Download the PHP package survos/followthemoney without Composer

survos/followthemoney

Based on FollowTheMoney — original project and documentation and the original Python library, alephdata/followthemoney. Browse the upstream schema explorer for the model we reuse. This PHP package is maintained independently by Survos.

PHP 8.5 library for the FollowTheMoney entity model, with no Python process, Symfony kernel, database or search service required at runtime. MIT licensed.

The package ships all 70 schemas from upstream revision c70a89ffa59ea785956f6095643310da3735b7fb, including resolved inheritance, reverse properties, relationship edges and display metadata. The snapshot and its SHA-256 are in resources/. This is an independent PHP implementation, not an official FollowTheMoney release or a complete port of its Python tools.

Installation

Install the published package on Packagist:

Requires PHP 8.5, ext-intl and ext-mbstring. No Symfony or Python runtime is required. For Symfony applications that process JSONL files, also install survos/jsonl-bundle.

For development inside the Survos monorepo, the mono root already registers the library's namespace and includes its tests. Other local applications may use a Composer path repository pointing to mono/lib/followthemoney.

Entities and relationships

All properties are ordered, deduplicated lists of strings, including dates, numbers and references. name reads the first value and replaces the name list on assignment; names accesses/replaces the entire name property. Neither parses a personal name into given/family names. Use get(), add() and set() for every schema property. Updates are atomic when a cleaner throws. Schema and property metadata are immutable; the model can be loaded once and reused.

Identifiers are supplied by the application. Do not use a person's name alone as an identity key. This library does not infer that similar names identify the same person, or infer gender from names.

Import, normalization and validation

$model->fromArray($wire) imports already normalized FtM wire data, preserving string values and optional top-level metadata. Unknown properties and writes to reverse stubs fail. Use the existing survos/jsonl-bundle for streaming files, gzip and sidecars:

The bundle is an optional application dependency. It owns file IO; this library owns the FtM model and row conversion. There is no duplicate JSONL reader/writer.

For raw input use add(), set() or fromArray($wire, normalized: false). As in Python, a cleaner returning null drops an invalid/empty input; unknown properties, unsupported cleaners and self-references throw. For applications that must report rejected input, call normalizer->clean($property, $raw) first and retain the original input and rejection separately. Normalized imports do not silently re-clean historical or externally supplied values.

validate() reports missing required properties, excessive lengths and malformed reference IDs. Pass an ID resolver to check dangling references and schema ranges:

It is a structural validator, not a claim that arbitrary imported strings have passed every upstream type validator. Passing Entity objects to add() checks the target schema immediately. ID-only references are checked with the resolver. Incomplete extraction fragments can exist before validation.

Cleaning compatibility

Python-generated fixture tests cover the implemented cleaning profile:

• Unicode NFC and unsafe-character removal; strings, text, HTML, numbers and checksums.
• Names and addresses, without OCR spelling correction or title inference.
• Prefix dates, including partial precision, following prefixdate 0.6.1.
• Entity IDs, gender aliases, email addresses and Wikidata Q-identifiers.
• Country/language codes and English labels from the pinned upstream/Babel tables; topic codes.
• Basic HTTP/HTTPS/FTP URLs, including a default HTTP scheme and empty path.

This is not universal normalization parity. Arbitrary multilingual country names, fuzzy matching, advanced URL repairs, phone/IP/IBAN/MIME/JSON cleaning, non-QID identifier formats and custom date formats are not implemented. Already-normalized values of these types can be imported and exported without loss. Raw unsupported types/formats throw rather than being treated as plain text. Register a cleaner with normalizer->register($type, callable) when an application needs one. A cleaner receives the raw value and Property, and returns a string or null.

Raw HTML remains HTML. Escape values in views; this library does not sanitize HTML for browser rendering. The properties wire object and returned arrays are copies.

Custom schemas

Custom schemas support multiple parents, inherited metadata and entity ranges. An entity property may declare reverse: ['name' => 'mentions', 'label' => 'Mentions']; the corresponding read-only stub is generated on its range and descendants. Cycles, unknown parents/types/ranges, replacement of upstream schemas and reverse name collisions fail. Extensions use native PHP arrays; Symfony YAML can supply the same definitions without adding a YAML dependency to this library.

OCR evidence

TextEvidence stores a W3C Web Annotation with TextPositionSelector and TextQuoteSelector, plus a separate provenance envelope carrying the exact source SHA-256, extractor version and optional confidence. Offsets are Unicode code points, end exclusive. Keep the original OCR text unchanged and retrieve the matching source version before applying offsets. ALTO boxes/page anchors can remain in the existing article store alongside this evidence.

Run php lib/followthemoney/examples/obituary.php from mono for the Emily/Lucy example. It writes FtM JSONL to stdout and evidence to stderr. Its source URI is a placeholder; no unverified death date or Wikidata identity is asserted.

Elasticsearch

Search\ElasticsearchProjection supplies a mapping, documents and streaming Bulk API NDJSON. It has no transport, network calls or service-container dependencies:

names and text support full-text search; schema, schemata, references and names.raw support filters. The original properties object uses Elasticsearch flattened: e.g. a term query on properties.deathDate finds 1953. Partial dates remain strings; range queries on this field are lexical, not chronological. Long property values remain in _source but values over 8191 characters are not keyword-indexed, avoiding Lucene term-size errors. Full-text fields still contain eligible text. Optional wire metadata is not copied into the strict search mapping.

The adapter is an independent projection, not OpenAleph's internal index format. Reindex from the authoritative entity store when mappings change. PostgreSQL, Doctrine, Messenger, permissions, review screens, poi-bundle lookups and MCP belong to consuming applications or a later Symfony bundle. No bundle is needed to use this library through normal Symfony service registration.

The integration test was run against Elasticsearch 9.5.3, including bulk indexing, full-text name search, exact partial-date queries, inherited-schema filters, relationship references and a long text value. Run it against your deployment with FTM_ELASTICSEARCH_URL=http://127.0.0.1:9200 vendor/bin/phpunit -c lib/followthemoney/phpunit.xml.dist (from mono). It creates and removes a uniquely named ftm-test-* index and is skipped when that environment variable is absent.

Tests and upstream refresh

From mono:

Standalone: composer install && composer test. Python is used only to regenerate schemas and expectations, never to execute PHPUnit or serve entity requests.

To reproduce the shipped resources, create an isolated Python environment, install tools/requirements.lock, clone upstream at the revision in resources/upstream.json, then run:

Run these from the library directory using that environment. PyICU may require ICU headers/pkg-config (on Homebrew set PKG_CONFIG_PATH to ICU's lib/pkgconfig). The Python package in the environment must match the checkout. Review schema, fixture and lockfile differences together on upgrades. Tests verify the snapshot hash and all reference ranges in addition to behavior. Fixture expectations come from Python; they are not produced by the PHP cleaner.

Attribution and references

Upstream model and portions of cleaning behavior: FollowTheMoney, MIT; license retained in resources/UPSTREAM-LICENSE. Prefix parsing: prefixdate, MIT; retained in resources/PREFIXDATE-LICENSE. No OpenAleph AGPL implementation was copied.

• FollowTheMoney schema explorer
• W3C Web Annotation model
• Elasticsearch flattened fields