autodocgenerator 1.4.9.5

This Project helps you to create docs for your projects

Project Overview – Auto‑Doc Generator

---

1. Project Title

Auto‑Doc Generator – A layered, orchestrated pipeline that creates a complete README.md (or any markdown documentation) from a source‑code repository by automatically chunking the code, prompting a large‑language model (LLM) for descriptive fragments, post‑processing the results, and enriching the output with vector embeddings.

---

2. Project Goal

Develop a hands‑free documentation tool that can be run locally or as a GitHub Action.
The software scans a repository, compresses and splits the source into manageable fragments, asks a Groq‑hosted LLM to produce markdown snippets for each fragment, optionally re‑orders those snippets using embeddings, caches intermediate results, and finally emits a polished README.md.

The tool solves two recurring problems for developers and CI pipelines:

1. Manual, out‑of‑date documentation – documentation is generated directly from the current state of the code base, ensuring it never lags behind.
2. Time‑consuming, error‑prone doc writing – the LLM handles the natural‑language summarisation while the pipeline guarantees reproducibility, caching and progress reporting.

---

3. Core Logic & Principles

Layer / Component	Responsibility	Core Principle
Entry (CLI / Action) autodocgenerator.auto_runner.run_file.__main__	Parses autodocconfig.yml, builds configuration objects, creates the central Manager, and starts the pipeline.	Single entry point –‑ deterministic start‑up, works both locally (python -m …) and in CI.
Orchestrator autodocgenerator.manage.Manager	Holds Config, CacheSettings, LLM (GPTModel), embedding service, UI (logger & progress), and coordinates every stage.	Centralised state machine –‑ all shared objects flow through the manager, enabling incremental runs via cache.
Pre‑processor	• CodeMix.build_repo_content() – walks the repo, applies ignore patterns, builds a single string with file‑level markers. • compressor.compress_to_one() – optional global summarisation. • spliter.split_data() – chops the huge string into ≤ max_symbols chunks.	Chunk‑first, compress‑later –‑ ensures the LLM never receives payloads larger than the provider limits while keeping context boundaries.
Engine (LLM Wrapper) engine.models.gpt_model.{GPTModel, AsyncGPTModel}	Thin wrapper around the Groq API exposing ask / ask_async. Handles multiple API keys, model fallback and request history.	Provider‑agnostic interface –‑ the rest of the code only needs ask(prompt).
Factory factory.base_factory.DocFactory + concrete BaseModule subclasses	Plug‑in system that creates additional markdown sections (intro, custom modules, link tables, etc.). Each module receives the shared info object and the LLM instance, returns a markdown fragment.	Extensibility –‑ new documentation pieces are added by implementing a subclass and exposing it in the YAML config.
Post‑processor	• postprocessor.custom_intro – generates a custom introductory block. • postprocessor.sorting – extracts anchors, asks the LLM for a CSV ordering, optionally re‑orders via vector similarity.	Semantic ordering –‑ the final document follows a logical flow rather than raw chunk order.
Embedding Layer postprocessor.embedding.Embedding	Calls the Google Gemini embedding API, stores a dense vector on each DocContent. The vectors are later used to compute similarity to a root vector for ordering.	Content‑driven similarity –‑ sections that talk about the same concept appear together.
Schema / Cache	Pydantic models (CacheSettings, DocHeadSchema, DocContent) persisted as .auto_doc_cache_file.json.	Incremental builds –‑ if the repository has not changed, the manager re‑uses cached fragments, saving API calls and time.
UI ui.progress_base.ConsoleGitHubProgress & ui.logging.BaseLogger	Console progress bar (Rich‑compatible) and structured logging (debug / info / error).	Visibility –‑ users see real‑time progress and detailed logs, both locally and in CI.
Config config.config.Config & engine.config constants	Holds global settings, environment‑variable validation (GROQ_API_KEYS, GOOGLE_EMBEDDING_API_KEY, GITHUB_EVENT_NAME), prompt templates, thresholds, feature flags.	Centralised, declarative configuration –‑ all behaviour can be toggled from autodocconfig.yml.

Functional Flow (high‑level)

1. Initialisation – CLI reads the YAML, validates env‑vars, builds a Config object and instantiates Manager.
2. Git status check – Manager.check_git_status() decides whether a fresh run is required (based on the last processed commit stored in CacheSettings).
3. Source aggregation – CodeMix creates a single markdown‑ish representation of the repo (CacheSettings.code_mix).
4. Optional global compression – compressor.compress_to_one summarises the whole repo into a “global info” chunk (CacheSettings.global_info).
5. Chunking – split_data produces size‑bounded fragments. Each fragment is sent to the LLM (GPTModel.ask) and the returned markdown is stored as a DocContent in DocHeadSchema.
6. Factory‑driven sections – All BaseModule subclasses (intro, custom links, user‑defined modules) generate additional markdown fragments that are merged into the same schema.
7. Ordering – If enabled, the LLM is asked to propose a CSV order of section titles; anchors are extracted and, optionally, a vector‑based similarity sort refines the order.
8. Embedding – Each DocContent is vectorised via the Google Embedding API; vectors are kept in memory and can be persisted for downstream tooling.
9. Cache clean‑up – Mutable temporary fields in CacheSettings are cleared to keep the cache file small.
10. Persist output – Manager.save() writes the final markdown to .auto_doc_cache/output_doc.md and updates the JSON cache; a CI step may copy the file to the repository root as README.md.

All stages read and write to the shared CacheSettings and DocHeadSchema objects, guaranteeing a single source of truth throughout the run.

---

4. Key Features

• Full‑repo scan with configurable ignore patterns (files, directories, extensions).
• Automatic chunking respecting a maximum token / symbol limit for LLM calls.
• LLM‑driven summarisation using Groq‑hosted models; supports key rotation and model fallback.
• Plug‑in factory for custom markdown modules (intro, link tables, user‑defined sections).
• Optional global compression to produce an overarching project description.
• Semantic re‑ordering via LLM‑generated CSV ordering and optional vector‑similarity sorting.
• Embedding generation with Google Gemini embedding API; vectors stored per section for future retrieval or similarity search.
• Caching layer (.auto_doc_cache_file.json) that stores intermediate results, enables incremental builds, and reduces API usage.
• CLI & GitHub Action entry points –‑ one command works both locally and in CI pipelines.
• Progress & logging UI (Rich‑based progress bar, structured logger) for transparent execution.
• Extensible architecture –‑ add new sections by subclassing BaseModule; swap LLM or embedding providers by implementing the same interface.

---

5. Dependencies

Category	Packages	Purpose
Core runtime	python >=3.9	Primary interpreter.
LLM access	groq (or the underlying HTTP client)	Calls the Groq LLM API (ask, ask_async).
Embedding	google-generativeai (Gemini embedding endpoint)	Generates 768‑dimensional vectors for each markdown fragment.
Data models & validation	pydantic	Typed schemas (CacheSettings, DocHeadSchema, DocContent).
CLI framework	cleo (or typer)	Provides the python -m autodocgenerator.auto_runner.run_file command interface.
Progress & logging	rich	Console progress bar and colourful logs.
File system utilities	pathlib, yaml (PyYAML)	Reads autodocconfig.yml, traverses the repository.
HTTP / async support	httpx (optional, used by Groq wrapper)	Async requests to the LLM API.
Testing (optional)	pytest, pytest-mock	Unit‑test suite for the pipeline.
CI integration	No additional packages; the entry point is invoked from a reusable GitHub Action workflow (reuseble_agd.yml).

All dependencies are pure‑Python and available on PyPI. The project can be installed via a standard pip install -e . after cloning the repository.

---

In summary, the Auto‑Doc Generator is a modular, cache‑aware pipeline that turns any source‑code repository into a high‑quality markdown documentation file with minimal human effort. Its layered architecture, clear separation of concerns, and plug‑in points make it easy to adapt to new LLM providers, embedding services, or custom documentation sections.

Executive Navigation Tree

• 📂 Initialization

  • init-welcome
  • git-status-evaluation
  • configuration-parsing
  • projectsettings-class
  • pyproject-toml
  • autodoc-setup-options

• ⚙️ Installation

  • install-script
  • install-sh
  • install-workflow-setup

• 📂 Manager

  • manager-class
  • manager-init-folder-system
  • manager-class-usage

• 📂 Custom Modules

  • custom-modules-implementation
  • intro-modules-implementation

• 📂 Model

  • model-base
  • base-module-interface
  • gptmodel-construction
  • gptmodel-generate-answer

• 📂 Embedding

  • embedding-class
  • embedding-ordering
  • vector-sort-helpers
  • get-order

• 📂 Generation

  • generate-code-file
  • generate-global-info
  • doc-factory-constructor
  • doc-factory-generate-doc
  • factory-generate-doc
  • gen_doc-function
  • gen_doc-logic-flow
  • gen_doc-data-contract
  • gen-doc-parts
  • generete-doc-parts
  • write-docs-by-parts
  • custom-description-loop
  • standalone-custom-description
  • CONTENT_DESCRIPTION
  • global-intro-generation
  • link-intro-generation
  • html-link-extraction

• 📂 Parsing

  • "\'?(.*?)"\'?
  • extract-links-from-start
  • split-text-by-anchors
  • parse-answer
  • spliter-function

• 📂 Compression

  • compress-function
  • compress-and-compare
  • compress-to-one

• 📂 Schema & Utilities

  • schema-classes
  • codemix-class
  • logging-infrastructure
  • progress-abstraction

• 📂 Finalization

  • finalize
  • have-to-change

Welcome Banner & Logger Instantiation

Functional Role
The module prints a colored ASCII logo and status line when the package is imported, then creates a global logger instance for the whole library.

Visible Interactions

• Uses print (stdout) for the banner – no external I/O.
• Imports BaseLogger and related classes from autodocgenerator.ui.logging to construct logger.
• Exposes logger at package level so downstream modules can from autodocgenerator import logger and share a single configured logger.

Step‑by‑Step Logic

1. Define _print_welcome – local helper.
2. Inside, set ANSI colour/format constants.
3. Compose ascii_logo string with colour codes.
4. Print logo and a status line showing library name, version, and ready state.
5. Call _print_welcome() immediately on import.
6. Import logger classes from .ui.logging.
7. Instantiate BaseLogger → logger.
8. Attach a BaseLoggerTemplate via logger.set_logger to configure format/level.

Data Contract

Entity	Type	Role	Notes
_print_welcome	function	Emits banner on import	No parameters, no return value
BLUE, BOLD …	str	ANSI escape sequences	Used only inside the function
ascii_logo	str	Formatted logo text	Multi‑line string
logger	BaseLogger (instance)	Global logging object	Accessible as autodocgenerator.logger
BaseLoggerTemplate	class	Logging format/template	Passed to logger.set_logger
print	built‑in	Output side‑effect	Writes to standard output

Critical Note – The banner prints every time the package is imported, which may be undesirable in non‑interactive contexts (e.g., automated CI). Adjust by guarding the call with an environment flag if needed.

Git Status Evaluation (autodocgenerator.auto_runner.check_git_status)

Functional Role – Determines whether the repository has changed since the last documented commit and instructs the Manager to rebuild documentation accordingly.

Visible Interactions

Entity	Type	Role	Notes
Manager	class instance	Receives CacheSettings.last_commit; calls manager.check_sense_changes	Imported from autodocgenerator.manage
CacheSettings	pydantic model	Stores last_commit; mutated in‑place	Imported from autodocgenerator.schema.cache_settings
CheckGitStatusResultSchema	pydantic model	Returned result (need_to_remake, remake_gl_file)	Imported from same module
GITHUB_EVENT_NAME	str env constant	Bypasses diff check for manual workflow runs	Imported from engine.config.config
subprocess	stdlib	Executes git commands	Used in get_diff_by_hash, get_detailed_diff_stats, get_git_revision_hash

Logic Flow

1. Environment guard – If GITHUB_EVENT_NAME == "workflow_dispatch" or manager.cache_settings.last_commit is empty, set last_commit to current HEAD hash and force a full rebuild (need_to_remake=True, remake_gl_file=True).
2. Otherwise, call get_detailed_diff_stats with the stored hash to collect per‑file change stats.
3. Pass the list of dicts to manager.check_sense_changes, which decides if a partial or full regeneration is required.
4. Return the CheckGitStatusResultSchema produced by the manager.

Note – The function assumes git is available and the working directory is the repository root; no fallback is implemented.

---

Configuration Parsing (autodocgenerator.auto_runner.config_reader)

Functional Role – Loads autodocconfig.yml content, builds the runtime Config, a list of custom modules, and a StructureSettings object that controls downstream pipeline behavior.

Visible Interactions

Entity	Type	Role	Notes
yaml.safe_load	function	Parses raw YAML string	Imported from yaml
Config / ProjectBuildConfig	classes	Hold global settings, ignore patterns, additional info	Imported from ..config.config
BaseModule, CustomModule, CustomModuleWithOutContext	classes	Represent user‑defined documentation fragments	Imported from autodocgenerator.factory.modules.general_modules
StructureSettings	class (local)	Toggles features like intro links, ordering, global file usage	Instantiated per run
list[BaseModule]	runtime list	Ordered collection of modules to feed the DocFactory	Constructed from custom_descriptions

Logic Flow

1. yaml.safe_load converts the file text to a dict.
2. Core fields (ignore_files, language, project_name, project_additional_info) populate a fresh Config instance via fluent setters.
3. Each ignore pattern is registered with config.add_ignore_file.
4. Project‑specific key‑value pairs are added through config.add_project_additional_info.
5. custom_descriptions is transformed into a list of BaseModule subclasses: entries beginning with % become CustomModuleWithOutContext, others become CustomModule.
6. structure_settings dict is applied to a new StructureSettings instance via load_settings.
7. The tuple (config, custom_modules, structure_settings_object) is returned for the Manager to consume.

Critical Warning – No validation is performed on the shape of custom_descriptions; malformed entries may raise runtime errors.

ProjectSettings – Prompt Builder

Entity	Type	Role	Notes
project_name	str (ctor)	Identifier injected into the system prompt.
info	dict[str, str]	Arbitrary key‑value pairs added via add_info.
Property prompt	str	Concatenates BASE_SETTINGS_PROMPT, project name, and all info entries (each on its own line).

Logic Flow

1. Initialise with project_name.
2. add_info(key, value) stores custom metadata.
3. Accessing prompt builds the final system‑prompt string on‑the‑fly.

pyproject.toml – Package Definition

Entity	Type	Role	Notes
[project]	metadata table	Describes the Python package (name, version, description, authors, license, readme, Python requirement).	requires‑python = ">=3.11,<4.0".
dependencies	list	Runtime libraries required by Auto‑Doc Generator.	Includes groq, google‑genai, rich, etc.
[tool.poetry]	configuration table	Excludes the cache file from distribution.	exclude = [".auto_doc_cache_file.json"].
[build-system]	table	Build backend specification for PEP 517.	Uses poetry-core.

Data Contract – When the project is built, the build system reads pyproject.toml to resolve the exact version constraints listed under dependencies. No runtime code interacts with this file; it serves solely as static package metadata. The file is a YAML document that defines the behavior of the documentation generator. The top‑level keys and their possible values are:

project_name – a string that sets the name of the project shown in the generated documentation.

language – the language code (e.g., en) used for the output text.

ignore_files – a list of glob patterns and directory names that the generator will skip. Typical entries include build folders (dist), Python byte‑code caches (*.pyc, __pycache__), virtual‑environment directories (venv, .env), IDE configuration folders (.vscode, .idea), database files, log files, coverage reports, version‑control metadata, static assets, and any markdown files you do not want to process.

build – a subsection containing parameters that control the generation process:

• save_logs – boolean (true/false) indicating whether logs should be persisted.
• log_level – numeric level (e.g., 2) that sets the verbosity of logging.
• threshold_changes – an integer that defines the change size limit (in characters) for triggering a full regeneration.

structure – a subsection that shapes the layout of the final document:

• include_intro_links – boolean to add navigation links at the start.
• include_intro_text – boolean to include introductory explanatory text.
• include_order – boolean to preserve the order of processed files.
• use_global_file – boolean to merge content into a single global file.
• max_doc_part_size – maximum number of characters per documentation segment.

project_additional_info – a mapping for extra project metadata. In the example a global idea entry provides a short description of the project’s purpose.

custom_descriptions – a list of free‑form strings that the generator will incorporate as custom sections. These can be instructions, usage guides, or any other explanatory paragraphs you want to appear in the output.

When creating the file, follow standard YAML syntax: use proper indentation (two spaces per level) and enclose strings in quotes if they contain special characters. Ensure each top‑level key is present (or omitted if defaults are acceptable) and provide the desired values according to the descriptions above.

install.ps1 – CI Bootstrap Generator

Entity	Type	Role	Notes
.github/workflows/autodoc.yml	file (generated)	GitHub Actions workflow that re‑uses a remote reusable workflow.	Inserts secret GROCK_API_KEY.
autodocconfig.yml	file (generated)	Default configuration for the Auto‑Doc Generator.	Populated with project name, language, ignore patterns, and build/structure flags.
PowerShell commands	script	Creates workflow directory, writes the two files, and echoes a success message.	Uses here‑strings (@' … '@) to avoid variable expansion.

Logic Flow

1. Ensure .github/workflows exists (New-Item -Force).
2. Write static workflow YAML to autodoc.yml.
3. Derive the current folder name, embed it in a YAML config string, and write autodocconfig.yml.
4. Output a green “Done!” banner.

Information not present in the provided fragment – No validation of write permissions or error handling for I/O failures.

install.sh – CI Bootstrap Script

Entity	Type	Role	Notes
.github/workflows/autodoc.yml	generated file	GitHub‑Actions workflow that re‑uses the remote reuseble_agd.yml and injects the secret GROCK_API_KEY.	Uses a here‑document (cat <<EOF).
autodocconfig.yml	generated file	Default configuration for the Auto‑Doc Generator; contains project name, language, ignore patterns and build/structure flags.	project_name is derived from basename "$PWD".
mkdir -p .github/workflows	command	Guarantees the target directory exists before writing files.	Idempotent.
echo "✅ Done! …"	command	User feedback on successful creation of each file.	No error handling.

Assumption – The script runs with write permission in the repository root; any I/O error is not caught.

Logic Flow

1. Ensure .github/workflows exists.
2. Write a static workflow YAML to autodoc.yml, escaping the first $ so the secret placeholder remains intact.
3. Emit a success banner.
4. Write autodocconfig.yml with a YAML block that populates ignore lists and flags, interpolating the current directory name.
5. Echo a second success banner.

The script does not validate the generated content, nor does it check for existing files before overwriting.

To set up the installation workflow for both Windows PowerShell and Linux‑based environments, follow these steps:

Windows (PowerShell)

1. Run the remote installer
   Execute the following command in an elevated PowerShell session:

   irm https://raw.githubusercontent.com/Drag-GameStudio/ADG/main/install.ps1 | iex
   

   This command fetches the PowerShell installation script directly from the repository and pipes it to the PowerShell interpreter for immediate execution.

2. Verification
   After the command completes, confirm that the required components have been installed by checking the presence of the expected binaries or by running a version check command provided by the script.

Linux / macOS (Bash)

1. Run the remote installer
   In a terminal, issue the following command:

   curl -sSL https://raw.githubusercontent.com/Drag-GameStudio/ADG/main/install.sh | bash
   

   The curl request downloads the Bash installer script and streams it directly to bash for execution.

2. Verification
   Once the script finishes, verify the installation by invoking any provided test commands or by confirming that the installed executables are available in your PATH.

GitHub Actions Integration

To automate the installation within a GitHub Actions workflow, you must provide an API key from the Grock service as a secret:

1. Create the secret

   • Navigate to your repository’s Settings → Secrets and variables → Actions.
   • Add a new secret named GROCK_API_KEY.
   • Paste the API key you obtained from the Grock documentation (see https://grockdocs.com).

2. Reference the secret in the workflow
   In your workflow YAML, you can expose the secret to steps that need it:

   env:
     GROCK_API_KEY: ${{ secrets.GROCK_API_KEY }}
   

   Ensure any scripts or commands that interact with the Grock API reference this environment variable.

Summary of Commands

Platform	Command
PowerShell (Windows)	irm https://raw.githubusercontent.com/Drag-GameStudio/ADG/main/install.ps1 | iex
Bash (Linux/macOS)	curl -sSL https://raw.githubusercontent.com/Drag-GameStudio/ADG/main/install.sh | bash

By following the above steps, you will have a reproducible installation process for both local development and CI pipelines, with the required API key securely supplied via GitHub Actions secrets.

Manager – Orchestrator Core

Entity	Type	Role	Notes
project_directory	str	Root of the repo to document
config	Config	Holds ignore patterns, language, logging flags
llm_model	Model	Groq‑based LLM client used throughout the pipeline
embedding_model	Embedding	Google embedding wrapper for vectorising sections
progress_bar	BaseProgress	Tracks overall and sub‑task progress
logger	BaseLogger	Writes info/warn/error logs to report.txt
doc_info	DocInfoSchema	In‑memory container for code_mix, global_info, doc
cache_settings	CacheSettings	Persistent JSON cache (.auto_doc_cache_file.json)	Loaded/updated in init_folder_system

Manager.__init__ – Construction

1. Instantiates DocInfoSchema and stores injected dependencies.
2. Creates a file logger (FileLoggerTemplate) pointing to report.txt.
3. Calls init_folder_system to ensure /.auto_doc_cache exists and loads/creates the cache JSON.

init_folder_system – Cache Bootstrap

• Creates cache directory if missing.
• Writes a fresh CacheSettings JSON when the cache file does not exist.
• Deserialises the file into self.cache_settings via CacheSettings.model_validate_json. !noinfo

Custom Modules (CustomModule, CustomModuleWithOutContext)

Class	Constructor Arg	generate Behaviour
CustomModule	discription: str	Calls generete_custom_discription(split_data(...), model, self.discription, language)
CustomModuleWithOutContext	discription: str	Calls generete_custom_discription_without(model, self.discription, language)

Both rely on external generete_custom_discription* helpers; the fragment supplies only the call signatures.

Intro Modules (IntroLinks, IntroText)

Class	generate Steps
IntroLinks	Retrieves HTML links via get_all_html_links(info["full_data"]), then formats them with get_links_intro(links, model, language).
IntroText	Produces introductory text via get_introdaction(info["global_info"], model, language).

Data Contract Table

Entity	Type	Role	Notes
info	dict	Input context (keys used: code_mix, language, full_data, global_info)	Missing keys result in None passed to helpers.
model	Model	LLM client for all helper calls	No direct usage shown here.
Return	str	Markdown fragment for the respective module	Inserted into DocHeadSchema by DocFactory.

Warning – The fragment does not validate presence of required keys; callers must ensure info contains them. ## Model & ParentModel – Shared Contract

Entity	Type	Role	Notes
ParentModel	abstract base	Stores api_keys, history, rotation state, and enforces abstract methods.	models_list shuffled if use_random.
History	class	Holds system_prompt and mutable history list.	add_to_history(role, content) appends messages.
Model	concrete subclass	Provides thin wrappers: get_answer_without_history → generate_answer; get_answer adds user prompt to history then calls generate_answer.	Default generate_answer returns "answer" (placeholder).

Assumption – All logging classes (InfoLog, WarningLog, ErrorLog) and the Groq client behave as their names imply; no internal details are inferred beyond the shown calls.

BaseModule Abstract Interface

Entity	Type	Role	Notes
BaseModule	class (ABC)	Blueprint for plug‑in generators	Sub‑classes must implement generate(info: dict, model: Model)

Assumption – The abstract method returns a string representing a markdown fragment; the exact format is not enforced by the snippet. ## GPTModel – LLM Wrapper Construction

Entity	Type	Role	Notes
GPTModel	class (subclass of Model)	Instantiates a Groq client, loads API keys, prepares model rotation, attaches logger.	api_key defaults to GROQ_API_KEYS; models_list shuffled when use_random=True.
self.client	Groq	Performs chat.completions.create calls.	Re‑created on key rotation.
self.logger	BaseLogger	Emits InfoLog / WarningLog / ErrorLog.	Created per instance.

The constructor stores history, api_keys, regen_models_name (shuffled model list) and sets index counters (current_model_index, current_key_index). No external I/O occurs beyond client init. ## GPTModel.generate_answer – Prompt Execution Logic

Entity	Type	Role	Notes
with_history	bool	Determines whether to prepend self.history.history to the request.	If False and prompt supplied, messages = prompt.
messages	list[dict]	Payload sent to Groq.	Contains role/content pairs.
model_name	str	Selected model from regen_models_name.	Rotated on failure.
chat_completion	Groq response	Holds choices[0].message.content.	Returned as result.
result	str	Final LLM answer.	Logged at level 2; empty string returned if None.

Logic flow

1. Log start.
2. Choose message source based on with_history.
3. Enter retry loop:
   • Fail‑fast if regen_models_name empty → ModelExhaustedException.
   • Attempt self.client.chat.completions.create(messages=messages, model=model_name).
   • On exception: log warning, rotate API key (current_key_index) and, if wrapped, rotate model index. Re‑instantiate Groq with new key, repeat.
4. Extract result, log success and raw answer, return it (or "" if None).

Embedding – Gemini Vectoriser

Entity	Type	Role	Notes
api_key	str (ctor arg)	Auth for Google GenAI	Stored in self.client
self.client	genai.Client	API wrapper
get_vector(prompt)	list[float]	Calls embed_content with model gemini‑embedding‑2‑preview (768‑dim)	Raises Exception if embeddings is None

Warning – The method returns list(text_response.embeddings[0])[0][1], assuming the first embedding element is a tuple‑like structure; any format change will break the call.

create_embedding_layer & order_doc – Vectorisation & Re‑ordering

• Iterates over self.doc_info.doc.parts and calls init_embedding(self.embedding_model) to attach embeddings.
• Calls get_order with the LLM to obtain a new ordering list, then assigns it back to content_orders.

Helper Functions – Vector Distance & Sorting

• bubble_sort_by_dist(arr) – classic bubble sort on a list of (id, distance) tuples.
• get_len_btw_vectors(v1, v2) – Euclidean norm via np.linalg.norm.
• sort_vectors(root_vector, other) – Computes distance from root_vector to each vector in other (dict id → vector), returns IDs ordered by ascending distance.

All functions are pure and return plain Python collections; they do not log.

get_order – LLM‑Driven Title Re‑ordering

Requests the LLM to sort a list of section titles semantically.

Entity	Type	Role	Notes
model	Model (sub‑class of ParentModel)	LLM backend providing get_answer_without_history	No history retained across calls
chanks	list[str]	Raw titles extracted from anchors	Passed verbatim into the prompt
Return	list[str]	Ordered titles (comma‑separated list trimmed)	Used later to align sections

Logic Flow

1. Log start via BaseLogger.
2. Build a user‑role prompt requesting a comma‑separated, exact list of sorted titles, preserving leading “#”.
3. Call model.get_answer_without_history(prompt).
4. Split the LLM response on commas, strip whitespace, produce new_result.
5. Log the final ordered list and return it.

Assumption – The LLM obeys the “return ONLY a comma‑separated list” instruction; any deviation will be propagated unchanged.

---

generate_code_file – Repo Snapshot

• Uses CodeMix(project_directory, config.ignore_files) to walk the repository and produce a single string (code_mix).
• Stores result in self.doc_info.code_mix.
• Logs start/end and advances progress_bar.

generate_global_info – Optional Global Summary

• If is_reusable and a cached global_info exists, re‑uses it.
• Otherwise splits code_mix with split_data(full_code_mix, max_symbols).
• Calls compress_to_one (LLM + progress) to obtain a compressed markdown fragment.
• Saves to self.doc_info.global_info and updates progress.

DocFactory.__init__ – Construction

Entity	Type	Role	Notes
modules	*BaseModule	Collection of user‑provided generators	Stored as self.modules
with_splited	bool	Controls post‑generation splitting	Default True
logger	BaseLogger	Centralised logger instance	Created via BaseLogger()

DocFactory.generate_doc – Core Logic

Entity	Type	Role	Notes
info	dict	Shared context (e.g., code_mix, language)	Passed unchanged to each module
model	Model	LLM client used by modules	No direct calls in this fragment
progress	BaseProgress	Tracks sub‑task progress	create_new_subtask, update_task, remove_subtask
doc_head	DocHeadSchema	Accumulator for generated parts	add_parts(key, DocContent)

Step‑by‑step flow

1. Initialise empty DocHeadSchema.
2. progress.create_new_subtask("Generate parts", len(self.modules)).
3. Iterate self.modules (index i, element module):
   • Call module.generate(info, model) → module_result.
   • If self.with_splited is True:
     • split_text_by_anchors(module_result) → splited_result (dict of anchor → fragment).
     • For each el in splited_result: doc_head.add_parts(el, DocContent(content=splited_result[el])).
   • Else: construct task_name = f"{module.__class__.__name__}_{i}" and add whole result.
   • Log two InfoLog entries (module success, raw output).
   • progress.update_task().
4. After loop, progress.remove_subtask().
5. Return populated doc_head.

factory_generate_doc – Plugin Module Execution

• Builds info dict (language, full_data, code_mix, global_info).
• Logs the module list and input keys.
• Invokes doc_factory.generate_doc(info, llm_model, progress_bar) – the fragment documented earlier.
• Prepends or appends the returned DocHeadSchema to the existing document based on to_start.

gen_doc – Orchestrator for Documentation Generation

Entity	Type	Role	Notes
project_path	str	Root directory of the repository to document	Passed unchanged to Manager.
config	Config (custom class)	Holds ignore patterns, language, project metadata	Created by read_config.
custom_modules	list[BaseModule]	User‑defined markdown generators	Instances of CustomModule or CustomModuleWithOutContext.
structure_settings	StructureSettings (local)	Flags controlling optional steps (global file, intro sections, ordering)	Populated by read_config.
Return	str	Full assembled markdown document or empty string when no rebuild needed	Obtained via manager.doc_info.doc.get_full_doc().

Step‑by‑Step Logic Flow

1. Model Instantiation – GPTModel receives GROQ_API_KEYS; Embedding receives GOOGLE_EMBEDDING_API_KEY.
2. Manager Construction – Manager(project_path, config, llm_model, embedding_model, progress_bar) creates the central orchestrator, storing all supplied objects.
3. Git Status Check – check_git_status(manager) returns a CheckGitStatusResultSchema with booleans need_to_remake / remake_gl_file.
4. Early Exit – If both flags are False, the function returns "" (no documentation rebuild).
5. Source Extraction – manager.generate_code_file() builds the raw code snapshot (code_mix).
6. Global Info (optional) – If structure_settings.use_global_file is true, manager.generate_global_info compresses the snapshot; the is_reusable flag is the inverse of remake_gl_file.
7. Chunked Documentation – manager.generete_doc_parts splits the code into chunks (size limited by structure_settings.max_doc_part_size) and queries the LLM for markdown fragments.
8. Custom Module Generation – manager.factory_generate_doc(DocFactory(*custom_modules)) runs each user‑provided BaseModule to produce additional markdown sections.
9. Optional Ordering – If structure_settings.include_order is true, manager.order_doc() re‑orders sections via an LLM‑driven pass.
10. Intro Sections (optional) – IntroText and/or IntroLinks are instantiated based on flags and injected at the document start via a second factory_generate_doc call (with_splited=False, to_start=True).
11. Embedding Layer – manager.create_embedding_layer() computes vector embeddings for all markdown parts.
12. Cache Cleanup – manager.clear_cache() resets mutable cache fields.
13. Persist & Return – manager.save() writes the final markdown and cache files; the assembled document string is returned.

Data Contract Summary

Entity	Type	Role	Notes
manager	Manager	Core pipeline controller	Holds CacheSettings, DocHeadSchema, progress logger, etc.
change_info	CheckGitStatusResultSchema	Result of Git diff analysis	Attributes need_to_remake: bool, remake_gl_file: bool.
structure_settings.use_global_file	bool	Toggles global‑file generation	Determines step 6.
structure_settings.max_doc_part_size	int	Maximum symbols per chunk for LLM calls	Controls step 7.
structure_settings.include_order	bool	Enables LLM‑based re‑ordering	Controls step 9.
structure_settings.include_intro_text / include_intro_links	bool	Controls inclusion of intro modules	Affects step 10.
manager.doc_info.doc	DocHeadSchema (contains DocContent parts)	Aggregated markdown fragments	get_full_doc() concatenates all parts.

Critical Assumption – The function assumes all imported classes behave as their names suggest; no internal details are inferred beyond what is visible in the snippet.

gen_doc_parts – Pipeline Driver for Chunked Documentation

Entity	Type	Role	Notes
full_code_mix	str	Whole repository snapshot produced by CodeMix.
max_symbols	int	Maximum characters per chunk for split_data.
model	Model	LLM used throughout the pipeline.
project_settings	ProjectSettings	Shared prompt context.
language	str	Desired output language.
progress_bar	BaseProgress	UI feedback for chunk processing.
global_info	Any	Forwarded to write_docs_by_parts.
Return	str	Concatenated markdown of the entire repository.

Logic Flow

1. Chunk the repo via split_data(full_code_mix, max_symbols).
2. Initialise a sub‑task on progress_bar.
3. Iterate chunks, calling write_docs_by_parts for each; accumulate results in all_result.
4. After each chunk, keep a 3000‑character tail of the current result to feed as prev_info for the next call (preserves context).
5. Update progress bar, finally remove sub‑task and log completion.

---

generete_doc_parts – Chunked Documentation

• Calls gen_doc_parts (LLM per chunk) with language and optional global_info.
• Splits the concatenated result into anchor sections via split_text_by_anchors.
• Inserts each section into self.doc_info.doc as DocContent.

write_docs_by_parts – Part‑wise LLM Documentation Generator

Entity	Type	Role	Notes
part	str	Source code fragment to document.
model	Model	LLM used for generation (get_answer_without_history).
project_settings	ProjectSettings	Supplies global system prompt.
prev_info	str | None	Previous fragment output, used to keep continuity.
language	str	Target language for the generated text (default en).
global_info	str | None	Optional additional project‑wide context.
Return	str	Generated markdown for the fragment.

Logic Flow

1. Log start via BaseLogger.
2. Build a system‑message list: language, global project info, static part template (BASE_PART_COMPLITE_TEXT), optional global_info and prev_info.
3. Append the user message containing part.
4. Call model.get_answer_without_history(prompt).
5. Strip surrounding markdown fences (```) if present and return the clean answer.

Assumption – model.get_answer_without_history always returns a string; no error handling is shown in the fragment.

---

generete_custom_discription – Conditional Chunk Description

Iterates over splited_data (iterable of strings). For each chunk it sends a detailed prompt containing the chunk, a custom description request, and BASE_CUSTOM_DISCRIPTIONS. The loop stops when the LLM returns a result that does not contain !noinfo or “No information found”, or when such markers appear after position 30.

Entity	Type	Role	Notes
splited_data	Iterable[str]	Source fragments
model	Model	LLM
custom_description	str	Task description supplied by the caller
language	str	Language selector
Return	str	First satisfactory description, empty if none

generete_custom_discription_without – Stand‑Alone Description

Creates a single‑anchor response (mandatory <a name="CONTENT_DESCRIPTION"></a> tag) that rewrites custom_description. No source context is given.

Entity	Type	Role	Notes
model	Model	LLM
custom_description	str	Text to be rewritten
language	str	Language selector
Return	str	LLM answer respecting strict tag rules

get_introdaction – Global Documentation Intro

Builds a prompt using BASE_INTRO_CREATE and asks the LLM for a high‑level introduction based on global_data.

Entity	Type	Role	Notes
global_data	str	Full repository summary (or similar)
model	Model	LLM backend
language	str	Language selector
Return	str	Intro markdown fragment

get_links_intro – LLM‑Driven Links Intro

Calls a Model (typically GPTModel) with a three‑message prompt to create an introductory paragraph that lists the supplied links.

Entity	Type	Role	Notes
links	list[str]	Input list of anchor strings
model	Model	LLM provider	Must implement get_answer_without_history
language	str	System‑prompt language selector (default "en")
Return	str	Generated markdown intro	Logged before and after the call

get_all_html_links – HTML Anchor Collector

Extracts anchor names from a markdown string.

Entity	Type	Role	Notes
data	str	Source documentation	Expected to contain <a name="…"></a> tags
links	list[str]	Return value	Each entry is prefixed with # and filtered to length > 5
logger	BaseLogger	Side‑effect	Logs start, count, and list at level 1
pattern	str (regex)	Internal	r'<a name=["\']?(.*?)["\']?></a>'

Assumption – The function does not validate duplicate anchors.

extract_links_from_start – Anchor Detection in Chunk List

Identifies leading <a name=…></a> tags and returns a list of markdown links plus a flag indicating whether the first non‑anchor chunk must be discarded.

Entity	Type	Role	Notes
chunks	list[str]	Raw text fragments supplied by the caller	Each element is stripped before inspection
Return	tuple[list[str], bool]	(links, have_to_del_first)	links are #anchor strings; flag is True when any chunk lacks a valid anchor

Logic Flow

1. Initialise links = [], have_to_del_first = False.
2. Iterate over chunks.
3. Use regex ^<a name=["']?(.*?)["']?</a> to capture the anchor name.
4. If a name longer than 5 characters is found → prepend “#” and append to links.
5. If a chunk yields no anchor → set have_to_del_first = True.
6. Return the tuple.

Warning – The function assumes the first anchor appears at the very start of a chunk; otherwise have_to_del_first may be incorrectly set.

---

split_text_by_anchors – Chunk Segmentation by Anchor Tags

Splits a full markdown document into a dictionary keyed by anchor links.

Entity	Type	Role	Notes
text	str	Complete markdown payload containing <a name=…></a> markers	May include leading non‑anchor content
Return	dict[str, str]	Mapping #anchor → chunk content	Keys derived from extract_links_from_start

Logic Flow

1. Regex (?=<a name=["']?[^"\'>\s]{6,200}["']?</a>) splits text while retaining delimiters.
2. Strip empty entries → result_chanks.
3. Call extract_links_from_start(result_chanks) → all_links, have_to_del_first.
4. If the first anchor appears far into the file (start_link_index > 10) or have_to_del_first is true, drop the first chunk (typically stray pre‑anchor text).
5. Verify len(all_links) == len(result_chanks); otherwise raise Exception("Somthing with anchors").
6. Build the result dict by pairing each link with its corresponding chunk.

Critical – Mismatch between detected links and chunks aborts the pipeline, ensuring anchor integrity.

---

parse_answer – Git‑Change Check Result Parser

Converts a pipe‑separated string into a typed schema.

Entity	Type	Role	Notes
answer	str	Expected format `"true	false"` etc.
Return	CheckGitStatusResultSchema	need_to_remake & remake_gl_file booleans	Instantiated directly

Logic Flow

1. splited = answer.split("|").
2. change_doc = splited[0] == "true"; change_global = splited[1] == "true".
3. Return schema with those booleans.

---

split_data – Size‑Bound Chunker

Entity	Type	Role	Notes
data	str	Full repository markdown (output of CodeMix).
max_symbols	int	Upper bound for each chunk’s character count.
Return	list[str]	Sequential fragments, each ≤ max_symbols.

Logic Flow (partial – file truncated)

1. Initialise split_objects.
2. Split data on the sentinel `"

compress – Chunk‑Level LLM Compression

Entity	Type	Role	Notes
data	str	Raw source fragment	May contain any file content.
project_settings	ProjectSettings	Supplies system prompt (prompt property).
model	Model	LLM wrapper exposing get_answer_without_history.
compress_power	int	Controls prompt‑generation intensity (passed to BASE_COMPRESS_TEXT).
Return	str	LLM‑produced summary of data.	Directly returned; no post‑processing.

Logic Flow

1. Assemble three messages: system → project prompt, system → compress‑size hint, user → data.
2. Call model.get_answer_without_history(prompt=prompt).
3. Return the raw answer string.

Assumption – The LLM obeys the “return ONLY a comma‑separated list” instruction; any deviation will be propagated unchanged.

compress_and_compare – Batch Compression & Merging

Entity	Type	Role	Notes
data	list[str]	Ordered fragments to compress.
model	Model	Same LLM used by compress.
project_settings	ProjectSettings	Shared prompt context.
compress_power	int (default 4)	Number of fragments merged per output slot.
progress_bar	BaseProgress	UI feedback; default instantiated.
Return	list[str]	Length = ⌈len(data)/compress_power⌉, each element = merged compressed text.

Logic Flow

1. Allocate an output list sized for the target groups.
2. Initialise a sub‑task on progress_bar.
3. Iterate data; for each element el compute curr_index = i // compress_power.
4. Append compress(el, …) + "\n" to the appropriate bucket.
5. Update progress; after loop, remove the sub‑task and return the bucket list.

compress_to_one – Recursive Global Summarisation

Entity	Type	Role	Notes
data	list[str]	Initial set of fragments (often output of split_data).
model	Model	LLM used for all compression steps.
project_settings	ProjectSettings	Global prompt source.
compress_power	int (default 4)	Base merging factor; may be reduced to 2 for small tails.
progress_bar	BaseProgress	UI feedback.
Return	str	Single markdown block representing the whole repository.

Logic Flow

1. Loop while len(data) > 1.
2. If the remaining list is shorter than compress_power + 1, set new_compress_power = 2; otherwise keep the original.
3. Replace data with compress_and_compare(data, …, new_compress_power).
4. Increment iteration counter.
5. When one element remains, return data[0].

Schema Classes – In‑Memory / Persistent Data Model

Class	Key Fields	Purpose
CacheSettings	last_commit: str, doc: DocInfoSchema	JSON‑persisted cache (.auto_doc_cache_file.json).
DocInfoSchema	global_info: str, code_mix: str, doc: DocHeadSchema	Holds raw repo text, optional global summary, and assembled doc parts.
DocHeadSchema	content_orders: list[str], parts: dict[str, DocContent]	Maintains ordered collection of generated markdown fragments.
DocContent	content: str, embedding_vector: list | None	Individual markdown block; can embed vectors via init_embedding.

Interaction Overview

• Manager (or other orchestrator) reads/writes CacheSettings to reuse previous runs.
• DocHeadSchema.add_parts(name, DocContent) is invoked by factories or write_docs_by_parts‑derived results.
• DocHeadSchema.get_full_doc() concatenates ordered parts for final output.

Warning – The fragment does not show persistence logic; it is assumed elsewhere that CacheSettings is serialized/deserialized.

CodeMix – Repository Snapshot Builder

Collects file‑system structure and file contents into a single markdown‑compatible string.

Entity	Type	Role	Notes
root_dir	str (ctor)	Base directory to walk	Resolved to absolute Path
ignore_patterns	list[str]	Glob patterns for exclusion	Defaults to [] or supplied list
Method should_ignore(path)	bool	Determines if path matches any ignore pattern	Checks full relative path, basename, and each part
Method build_repo_content()	str	Generates repository outline and file blocks	Returns a single string; logs ignored paths

Logic Flow

1. Initialise logger.
2. Append “Repository Structure:” header.
3. Walk root_dir.rglob("*") sorted; for each entry not ignored, compute depth → indentation, append directory/file name line.
4. Insert separator line ("="*20).
5. Second pass: for each file not ignored, emit <file path="..."> tag, then file text, then a stray newline placeholder ("\n"). Errors are captured as inline messages.
6. Join all pieces with newline characters and return.

Critical – The ignore logic uses fnmatch against the full relative path, basename, and each path component, ensuring comprehensive exclusion based on the supplied ignore_list.

BaseLogger & Log Templates

Entity	Type	Role	Notes
BaseLog	class	Holds raw message and numeric level; provides _log_prefix.	format() returns plain text; subclasses add level tags.
ErrorLog, WarningLog, InfoLog	subclasses of BaseLog	Format messages with [ERROR], [WARNING], [INFO] prefixes.	Use _log_prefix → timestamp.
BaseLoggerTemplate	class	Minimal logger; prints or writes formatted logs.	global_log respects log_level.
FileLoggerTemplate	subclass of BaseLoggerTemplate	Persists logs to a file path.	Opens file in append mode each call.
BaseLogger	singleton class	Central façade; holds a logger_template set via set_logger.	log() forwards to global_log.

Logic Flow

1. BaseLogger.__new__ guarantees a single instance.
2. Client creates a concrete template (e.g., FileLoggerTemplate) and registers it with BaseLogger.set_logger.
3. Calls to BaseLogger.log(ErrorLog("msg")) invoke logger_template.global_log, which prints or writes the prefixed string.

Assumption – No thread‑safety mechanisms are present; concurrent writes may interleave.

BaseProgress and Concrete Implementations

Entity	Type	Role	Notes
BaseProgress	abstract class	Defines UI API: create_new_subtask, update_task, remove_subtask.	Methods are stubs (...).
LibProgress	subclass	Wraps Rich Progress; tracks a base task and an optional sub‑task.	update_task advances current task or base task.
ConsoleGtiHubProgress	subclass	Simple console feedback via ConsoleTask.	Uses two ConsoleTask instances for general and sub‑tasks.
ConsoleTask	helper	Prints start message and incremental percent.	No external dependencies.

Logic Flow

1. Orchestrator instantiates a concrete progress (e.g., LibProgress).
2. For each pipeline stage it calls create_new_subtask(name, total_len).
3. After each unit of work update_task() is invoked, advancing the appropriate bar.
4. Upon completion remove_subtask() discards the sub‑task reference.

Warning – If create_new_subtask is never paired with remove_subtask, the base task may never finish.

save – Persist Output & Cache

• Writes the assembled markdown (self.doc_info.doc.get_full_doc()) to output_doc.md.
• Updates self.cache_settings.doc with the latest DocInfoSchema and rewrites the cache JSON.

Warning – The fragment does not perform explicit validation of keys inside info; callers must ensure required entries exist.

have_to_change – LLM‑Based Repository Change Evaluation

Queries the LLM whether documentation must be regenerated based on a diff and optional global info.

Entity	Type	Role	Notes
model	Model	LLM interface	Uses get_answer_without_history
diff	list[dict[str, str]]	Structured diff description	Inserted verbatim into the prompt
global_info	str | None	Optional repository‑wide summary	Added as a system message if present
Return	CheckGitStatusResultSchema	Result of parse_answer	Indicates doc rebuild needs

Logic Flow

1. Assemble a three‑message prompt: system prompt (BASE_CHANGES_CHECK_PROMPT), optional global info, and user diff.
2. Invoke LLM, obtain raw answer string.
3. Pass answer to parse_answer and return the schema.

---