Changelog — 6.0

Breaking changes

Package and namespace renamed

The package has been renamed from byjg/b8 to byjg/text-classifier and all namespaces have changed accordingly.

Before	After
B8\	ByJG\TextClassifier\
B8\Lexer\	ByJG\TextClassifier\Lexer\
B8\Degenerator\	ByJG\TextClassifier\Degenerator\
B8\Storage\	ByJG\TextClassifier\Storage\
B8\NaiveBayes\	ByJG\TextClassifier\NaiveBayes\
B8\NaiveBayes\Storage\	ByJG\TextClassifier\NaiveBayes\Storage\

Classes renamed

Before	After
B8	BinaryClassifier
ConfigB8	ConfigBinaryClassifier

classify() return type changed

Both classifiers now return a ClassificationResult object instead of a plain scalar or array. BinaryClassifier::classify() no longer returns a raw float; NaiveBayes::classify() no longer returns array<string, float>.

BinaryClassifier — before:

$score = $b8->classify($text);  // float|string
if (is_float($score) && $score > 0.8) { /* spam */ }

BinaryClassifier — after:

use ByJG\TextClassifier\ClassificationResult;

$result = $b8->classify($text);  // ClassificationResult|string
if ($result instanceof ClassificationResult && $result->score > 0.8) { /* spam */ }

NaiveBayes — before:

$scores = $nb->classify($text);  // array<string, float>
$top    = array_key_first($scores);

NaiveBayes — after:

$result = $nb->classify($text);  // ?ClassificationResult
echo $result->choice;   // winning category
echo $result->score;    // its score

NaiveBayes::classify() returns null instead of [] when no categories are trained yet.

PHP version requirement

Minimum PHP version raised from 8.1 to 8.3.

DBA storage handler changed

The Dba storage backend switched from db4 (BerkeleyDB) to gdbm. Existing db4 database files are not compatible — recreate them with createDatabase().

---

New features

ClassificationResult DTO

A new read-only value object returned by both classifiers:

Field	Type	Description
choice	string	Winning category ('spam', 'tech', …)
score	float	Final score of the winning category 0.0–1.0
scores	array<string, float>	All final scores, sorted descending
statScores	array<string, float>	Raw statistical scores before any LLM escalation
llmDecision	string|null	What the LLM chose, or null if not consulted
escalated	bool	true when the LLM was invoked

LLM-assisted classification with active learning

An LLM can now be injected directly into BinaryClassifier and NaiveBayes via their constructors. When the statistical model is uncertain, classify() escalates to the LLM internally. With autoLearn=true (default), the LLM decision is fed back as training data and the model is re-scored — no wrapper classes needed.

use ByJG\TextClassifier\Llm\ConfigLlm;
use ByJG\TextClassifier\Llm\OpenAiLlmClient;

$llm    = new OpenAiLlmClient($openai, 'gpt-4o-mini');
$config = (new ConfigLlm())
    ->setMinConfidence(0.65)
    ->setMinMargin(0.15)
    ->setAutoLearn(true);

$nb = new NaiveBayes($storage, $lexer, llm: $llm, configLlm: $config);

$result = $nb->classify($text);
// $result->escalated    — true if LLM was called
// $result->llmDecision  — what the LLM decided
// $result->statScores   — score before LLM involvement

New classes

Class	Description
ByJG\TextClassifier\ClassificationResult	Read-only result DTO returned by both classifiers
ByJG\TextClassifier\Llm\LlmInterface	Single-method interface: classify(string $text, array $categories): string
ByJG\TextClassifier\Llm\OpenAiLlmClient	OpenAI-compatible LLM client (works with Ollama too)
ByJG\TextClassifier\Llm\ConfigLlm	Fluent config for LLM escalation thresholds and autoLearn

NaiveBayes multi-class engine

ByJG\TextClassifier\NaiveBayes\NaiveBayes is a new multi-class Naive Bayes classifier that supports any number of user-defined categories, sharing the same lexer pipeline as BinaryClassifier. Backed by in-memory (Memory) or SQL (Rdbms) storage.

NaiveBayes::getCategories()

New public method returning the list of categories currently present in storage.

---

New dependencies

Package	Version
byjg/llm-api-objects	^6.0
openai-php/client	^0.19