AI based code generation pipeline

A command line tool and framework for systematic code generation using AI

In silence, code weaves,
Through prompts, AI breathes life anew,
Scripts bloom, knowledge leaves.

Git guards every step,
In the dance of creation,
Change blooms, watched and kept.
– ChatGPT

Basic idea

This is a command line tool to generate files using an AI - either ChatGPT or a model with a similar chat completion interface. It can be used to generate code, documentation, or other text files. Each run of the command line tool can take several prompt files with instructions what to generate as argument, but can also take other source files to be processed as further input. The output is written to a text file.

It can be used to solve complex tasks with several steps by chaining several runs of the tool - e.g. create an OpenAPI document from a specification, then generate an interface from that, a test and an implementation class. Of course, manual inspection and editing of the generated files will usually be necessary.

I suggest to inspect the intermediate and final results and committing them into a version control system like Git. That ensures manual checks when they are regenerated, and minimizes regeneration.

If you have questions, please don't hesitate to contact me , if you have fun you can also ask the Helper for the AI based Code Generation Pipeline which is an OpenAI GPT that is fed with all the documentation, or - if you have set up the OpenAI API key for it - you can ask it questions about itself with aigenpipeline --helpai <question>.

Some example usages

Basic Generation: Generating an OpenAPI document interface, tests, and implementation class could be achieved by chaining runs of AIGenPipeline, each with the appropriate prompt and input files. For example:
```
aigenpipeline -p openapi_prompt.txt -o generated_openapi.yaml
aigenpipeline -p interface_prompt.txt -i generated_openapi.yaml -o generated_interface.java
aigenpipeline -p test_prompt.txt -i generated_interface.java -o generated_test.java
```
By the way: if you want to change an existing output file, you can add the argument -upd. Then the current content of the output file will be given to the AI, and it will be instructed to check and update the file according to the new input files, minimizing the changes.

Explanation / Query: After generating an output, if there are questions or the need for clarification on how to improve it:

aigenpipeline -p interface_prompt.txt -i generated_openapi.yaml -o generated_interface.java --explain \
     "Why wasn't a @GET annotation used in method foo? How would I have to change the prompt to make sure it's 
used?"

Force Regeneration: Normally a versioning mechanism (see below) ensures the result is not regenerated unless the input changes. The -f flag disables this checking:
```
aigenpipeline -f -o specs/openapi.yaml api_interface_prompt.txt src/main/java/foo/MyInterface.java
```
Alternatively, the output files could just be removed before running the tool, or the version comments in suitable input files or prompt file(s) could be changed.
Generate parts of a file: If you want to combine manually written and ai generated parts in one file, you can use the -wp <marker> option to replace a part of the output file. The marker should occur in exactly two lines of the already existing output file - the lines between them are replaced by the AI generated text. The first line must also contain the version commment (see below).
```
aigenpipeline -wp generatedtable -o outputfile.md -p maketablefromjson.txt input.json
```
could e.g. replace the part between the lines  and  in:
```
This is the hand generated part

| Column1 | Column2 |
|---------|---------|
| value1  | value2  |

Here can be more handwritten stuff that is untouched.
```

Caching and versioning

Since generation takes time, costs a little and the results often have to be manually checked, the tool takes precaution not to regenerate the output files unless the input changes.

In the simplest case it can just do nothing when the output file is already there. If input files are changed, the output file would have to be removed to enforce regeneration.

Another work mode is checking whether the output file is newer than the input files (e.g. Makefile style usage) But this has the problem that that does not sensibly work if the output files are checked in with a version control system with Git, and would also regenerate files in case of minor changes in the input files.

The main suggested work mode (which is also the default) is to provide the input and prompt files with version comments that declare the version of the input file, and only generate the output file if it is not present, or regenerate the output file if the versions have changed. A version comment in the output file will be generated, which declares a version for the output file (a hash of it's content) and the versions of the used input files. The version comment in input and prompt files can be manually changed to force regeneration of the output file.

Structure of version comments

A version comment can e.g. look like this:

/* AIGenVersion(ourversion, inputfile1@version1, inputfile2@version2, ...) */

To declare a version for a manually created prompt / source file you can just put in a version comment like AIGenVersion(1.0) and change the version number each time you want to force regeneration.

The comment syntax (in this case /* */) is ignored - we just look for the AIGenVersion via regular expression. A version comment will be written at the start or end of the output file; that and the used comment syntax is determined by the file extension.

Hint files (given with the --hint option) are not used for version checking, but only to give temporary additional information to the AI, such as instructions to focus on specific things for an update. Use with care.

Using different large language models

While the tool defaults to using the OpenAI chat completion service, it is possible to use other services / LLM as well. I tried with Anthropic Claude text generation and some local models run with the nice LM Studio. See using other models for some examples.

Configuration files

The tool can read configuration files with common configurations (e.g. which AI backend to use). These should simply contain command line options; we'll split it at whitespaces just like in bash. Also, there can be an environment variable AIGENPIPELINE_CONFIG that can contain options.

Configuration files can be given explicitly (option -cf / --configfile) or the tool can scan for files named .aigenpipeline upwards from the output file directory. The search for .aigenpipeline files can be switched off with the -cn / --confignoscan option. If that option is given in one of these configuration files, that aborts the scanning further upwards in the directory tree.

The order these configurations are processed is: environment variable, .aigenpipeline files from top to bottom, command line arguments. Thus, the later override the earlier one. Explicitly given configuration files are processed at the point where the argument occurs when processing the command line arguments. The option -cp / --configprint gives an overview of the used files / sources of configuration.

Command files

While the startup time of aigenpipeline is low in comparison to the actual LLM calls, it can still hurt if there are many files to process and most AI calls can be skipped because there were no changes in inputs or prompts. Thus, if you just give one file as argument, it'll be read s command file containing a number of command lines that are executed in sequence. Those contain a number of command lines that are executed in sequence. Empty lines separate individual command lines, and lines starting with a # are ignored (comments). For example:

#!/usr/bin/env ../../bin/aigenpipeline
-m gpt-4o-mini -p 0dialogelements.prompt README.md -o dialogelements.txt

# command lines are separated by empty lines, and comment lines starting with # are ignored
-m gpt-4o-mini -p 1html.prompt README.md dialogelements.txt -o differentialReTranslation.html

# Each command line can be split over several lines if convenient.
-m gpt-4o-mini -p 2css.prompt 
README.md differentialReTranslation.html 
-o differentialReTranslation.css

Other features

If you are not satisfied with the result, the tool can also be used to ask the AI for clarification: ask a question about the result or have it make suggestions how to improve the prompt. This mode will not write the output file, but recreate the conversation that lead to the output file and ask the AI for a clarification or suggestion in form of a chat continuation.

It's also possible to ask the tool questions about itself - use the -ha / --helpai option with a question, and it tries to answer that from this documentation. There is also an OpenAI GPT that can be asked.

How can I download or install it?

The download and installation links are given on the release pages on Github. You can either:

build it from the source,
grab a zip with the command line script (incl. a prebuilt jar) from the command line package (on maven central)
or use the framework jar directly with the class AIGenerationTask .

Limitations and additional ideas (not implemented yet)

This means the resulting file is always completely regenerated or not at all. How to make differential changes? One idea would be to give the previous file to the AI and ask it to make minimal changes.

Command Usage

Usage:
aigenpipeline [options] [<input_files>...]
or
ai-gen-pipeline <command_file>

The AIGenPipeline tool generates content using an AI based on a prompt and input files.
It can also update or improve existing content, and it only calls the AI if the input or prompt files have changed.

If it's called with a command file, it reads a number of command lines from that file. Empty lines separate individual command lines.
Lines starting with a # are ignored (comments). This saves the startup time when calling the tool multiple times.

Options:

  General options:
    -h, --help               Show this help message and exit.
    -ha, --helpai <question> Answer a question about the tool from the text on its documentation site and exit.
                             Use as last argument - rest of command line counts as question.
    --version                Show the version of the AIGenPipeline tool and exit.
    -c, --check              Only check if the output needs to be regenerated based on input versions without actually 
                             generating it. The exit code is 0 if the output is up to date, 1 if it needs to be 
                             regenerated.
    -n, --dry-run            Enable dry-run mode, where the tool will only print to stderr what it would do without 
                             actually calling the AI or writing any files.
    -v, --verbose            Enable verbose output to stderr, providing more details about the process.

  Input / outputs:
    -o, --output <file>      Specify the output file where the generated content will be written.
    -ifp, --infileprompt <marker> <file>  The output and the prompt are in the same file, the marker is used in separating the parts.
    -upd, --update           Gives the current content of the output as hint to the AI that it should be updated / improved.
    --hint <file>            Gives this file as additional clue to the AI (special filename - is stdin), e.g. to tell
                             it to focus on something for an --update. This is not used for version checking.
    -p, --prompt <file>      Reads a prompt from the given file.
    -s, --sysmsg <file>      Optional: Reads a system message from the given file instead of using the default.
    -k <key>=<value>         Sets a key-value pair replacing ${key} in prompt files with the value. 
    -os, --outputscan <pattern>  Searches for files matching the ant-like pattern and scans them for AIGenPromptStart markers.
                             The infile prompts in these files are processed (see -ifp).
    -dd, --dependencydiagram Print a dependency diagram (Mermaid graph) of the scanned files and exit.

  AI Generation control:
    -f, --force              Force regeneration of output files, ignoring any version checks - same as -ga.
    -ga, --gen-always        Generate the output file always, ignoring version checks.
    -gn, --gen-notexists     Generate the output file only if it does not exist.
    -go, --gen-older         Generate the output file if it does not exist or is older than any of the input files.
    -gv, --gen-versioncheck  Generate the output file if the version of the input files has changed. (Default.)
    -wv, --write-version     Write the output file with a version comment. (Default.)
    -wvf, --write-versionfile Write the version comment to a separate file named like the output file with .version appended.
    -wo, --write-noversion   Write the output file without a version comment. Not compatible with default -gv .
    -wp, --write-part <marker> Replace the lines between the first occurrence of the marker and the second occurrence.
                             If a version marker is written, it has to be in the first of those lines and is changed there.
                             It is an error if the marker does not occur exactly twice; the output file has to exist.
    -e, --explain <question> Asks the AI a question about the generated result. This needs _exactly_the_same_command_line_
                             that was given to generate the output file, and the additional --explain <question> option.
                             It recreates the conversation that lead to the output file and asks the AI for a 
                             clarification. The output file is not written, but read to recreate the conversation.
                             Use as last argument - rest of command line counts as question.

  Configuration files:
    -cf, --configfile <file> Read configuration from the given file. These contain options like on the command line.
    -cn, --confignoscan      Do not scan for `.aigenpipeline` config files.
    -cne, --configignoreenv  Ignore the environment variable `AIGENPIPELINE_CONFIG`.
    -cp, --configprint       Print the collected configurations and exit.

  AI backend settings:
    -u, --url <url>          The URL of the AI server. Default is https://api.openai.com/v1/chat/completions .
                             In the case of OpenAI the API key is expected to be in the environment variable 
                             OPENAI_API_KEY, or given with the -a option.
    -a, --api-key <key>      The API key for the AI server. If not given, it's expected to be in the environment variable 
                             OPENAI_API_KEY, or you could use a -u option to specify a different server that doesnt need
                             an API key. Used in "Authorization: Bearer <key>" header.
    -org, --organization <id> The optional organization id in case of the OpenAI server.
    -m, --model <model>      The model to use for the AI. Default is gpt-4o .
    -t <maxtokens>           The maximum number of tokens to generate.

Arguments:
  [<input_files>...]       Input files to be processed into the output file. 

Examples:
  Generate documentation from a prompt file:
    aigenpipeline -p prompts/documentation_prompt.txt -o generated_documentation.md src/foo/bar.java src/foo/baz.java

  Force regenerate an interface from an OpenAPI document, ignoring version checks:
    aigenpipeline -f -o specs/openapi.yaml -p prompts/api_interface_prompt.txt src/main/java/foo/MyInterface.java

  Ask how to improve a prompt after viewing the initial generation of specs/openapi.yaml:
    aigenpipeline -o PreviousOutput.java -p prompts/promptGenertaion.txt specs/openapi.yaml --explain "Why did you not use annotations?"

  Scan for files with infile prompts and (re-)generate the AI generated parts of those files:
    aigenpipeline -os "src/site/**/*.md"

Infile prompts:
  The idea is that files can contain both the prompt that has been used to generate their AI generated part(s) in a
  comment, and also instructions like the used input files or other settings. In such a file you would have e.g. 
  
  <!-- AIGenPromptStart(somemarker)
  (Here would come the prompt for the AI)
  AIGenCommand(somemarker)
  data.txt
  AIGenPromptEnd(somemarker) -->
  (Here is the generated content placed after calling aigenpipeline.)
  <!-- AIGenEnd(somemarker) -->
  
  That also means that it's not necessary to write a script that processes each of those files, but the tool can scan
  for AIGenPromptStart markers, as in the example `aigenpipeline -os "src/site/**/*.md"` above.

Configuration files:
  These contain options like on the command line. The environment variable `AIGENPIPELINE_CONFIG` can contain options.
  If -cn is not given, the tool scans for files named .aigenpipeline upwards from the output file directory.
  The order these configurations are processed is: environment variable, .aigenpipeline files from top to bottom,
  command line arguments. Thus the later override the earlier one, as these get more specific to the current call.
  Lines starting with a # are ignored in configuration files (comments).

Note:
  It's recommended to manually review and edit generated files. Use version control to manage and track changes over time. 
  More detailed instructions and explanations can be found at https://aigenpipeline.stoerr.net/ .