README

PHP symbol extraction for PapiAI ingestion: the Depth::Api ontology.

Reduces a PHP file to what it declares and what it reaches for, so a model can see the shape of a whole system without paying for its method bodies.

Install

composer require papi-ai/ingest-php-symbols

Usage

Pass it to the ingestor and ask for Depth::Api:

use PapiAI\Ingest\Depth;
use PapiAI\Ingest\IngestRequest;
use PapiAI\Ingest\NativeIngestor;
use PapiAI\Ingest\PhpSymbols\PhpSymbolExtractor;

$ingestor = new NativeIngestor(new PhpSymbolExtractor());
$digest = $ingestor->ingest(new IngestRequest('/path/to/repo', Depth::Api, tokenBudget: 4000));

echo $digest->toPrompt();

What a block looks like

papi-ai/ingest's own NativeIngestor.php, 1,500 lines of file reduced to this:

namespace PapiAI\Ingest;

class NativeIngestor implements RepositoryIngestorInterface
  __construct(?SymbolExtractorInterface $extractor = ..., TokenEstimatorInterface $estimator = ...)
  supports(IngestRequest $request): bool
  ingest(IngestRequest $request): RepositoryDigest
-> HeuristicTokenEstimator, IngestRequest, PathFilter, RepositoryDigest, RepositoryIngestorInterface,
   SymbolExtractorInterface, TokenEstimatorInterface, TreeBuilder

The -> line is the point. Signatures tell you what a class accepts; the edges tell you what it is built out of, including types it only ever instantiates internally, which no signature would reveal.

Measured across papi-ai/ingest's own source: 486 tokens at Api against 8,763 at Full, or 5.5%. A repository that would not fit in a context window at all fits comfortably, whole.

What it reports

Namespace, once, at the top.
Every class, interface, trait and enum, with what it extends and implements.
Public method signatures only, with parameter and return types, variadics marked and defaults shown as present rather than spelled out. Private and protected methods are implementation: they cost tokens without telling a reader how the pieces connect.
Top-level functions.
Outgoing edges: types in signatures, typed properties, and classes reached with new. Sorted, deduplicated, with scalars and self/static/parent left out.

Types print as they were written in the source, not fully qualified, because RepositoryDigest says as much as PapiAI\Ingest\RepositoryDigest to anyone reading the file it came from, for a quarter of the tokens.

Behaviour

Deterministic: the same bytes always give the same block. A file that does not parse yields no block rather than an exception, so it still appears in the tree with its path and size intact.

License

MIT, Marcello Duarte

papi-ai / ingest-php-symbols

Maintainers

Package info

Statistics

Security