optimize commerce keyword filterlist

This commit is contained in:
team 1
2026-05-03 19:25:21 +02:00
parent 98447703ac
commit 2520ca06af
4 changed files with 190 additions and 1 deletions

View File

@@ -64,6 +64,9 @@ parameters:
- würde - würde
- ich - ich
- gerne - gerne
- welchem
- kann
- mit
- mein - mein
- größer - größer
- zeige - zeige
@@ -121,6 +124,7 @@ parameters:
- kostet - kostet
- kosten - kosten
- ua - ua
- ein
- also - also
- gut - gut
- gute - gute
@@ -129,6 +133,7 @@ parameters:
- gutes - gutes
- passen - passen
- passend - passend
- was
search_control_tokens: search_control_tokens:
- shop - shop

View File

@@ -220,6 +220,42 @@ parameters:
- '- This block is generated from the current user question and is stricter than broad product-selection wording.' - '- This block is generated from the current user question and is stricter than broad product-selection wording.'
- '- For measurement-parameter questions, technical suitability requires explicit positive evidence for the requested parameter in the same source record.' - '- For measurement-parameter questions, technical suitability requires explicit positive evidence for the requested parameter in the same source record.'
- '- Similar water-treatment parameters, abbreviations, units, product families, search queries, or ranking positions are not enough.' - '- Similar water-treatment parameters, abbreviations, units, product families, search queries, or ranking positions are not enough.'
product_specific_rules:
- '- Verify every recommended product independently against the requested measurement parameter.'
- '- If a retrieved RAG record mentions several products, only use a product for the requested parameter when that product is named in the same sentence, bullet, table row, or clearly bounded product section as the parameter evidence.'
- '- Do not transfer measurement suitability from one product, variant, indicator, category, tag, heading, or nearby paragraph to another product.'
- '- Generic category, umbrella-topic, device-class, product-family, document-title, tag, or application-area terms are not enough to prove a specific measurement parameter for a specific product.'
generic_request_patterns:
- '/\b(?:mit|für|fuer|zur|zum)\s+(?:dem\s+)?(?:messparameter|parameter|messwert|messgröße|messgroesse)\s+(?<parameter>[^?.!,;\n]{2,80})/iu'
- '/\b(?:messparameter|parameter|messwert|messgröße|messgroesse)\s*(?:für|fuer|von|zur|zum|:)\s*(?<parameter>[^?.!,;\n]{2,80})/iu'
generic_positive_context_terms:
- Messung
- messen
- misst
- Messbereich
- Messparameter
- Messgröße
- Messgroesse
- Bestimmung
- bestimmen
- Analyse
- analysiert
- überwachen
- ueberwachen
- Indikator für
- Indikator fuer
- Reagenz für
- Reagenz fuer
- Sensor
- Elektrode
generic_negative_context_terms:
- Betriebsbereich
- Betriebsumgebung
- Einsatzbedingungen
- störungsfrei
- stoerungsfrei
generic_safe_no_evidence_answer_template_de: Ich finde in den bereitgestellten Quellen keinen sicher belegten Treffer für die Messung von {label}.
generic_safe_no_accessory_evidence_answer_template_de: Ich finde in den bereitgestellten Quellen keinen sicher belegten Indikator oder ein Reagenz für die Messung von {label}.
parameters: parameters:
- id: ph - id: ph
label: pH / pH-Wert label: pH / pH-Wert

View File

@@ -610,6 +610,7 @@ final readonly class PromptBuilder
$negativeContextTerms = $this->extractMeasurementGuardStringList($guard, 'negative_context_terms'); $negativeContextTerms = $this->extractMeasurementGuardStringList($guard, 'negative_context_terms');
$nonEquivalentTerms = $this->extractMeasurementGuardStringList($guard, 'non_equivalent_terms'); $nonEquivalentTerms = $this->extractMeasurementGuardStringList($guard, 'non_equivalent_terms');
$label = $this->normalizeBlockText((string) ($guard['label'] ?? 'requested measurement parameter')); $label = $this->normalizeBlockText((string) ($guard['label'] ?? 'requested measurement parameter'));
$strictNoEvidence = (bool) ($guard['strict_no_evidence'] ?? true);
$resolvedRequestedRole = $requestedRole ?? $this->resolveRequestedProductRole($prompt); $resolvedRequestedRole = $requestedRole ?? $this->resolveRequestedProductRole($prompt);
$safeNoEvidenceAnswer = $this->normalizeBlockText((string) ( $safeNoEvidenceAnswer = $this->normalizeBlockText((string) (
$resolvedRequestedRole === 'accessory_or_consumable' $resolvedRequestedRole === 'accessory_or_consumable'
@@ -650,6 +651,7 @@ final readonly class PromptBuilder
} }
$rules = $this->config->getMeasurementEvidenceIntroRules(); $rules = $this->config->getMeasurementEvidenceIntroRules();
$rules = array_merge($rules, $this->config->getMeasurementEvidenceProductSpecificRules());
$rules[] = '- User requested measurement parameter: ' . $label . '.'; $rules[] = '- User requested measurement parameter: ' . $label . '.';
$rules[] = '- Positive parameter terms for this request: ' . implode(', ', $positiveTerms) . '.'; $rules[] = '- Positive parameter terms for this request: ' . implode(', ', $positiveTerms) . '.';
if ($positiveContextTerms !== []) { if ($positiveContextTerms !== []) {
@@ -666,7 +668,11 @@ final readonly class PromptBuilder
$rules[] = '- RAG/URL evidence scan for this exact parameter: ' . ($knowledgeHasEvidence ? 'explicit positive evidence found.' : 'no explicit positive evidence found.'); $rules[] = '- RAG/URL evidence scan for this exact parameter: ' . ($knowledgeHasEvidence ? 'explicit positive evidence found.' : 'no explicit positive evidence found.');
$rules = array_merge($rules, $shopEvidenceLines); $rules = array_merge($rules, $shopEvidenceLines);
if (!$knowledgeHasEvidence && !$shopHasEvidence) { if (!$strictNoEvidence && !$knowledgeHasEvidence && !$shopHasEvidence) {
$rules[] = '- The deterministic exact-term scan did not find product-specific evidence. The answer may still use a clearly equivalent named measurement parameter from the same source record, but must not infer suitability from generic categories, document titles, tags, search terms, neighbouring products, or broad umbrella-topic wording.';
}
if ($strictNoEvidence && !$knowledgeHasEvidence && !$shopHasEvidence) {
$rules[] = '- Mandatory answer behavior: do not recommend a product as suitable for this measurement parameter.'; $rules[] = '- Mandatory answer behavior: do not recommend a product as suitable for this measurement parameter.';
if ($safeNoEvidenceAnswer !== '') { if ($safeNoEvidenceAnswer !== '') {
$rules[] = '- Start the answer with this meaning in the user language: ' . $safeNoEvidenceAnswer; $rules[] = '- Start the answer with this meaning in the user language: ' . $safeNoEvidenceAnswer;
@@ -724,14 +730,114 @@ final readonly class PromptBuilder
foreach ($requestTerms as $term) { foreach ($requestTerms as $term) {
if ($this->containsMeasurementTerm($normalizedPrompt, $term)) { if ($this->containsMeasurementTerm($normalizedPrompt, $term)) {
$parameter['strict_no_evidence'] = true;
return $parameter; return $parameter;
} }
} }
} }
return $this->resolveGenericRequestedMeasurementGuard($prompt);
}
private function resolveGenericRequestedMeasurementGuard(string $prompt): ?array
{
foreach ($this->config->getMeasurementEvidenceGenericRequestPatterns() as $pattern) {
if (@preg_match($pattern, $prompt, $matches) !== 1) {
continue;
}
$rawParameter = $matches['parameter'] ?? ($matches[1] ?? '');
if (!is_scalar($rawParameter)) {
continue;
}
$parameterTerms = $this->buildGenericMeasurementParameterTerms((string) $rawParameter);
if ($parameterTerms === []) {
continue;
}
$label = implode(' / ', $parameterTerms);
return [
'id' => 'generic_' . substr(sha1($label), 0, 12),
'label' => $label,
'request_terms' => $parameterTerms,
'positive_terms' => $parameterTerms,
'positive_context_terms' => $this->config->getMeasurementEvidenceGenericPositiveContextTerms(),
'negative_context_terms' => $this->config->getMeasurementEvidenceGenericNegativeContextTerms(),
'non_equivalent_terms' => [],
'safe_no_evidence_answer_de' => $this->renderMeasurementEvidenceTemplate(
$this->config->getMeasurementEvidenceGenericSafeNoEvidenceAnswerTemplate(),
$label
),
'safe_no_accessory_evidence_answer_de' => $this->renderMeasurementEvidenceTemplate(
$this->config->getMeasurementEvidenceGenericSafeNoAccessoryEvidenceAnswerTemplate(),
$label
),
'strict_no_evidence' => false,
];
}
return null; return null;
} }
/**
* @return string[]
*/
private function buildGenericMeasurementParameterTerms(string $rawParameter): array
{
$rawParameter = $this->normalizeBlockText($rawParameter);
if ($rawParameter === '') {
return [];
}
$terms = [];
$withoutParentheses = preg_replace('/\([^)]*\)/u', ' ', $rawParameter) ?? $rawParameter;
$this->appendGenericMeasurementParameterParts($terms, $withoutParentheses);
if (preg_match_all('/\(([^)]{1,40})\)/u', $rawParameter, $matches) === 1) {
foreach ($matches[1] as $parenthetical) {
$this->appendGenericMeasurementParameterParts($terms, (string) $parenthetical);
}
}
return array_values(array_unique($terms));
}
/**
* @param string[] $terms
*/
private function appendGenericMeasurementParameterParts(array &$terms, string $value): void
{
$value = $this->normalizeBlockText($value);
if ($value === '') {
return;
}
$parts = preg_split('/\s*(?:,|;|\/|\boder\b|\bund\b|\bor\b|\band\b)\s*/iu', $value) ?: [$value];
foreach ($parts as $part) {
$part = $this->normalizeBlockText((string) $part);
$part = trim($part, " \t\n\r\0\x0B-–—:()[]{}\"'`“”„");
if ($part === '' || preg_match('/[\p{L}\p{N}]/u', $part) !== 1) {
continue;
}
if (mb_strlen($part, 'UTF-8') < 2 || in_array($part, $terms, true)) {
continue;
}
$terms[] = $part;
}
}
private function renderMeasurementEvidenceTemplate(string $template, string $label): string
{
return strtr($template, ['{label}' => $label]);
}
/** /**
* @return string[] * @return string[]
*/ */

View File

@@ -581,6 +581,48 @@ final class PromptBuilderConfig
return $this->getRequiredStringList('measurement_evidence_guard.intro_rules'); return $this->getRequiredStringList('measurement_evidence_guard.intro_rules');
} }
/**
* @return string[]
*/
public function getMeasurementEvidenceProductSpecificRules(): array
{
return $this->getRequiredStringList('measurement_evidence_guard.product_specific_rules');
}
/**
* @return string[]
*/
public function getMeasurementEvidenceGenericRequestPatterns(): array
{
return $this->getRequiredStringList('measurement_evidence_guard.generic_request_patterns');
}
/**
* @return string[]
*/
public function getMeasurementEvidenceGenericPositiveContextTerms(): array
{
return $this->getRequiredStringList('measurement_evidence_guard.generic_positive_context_terms');
}
/**
* @return string[]
*/
public function getMeasurementEvidenceGenericNegativeContextTerms(): array
{
return $this->getRequiredStringList('measurement_evidence_guard.generic_negative_context_terms');
}
public function getMeasurementEvidenceGenericSafeNoEvidenceAnswerTemplate(): string
{
return $this->getRequiredString('measurement_evidence_guard.generic_safe_no_evidence_answer_template_de');
}
public function getMeasurementEvidenceGenericSafeNoAccessoryEvidenceAnswerTemplate(): string
{
return $this->getRequiredString('measurement_evidence_guard.generic_safe_no_accessory_evidence_answer_template_de');
}
/** /**
* @return array<int, array<string, mixed>> * @return array<int, array<string, mixed>>
*/ */