of top of the range. The important thing query was whether or not the generated code would observe the identical patterns and construction. Based mostly on my findings, it does.
To recap, right here’s a fast spotlight of the Node.js API boilerplate’s key options:
- Vertical slicing structure based mostly on
DDD&MVCideas - Companies enter validation utilizing
ZOD - Decoupling software elements with dependency injection (
InversifyJS) - Integration and
E2Etesting with Supertest - Multi-service setup utilizing
Dockercompose
Over the previous month, I’ve spent my weekends formalizing the answer and implementing the required code-generation logic. Beneath, I’ll share the small print.
Implementation Overview
Let’s discover the specifics of the implementation. All Code Era logic is organized on the venture root stage, contained in the llm-codegen folder, guaranteeing simple navigation. The Node.js boilerplate code has no dependency on llm-codegen, so it may be used as an everyday template with out modification.
It covers the next use instances:
- Producing clear, well-structured code for brand new module based mostly on enter description. The generated module turns into a part of the Node.js REST API software.
- Creating database migrations and increasing seed scripts with fundamental knowledge for the brand new module.
- Producing and fixing E2E checks for the brand new code and guaranteeing all checks move.
The generated code after the primary stage is clear and adheres to vertical slicing structure ideas. It consists of solely the required enterprise logic for CRUD operations. In comparison with different code technology approaches, it produces clear, maintainable, and compilable code with legitimate E2E checks.
The second use case entails producing DB migration with the suitable schema and updating the seed script with the required knowledge. This activity is especially well-suited for LLM, which handles it exceptionally effectively.
The ultimate use case is producing E2E checks, which assist affirm that the generated code works appropriately. In the course of the operating of E2E checks, an SQLite3 database is used for migrations and seeds.
Primarily supported LLM shoppers are OpenAI and Claude.
Methods to Use It
To get began, navigate to the basis folder llm-codegen and set up all dependencies by operating:
npm i
llm-codegen doesn’t depend on Docker or some other heavy third-party dependencies, making setup and execution simple and easy. Earlier than operating the software, be sure that you set a minimum of one *_API_KEY atmosphere variable within the .env file with the suitable API key on your chosen LLM supplier. All supported atmosphere variables are listed within the .env.pattern file (OPENAI_API_KEY, CLAUDE_API_KEY and so on.) You should use OpenAI, Anthropic Claude, or OpenRouter LLaMA. As of mid-December, OpenRouter LLaMA is surprisingly free to make use of. It’s attainable to register right here and acquire a token free of charge utilization. Nonetheless, the output high quality of this free LLaMA mannequin might be improved, as many of the generated code fails to move the compilation stage.
To begin llm-codegen, run the next command:
npm run begin
Subsequent, you’ll be requested to enter the module description and identify. Within the module description, you possibly can specify all essential necessities, reminiscent of entity attributes and required operations. The core remaining work is carried out by micro-agents: Developer, Troubleshooter, and TestsFixer.
Right here is an instance of a profitable code technology:
Beneath is one other instance demonstrating how a compilation error was fastened:
The next is an instance of a generated orders module code:
A key element is that you could generate code step-by-step, beginning with one module and including others till all required APIs are full. This strategy permits you to generate code for all required modules in only a few command runs.
How It Works
As talked about earlier, all work is carried out by these micro-agents: Developer, Troubleshooter and TestsFixer, managed by the Orchestrator. They run within the listed order, with the Developer producing many of the codebase. After every code technology step, a test is carried out for lacking recordsdata based mostly on their roles (e.g., routes, controllers, companies). If any recordsdata are lacking, a brand new code technology try is made, together with directions within the immediate concerning the lacking recordsdata and examples for every function. As soon as the Developer completes its work, TypeScript compilation begins. If any errors are discovered, the Troubleshooter takes over, passing the errors to the immediate and ready for the corrected code. Lastly, when the compilation succeeds, E2E checks are run. Each time a take a look at fails, the TestsFixer steps in with particular immediate directions, guaranteeing all checks move and the code stays clear.
All micro-agents are derived from the BaseAgent class and actively reuse its base technique implementations. Right here is the Developer implementation for reference:
Every agent makes use of its particular immediate. Try this GitHub hyperlink for the immediate utilized by the Developer.
After dedicating vital effort to analysis and testing, I refined the prompts for all micro-agents, leading to clear, well-structured code with only a few points.
In the course of the growth and testing, it was used with varied module descriptions, starting from easy to extremely detailed. Listed below are a number of examples:
- The module liable for library guide administration should deal with endpoints for CRUD operations on books.
- The module liable for the orders administration. It should present CRUD operations for dealing with buyer orders. Customers can create new orders, learn order particulars, replace order statuses or info, and delete orders which can be canceled or accomplished. Order should have subsequent attributes: identify, standing, positioned supply, description, picture url
- Asset Administration System with an "Belongings" module providing CRUD operations for firm property. Customers can add new property to the stock, learn asset particulars, replace info reminiscent of upkeep schedules or asset places, and delete information of disposed or bought property.Testing with gpt-4o-mini and claude-3-5-sonnet-20241022 confirmed comparable output code high quality, though Sonnet is dearer. Claude Haiku (claude-3–5-haiku-20241022), whereas cheaper and comparable in value to gpt-4o-mini, usually produces non-compilable code. General, with gpt-4o-mini, a single code technology session consumes a mean of round 11k enter tokens and 15k output tokens. This quantities to a price of roughly 2 cents per session, based mostly on token pricing of 15 cents per 1M enter tokens and 60 cents per 1M output tokens (as of December 2024).
Beneath are Anthropic utilization logs displaying token consumption:
Based mostly on my experimentation over the previous few weeks, I conclude that whereas there should still be some points with passing generated checks, 95% of the time generated code is compilable and runnable.
I hope you discovered some inspiration right here and that it serves as a place to begin on your subsequent Node.js API or an improve to your present venture. Ought to you will have options for enhancements, be happy to contribute by submitting PR for code or immediate updates.
In case you loved this text, be happy to clap or share your ideas within the feedback, whether or not concepts or questions. Thanks for studying, and completely happy experimenting!
UPDATE [February 9, 2025]: The LLM-Codegen GitHub repository was up to date with DeepSeek API assist. It’s cheaper than
gpt-4o-miniand gives almost the identical output high quality, but it surely has an extended response time and generally struggles with API request errors.
Until in any other case famous, all photographs are by the writer
