Improve this page

Architecture notes

If you’re interested in contributing, here’s some general architectural notes that will hopefully help you find your way around the code base. (By the way, if you’d like to see more detail about specific aspects or topics, please let use know by creating a GitHib issue in the repo.)

Project anatomy

API Extractor’s code is separated into source code folders that reflect subsystems that can be arranged into a rough overall operational flow.

  • src/cli - the command-line interface (CLI) that gets things started

  • src/api - this folder contains the public API such as Extractor and ExtractorConfig. The CLI invokes these APIs the same way that an external consumer would; it doesn’t use any special internals. The TypeScript compiler gets configured in this stage, producing the ts.Program object that will be used below.

  • src/collector - The Collector acts as a central orchestrator that runs many of the stages below. Conceptually it is “collecting” all the API information in a central place, primarily CollectorEntity objects. This folder also has the MessageRouter class that routes errors and warnings based on the "messages" table from api-extractor.json.

  • src/analyzer - the core analyzer, which traverses the TypeScript compiler’s abstract syntax tree (AST) and produces the higher-level representations used by API Extractor. There are 4 major pieces of tech here:

    • The AstSymbol and AstDeclaration classes, which mirror the compiler’s ts.Symbol and ts.Declaration classes. The difference is that an AstDeclaration node is only generated for a subset of interesting nodes (e.g. classes, enums, interfaces, etc.) that will become “API items” in the documentation website and its api-extractor-model representation. This condensed tree omits all the intermediary ts.Declaration nodes (e.g. extends clauses, : tokens, and so forth). The AstSymbol.ts code comments provide some more detail about this very important data structure.

    • The ExportAnalyzer, which traverses chains of TypeScript import statements, eliminating the intermediary symbol aliases to build a flattened view as seen in the .d.ts rollup. The problem is that the compiler’s API makes it difficult to detect when this traversal leaves the working package (e.g. hops into the node_modules folder or compiler’s runtime library). That’s why this file has special handling for each kind of import syntax. The export * from construct is by far the most complicated form.

    • The Span class, which is a fairly lame but fairly effective utility for rewriting TypeScript source code while ignoring most of its meaning except for specific node types that we recognize. API Extractor does not use the compiler’s emitter to write .d.ts files, partially because those API were not public when we started, but also because it more faithfully preserves the original .d.ts inputs. The DtsRollupGenerator._modifySpan() function is a good illustration of how Span is used.

    • The AstReferenceResolver: Given a TSDoc declaration reference, this walks the the AstSymbolTable to find whatever it refers to.

  • src/enhancers - After the Collector has collected all the API objects and their metadata, we run a series of additional postprocessing stages called enhancers. The current ones are ValidationEnhancer (which applies some API validation rules) and DocCommentEnhancer which tunes up the TSDoc comments, for example expanding the @inheritDoc references.

  • src/generators - This folder implements API Extractor’s famous 3 output types: ApiReportGenerator, DtsRollupGenerator, and ApiModelGenerator.

  • src/schemas - This folder contains the api-extractor init template file, the JSON schema for api-extractor.json, and api-extractor-defaults.json which represents the default values for api-extractor.json settings.

Data flow

Another useful way to understand API Extractor is by examining what happens to a declaration as it gets transformed by each stage. Consider a simple function declaration that has two overloads:

import { Report } from 'reporting-package';

/** Declaration 1 */
export declare function add(report: Report, amount: number): void;

/** Declaration 2 */
export declare function add(report: Report, title: string): void;

Here’s how it gets processed:

  • Compiler stage: The TypeScript compiler engine parses the .d.ts file into two ts.Declaration objects (one for each overload) representing the parsed syntax. The compiler’s analyzer then makes an associated ts.Symbol which represents the function’s type. Each TypeScript type always becomes exactly one symbol, and in this case with two associated declarations (the two overloads). There will also be many “aliases” for this symbol. For example, if we write import { add } from "./math", the word add here becomes a symbol alias whose declaration is that import statement. If we follow the chain of symbol aliases (perhaps through many imports and exports), we will always reach the unique “followed symbol” corresponding to the original real definition of add().

  • Analyzer stage: API Extractor starts from your API entry point and follows each export to find its “followed symbol”. Then we make an AstSymbol and two AstDeclaration’s for add(). The analyzer also walks up and down the AST tree to fill out the context. For example, if the AstSymbol is a class, then we’ll create a child AstSymbol for each of its members. And if the class belongs to a namespace, then a parent AstSymbol is added representing the namespace.

    While following import statements, if we reach an external NPM package, the analysis stops there and produces an AstImport instead of a regular AstSymbol. This is because API Extractor understands package boundaries, and in fact is designed to be invoked separately on each project. Thus, in the above example, Report would become an AstImport instead of an AstSymbol. The analyzer’s overall job is to pick through the extremely detailed compiler data structures and produce a simplified tree of AstSymbol objects. This algorithm is the most complex stage of API Extractor, so we try to keep it isolated and single-purpose.

  • Collector stage: The collector builds the inventory of things that will end up as top-level items in the .d.ts rollup. We call these CollectorEntity objects, and there is one for our add() function, and another one for the Report import. So AstSymbol and AstImport can become a CollectorEntity. But note that AstDeclaration cannot, nor can AstModule (the analyzer’s representation of a .d.ts source file). To keep this straight, the analyzer’s objects inherit from the AstEntity base class if-and-only-if they can become a CollectorEntity. The CollectorEntity wraps AstEntity and appends some additional collector stage information:

    • Whether the entity is an export of your .d.ts rollup or just a local declaration.
    • The local name in the .d.ts rollup, since local declarations may need to get renamed by DtsRollupGenerator._makeUniqueNames() to avoid naming conflicts
    • The export name(s) which can be different from the local name. For example: export { A as B, A as C }.
  • Enhancers stage: The enhancers mostly work with the DeclarationMetadata, ApiItemMetadata, and SymbolMetadata objects. These objects are stored on AstSymbol and AstDeclaration, but they are entirely owned by the collector stage.

  • ApiReportGenerator and DtsRollupGenerator: These generators essentially just dump the CollectorEntity items into a big text file, but with different formatting. Other than trimming items according to release type, they don’t do much processing.

  • api-extractor-model stage: The @microsoft/api-extractor-model package is completely independent and does not rely on any of the other API Extractor types described above. It defines the portable .api.json file format. It has its own rich hierarchy inheriting from the ApiItem base class (mixin inheritance actually): ApiClass, ApiNamespace, ApiParameter, etc. In our example, the add() function will become an ApiFunction item in this representation. This model is designed to make it easy for third parties to generate documentation without having to understand the thorny compiler data structures. Thus the ApiModelGenerator takes our CollectorEntity for add() and converts it into an ApiFunction that will get serialized into .api.json.

    Recall that the analyzer internally used the AstReferenceResolver helper to look up TSDoc declaration references and find the target AstDeclaration. For the .api.json files, @microsoft/api-extractor-model provides an analogous ModelReferenceResolver helper that looks up ApiItem targets.

  • API Documenter stage: Okay, one final transformation happens here. It’s the last one! :-) When API Documenter loads up the .api.json file, it does not render it directly to .md files. First it converts the ApiFunction for our add() example function into a tree of TSDoc DocNode elements. Normally DocNode is used to represent doc comments. But it happens to be a full DOM-like structure that can represent rich text. Since the TSDoc comment for add() is already this kind of rich text, API Documenter cleverly reuses this representation to model an entire web page. This intermediate representation enables the markdown emitter to be decoupled from the documentation engine, and makes it easy in the future to output other formats such as HTML or React.

To summarize, for the humble add() function this pipeline produced a number of different representations:

  • AstDeclaration for the overload declarations
  • AstSymbol for the TypeScript type
  • CollectorEntity for the entry in the .d.ts file
  • DeclarationMetadata, ApiItemMetadata, and SymbolMetadata to annotate the symbol and declaration with more info
  • ApiFunction for the .api.json file
  • DocNode subtree for the documentation website