daveadmin
2d8d1c7409
Five MVP tools (Ask, Search, Summarize, Timeline, Redact) with email+password auth, Azure OpenAI gateway, evidence trail panel, and process-and-forget privacy default. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
632 lines
27 KiB
PHP
632 lines
27 KiB
PHP
<?php
|
|
declare(strict_types=1);
|
|
|
|
require_once __DIR__ . '/bootstrap.php';
|
|
require_once __DIR__ . '/AzureOpenAiGateway.php';
|
|
|
|
final class DbnLegalToolsService
|
|
{
|
|
private const MAX_PASTE_CHARS = 32000;
|
|
|
|
private DbnAzureOpenAiGateway $azure;
|
|
|
|
public function __construct(?DbnAzureOpenAiGateway $azure = null)
|
|
{
|
|
$this->azure = $azure ?: new DbnAzureOpenAiGateway();
|
|
}
|
|
|
|
public function search(string $query, string $language = 'en', int $limit = 6): array
|
|
{
|
|
$query = trim($query);
|
|
if (mb_strlen($query, 'UTF-8') < 3) {
|
|
dbnToolsAbort('Search query must be at least 3 characters.', 422, 'query_too_short');
|
|
}
|
|
$limit = max(1, min(10, $limit));
|
|
|
|
$trace = [
|
|
$this->trace('Query interpretation', 'Searching Dave Jr Legal private corpus plus the subscribed family-legal package.', 'complete'),
|
|
$this->trace('Search tools used', 'ClientRagPipeline::searchAll with keyword mode, private corpus enabled, shared package filter set to family-legal.', 'running'),
|
|
];
|
|
|
|
$client = dbnToolsRequireClient();
|
|
$package = $this->requireFamilyPackage((int)$client['id']);
|
|
|
|
$chunks = [];
|
|
$retrievalNote = 'ClientRagPipeline keyword retrieval';
|
|
try {
|
|
dbnToolsBootCaveau();
|
|
$gatewayUrl = 'http://10.0.1.10:4000';
|
|
try {
|
|
$config = getConfig();
|
|
$configured = trim((string)($config['ai_gateway']['url'] ?? ''));
|
|
if ($configured !== '') {
|
|
$gatewayUrl = $configured;
|
|
}
|
|
} catch (Throwable $e) {
|
|
// Retrieval still works in keyword mode without gateway config.
|
|
}
|
|
|
|
$rag = new ClientRagPipeline((int)$client['id'], $gatewayUrl, 30);
|
|
$chunks = $rag->searchAll($query, $limit, null, [
|
|
'search_private' => true,
|
|
'search_shared' => true,
|
|
'package_ids' => [(int)$package['id']],
|
|
'chunk_limit' => $limit,
|
|
'search_method' => 'keyword',
|
|
'min_private' => 0,
|
|
'include_beta_website' => true,
|
|
]);
|
|
} catch (Throwable $e) {
|
|
$retrievalNote = 'SQL keyword fallback after ClientRagPipeline error';
|
|
$trace[] = $this->trace('Search fallback', 'Pipeline retrieval failed; using direct SQL keyword fallback without storing the query.', 'warning');
|
|
$chunks = $this->fallbackKeywordSearch((int)$client['id'], $package, $query, $limit);
|
|
}
|
|
|
|
if (!$chunks) {
|
|
$fallback = $this->fallbackKeywordSearch((int)$client['id'], $package, $query, $limit);
|
|
if ($fallback) {
|
|
$chunks = $fallback;
|
|
$retrievalNote = 'SQL keyword fallback';
|
|
}
|
|
}
|
|
|
|
$hits = array_map(fn(array $chunk): array => $this->sourceFromChunk($chunk), array_slice($chunks, 0, $limit));
|
|
$confidence = $this->citationConfidence($hits);
|
|
|
|
$trace[1] = $this->trace('Search tools used', $retrievalNote . '; returned ' . count($hits) . ' source hit(s).', 'complete');
|
|
$trace[] = $this->trace('Evidence found', count($hits) ? 'Retrieved source excerpts for review.' : 'No matching source excerpts were found.', count($hits) ? 'complete' : 'warning');
|
|
$trace[] = $this->trace('Citation confidence', ucfirst($confidence) . ' confidence based on source count and retrieval scores.', $confidence === 'low' ? 'warning' : 'complete');
|
|
|
|
return [
|
|
'tool' => 'search',
|
|
'language' => $language,
|
|
'what_we_found' => count($hits) ? 'Found source excerpts from the legal corpus.' : 'No matching source excerpts were found.',
|
|
'hits' => $hits,
|
|
'evidence_trail' => $hits,
|
|
'what_remains_uncertain' => count($hits) ? 'Search results still need human review for legal relevance and currentness.' : 'The corpus may not contain enough evidence for this query.',
|
|
'next_practical_step' => count($hits) ? 'Open the strongest sources and confirm the cited sections before relying on them.' : 'Try a narrower query with statutory terms, party names, or dates.',
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => count($chunks),
|
|
'source_count' => count($hits),
|
|
'deployment' => null,
|
|
'citation_confidence' => $confidence,
|
|
],
|
|
'disclaimer' => dbnToolsDisclaimer($language),
|
|
];
|
|
}
|
|
|
|
public function ask(string $question, string $language = 'en'): array
|
|
{
|
|
$search = $this->search($question, $language, 7);
|
|
$hits = $search['hits'];
|
|
$trace = $search['trace'];
|
|
|
|
if (!$hits) {
|
|
$trace[] = $this->trace('Synthesis', 'Skipped answer synthesis because no evidence was found.', 'warning');
|
|
return [
|
|
'tool' => 'ask',
|
|
'language' => $language,
|
|
'answer' => $language === 'no'
|
|
? 'Jeg fant ikke nok kildestøtte i familie-rettskorpuset til å svare sikkert.'
|
|
: 'I did not find enough source support in the family-law corpus to answer safely.',
|
|
'what_we_found' => $search['what_we_found'],
|
|
'evidence_trail' => [],
|
|
'what_remains_uncertain' => $search['what_remains_uncertain'],
|
|
'next_practical_step' => $search['next_practical_step'],
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => 0,
|
|
'source_count' => 0,
|
|
'deployment' => null,
|
|
'citation_confidence' => 'low',
|
|
],
|
|
'disclaimer' => dbnToolsDisclaimer($language),
|
|
];
|
|
}
|
|
|
|
$this->azure->requireChat();
|
|
|
|
$context = $this->buildEvidenceContext($hits);
|
|
$locale = $language === 'no' ? 'Norwegian' : 'English';
|
|
$prompt = <<<PROMPT
|
|
Question:
|
|
{$question}
|
|
|
|
Evidence excerpts:
|
|
{$context}
|
|
|
|
Return JSON only with these keys:
|
|
{
|
|
"answer": "short direct answer in {$locale}",
|
|
"what_we_found": "plain-language summary of the supported finding",
|
|
"evidence_trail": [{"title":"source title","why_it_matters":"one sentence","citation":"visible source title or section"}],
|
|
"what_remains_uncertain": ["specific gaps or caveats"],
|
|
"next_practical_step": "one concrete next action"
|
|
}
|
|
PROMPT;
|
|
|
|
$system = $this->legalJsonSystemPrompt($language);
|
|
$raw = $this->azure->chatText([
|
|
['role' => 'system', 'content' => $system],
|
|
['role' => 'user', 'content' => $prompt],
|
|
], [
|
|
'json' => true,
|
|
'temperature' => 0.15,
|
|
'max_tokens' => 1300,
|
|
]);
|
|
|
|
$json = $this->azure->decodeJsonObject($raw);
|
|
if (!$json) {
|
|
$json = [
|
|
'answer' => $raw,
|
|
'what_we_found' => 'Azure returned a plain-text answer based on the retrieved excerpts.',
|
|
'evidence_trail' => [],
|
|
'what_remains_uncertain' => ['The response format could not be validated as structured JSON.'],
|
|
'next_practical_step' => 'Review the source excerpts manually before relying on the answer.',
|
|
];
|
|
}
|
|
|
|
$trace[] = $this->trace('Synthesis', 'Azure OpenAI generated an answer using only the retrieved source excerpts.', 'complete');
|
|
$trace[] = $this->trace('Uncertainty / missing evidence', $this->uncertaintySummary($json['what_remains_uncertain'] ?? []), 'complete');
|
|
$trace[] = $this->trace('Next practical step', (string)($json['next_practical_step'] ?? 'Review the evidence trail.'), 'complete');
|
|
|
|
return [
|
|
'tool' => 'ask',
|
|
'language' => $language,
|
|
'answer' => (string)($json['answer'] ?? ''),
|
|
'what_we_found' => (string)($json['what_we_found'] ?? ''),
|
|
'evidence_trail' => $hits,
|
|
'citation_notes' => $this->normalizeEvidenceTrail($json['evidence_trail'] ?? [], $hits),
|
|
'sources' => $hits,
|
|
'what_remains_uncertain' => $json['what_remains_uncertain'] ?? [],
|
|
'next_practical_step' => (string)($json['next_practical_step'] ?? ''),
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => count($hits),
|
|
'source_count' => count($hits),
|
|
'deployment' => $this->azure->chatDeployment(),
|
|
'citation_confidence' => $search['trace_metadata']['citation_confidence'] ?? 'medium',
|
|
],
|
|
'disclaimer' => dbnToolsDisclaimer($language),
|
|
];
|
|
}
|
|
|
|
public function summarize(string $text, string $language = 'en'): array
|
|
{
|
|
$text = $this->requirePasteText($text);
|
|
$this->azure->requireChat();
|
|
|
|
$locale = $language === 'no' ? 'Norwegian' : 'English';
|
|
$prompt = <<<PROMPT
|
|
Summarize this pasted case-preparation text in {$locale}. Do not invent missing facts.
|
|
|
|
Pasted text:
|
|
{$text}
|
|
|
|
Return JSON only:
|
|
{
|
|
"what_we_found": "plain-language summary",
|
|
"key_facts": ["fact"],
|
|
"dates": ["date or unknown"],
|
|
"parties": ["party or role"],
|
|
"legal_references_detected": ["reference"],
|
|
"what_remains_uncertain": ["uncertainty"],
|
|
"next_practical_step": "one concrete next action"
|
|
}
|
|
PROMPT;
|
|
|
|
$json = $this->runJsonTool($prompt, $language, 1300);
|
|
$trace = [
|
|
$this->trace('Query interpretation', 'Summarize pasted text without saving the text or output.', 'complete'),
|
|
$this->trace('Search tools used', 'No external corpus search; source is the user-pasted text.', 'complete'),
|
|
$this->trace('Evidence found', 'Evidence trail is limited to the pasted text supplied in this request.', 'complete'),
|
|
$this->trace('Citation confidence', 'Medium confidence for factual extraction; no external legal source verification was performed.', 'warning'),
|
|
$this->trace('Uncertainty / missing evidence', $this->uncertaintySummary($json['what_remains_uncertain'] ?? []), 'complete'),
|
|
$this->trace('Next practical step', (string)($json['next_practical_step'] ?? 'Review the summary against the original text.'), 'complete'),
|
|
];
|
|
|
|
return [
|
|
'tool' => 'summarize',
|
|
'language' => $language,
|
|
'what_we_found' => (string)($json['what_we_found'] ?? ''),
|
|
'key_facts' => $json['key_facts'] ?? [],
|
|
'dates' => $json['dates'] ?? [],
|
|
'parties' => $json['parties'] ?? [],
|
|
'legal_references_detected' => $json['legal_references_detected'] ?? [],
|
|
'evidence_trail' => [['title' => 'Pasted text', 'excerpt' => 'Processed in-memory only; not stored.']],
|
|
'what_remains_uncertain' => $json['what_remains_uncertain'] ?? [],
|
|
'next_practical_step' => (string)($json['next_practical_step'] ?? ''),
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => 1,
|
|
'source_count' => 1,
|
|
'deployment' => $this->azure->chatDeployment(),
|
|
],
|
|
'disclaimer' => dbnToolsDisclaimer($language),
|
|
];
|
|
}
|
|
|
|
public function timeline(string $text, string $language = 'en'): array
|
|
{
|
|
$text = $this->requirePasteText($text);
|
|
$this->azure->requireChat();
|
|
|
|
$locale = $language === 'no' ? 'Norwegian' : 'English';
|
|
$prompt = <<<PROMPT
|
|
Build a chronological timeline from this pasted text in {$locale}. Keep uncertain dates explicit.
|
|
|
|
Pasted text:
|
|
{$text}
|
|
|
|
Return JSON only:
|
|
{
|
|
"what_we_found": "short overview",
|
|
"events": [{"date":"YYYY-MM-DD, month/year, or unknown","actor":"actor or unknown","event":"event","source_excerpt":"short excerpt","confidence":"high|medium|low"}],
|
|
"evidence_trail": [{"title":"Pasted text","excerpt":"short relevant excerpt"}],
|
|
"what_remains_uncertain": ["uncertainty"],
|
|
"next_practical_step": "one concrete next action"
|
|
}
|
|
PROMPT;
|
|
|
|
$json = $this->runJsonTool($prompt, $language, 1600);
|
|
$events = is_array($json['events'] ?? null) ? $json['events'] : [];
|
|
$trace = [
|
|
$this->trace('Query interpretation', 'Extract dated events from pasted text without saving the text or output.', 'complete'),
|
|
$this->trace('Search tools used', 'No external corpus search; source is the user-pasted text.', 'complete'),
|
|
$this->trace('Evidence found', count($events) . ' event(s) identified.', count($events) ? 'complete' : 'warning'),
|
|
$this->trace('Citation confidence', 'Confidence is per event and based only on the pasted text.', 'complete'),
|
|
$this->trace('Uncertainty / missing evidence', $this->uncertaintySummary($json['what_remains_uncertain'] ?? []), 'complete'),
|
|
$this->trace('Next practical step', (string)($json['next_practical_step'] ?? 'Verify dates against original documents.'), 'complete'),
|
|
];
|
|
|
|
return [
|
|
'tool' => 'timeline',
|
|
'language' => $language,
|
|
'what_we_found' => (string)($json['what_we_found'] ?? ''),
|
|
'events' => $events,
|
|
'evidence_trail' => $json['evidence_trail'] ?? [['title' => 'Pasted text', 'excerpt' => 'Processed in-memory only; not stored.']],
|
|
'what_remains_uncertain' => $json['what_remains_uncertain'] ?? [],
|
|
'next_practical_step' => (string)($json['next_practical_step'] ?? ''),
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => count($events),
|
|
'source_count' => 1,
|
|
'deployment' => $this->azure->chatDeployment(),
|
|
],
|
|
'disclaimer' => dbnToolsDisclaimer($language),
|
|
];
|
|
}
|
|
|
|
public function redact(string $text, string $mode = 'standard'): array
|
|
{
|
|
$text = $this->requirePasteText($text);
|
|
$mode = $mode === 'strict' ? 'strict' : 'standard';
|
|
[$redacted, $entities] = $this->deterministicRedaction($text, $mode);
|
|
|
|
$categories = array_keys(array_filter($entities, fn(int $count): bool => $count > 0));
|
|
$trace = [
|
|
$this->trace('Query interpretation', 'Detect and redact sensitive identifiers from pasted text.', 'complete'),
|
|
$this->trace('Search tools used', 'Deterministic Norwegian privacy patterns first; no text was stored.', 'complete'),
|
|
$this->trace('Evidence found', count($categories) ? 'Detected categories: ' . implode(', ', $categories) . '.' : 'No deterministic sensitive categories were detected.', count($categories) ? 'complete' : 'warning'),
|
|
$this->trace('Citation confidence', 'High for emails and fødselsnummer-like values; medium for addresses and names.', 'complete'),
|
|
$this->trace('Uncertainty / missing evidence', 'Contextual names may need human review, especially in standard mode.', 'warning'),
|
|
$this->trace('Next practical step', 'Review the redacted output before sharing it outside the case team.', 'complete'),
|
|
];
|
|
|
|
return [
|
|
'tool' => 'redact',
|
|
'mode' => $mode,
|
|
'what_we_found' => 'Redacted deterministic privacy patterns from the pasted text.',
|
|
'redacted_text' => $redacted,
|
|
'detected_entity_categories' => $categories,
|
|
'entity_counts' => $entities,
|
|
'evidence_trail' => [['title' => 'Pasted text', 'excerpt' => 'Processed in-memory only; not stored.']],
|
|
'what_remains_uncertain' => ['Human review is still needed for names that depend on case context.'],
|
|
'next_practical_step' => 'Review the output and rerun in strict mode if the text will be shared broadly.',
|
|
'trace' => $trace,
|
|
'trace_metadata' => [
|
|
'chunk_count' => 1,
|
|
'source_count' => 1,
|
|
'deployment' => null,
|
|
],
|
|
'disclaimer' => 'Privacy support tool. Review before disclosure.',
|
|
];
|
|
}
|
|
|
|
private function requireFamilyPackage(int $clientId): array
|
|
{
|
|
$package = dbnToolsFetchPackage('family-legal');
|
|
if (!$package || empty($package['is_active'])) {
|
|
dbnToolsAbort('The family-legal corpus package is not active.', 503, 'package_unavailable');
|
|
}
|
|
if (!dbnToolsHasActiveSubscription($clientId, (int)$package['id'])) {
|
|
dbnToolsAbort('Dave Jr Legal does not have an active family-legal subscription.', 503, 'subscription_missing');
|
|
}
|
|
return $package;
|
|
}
|
|
|
|
private function runJsonTool(string $prompt, string $language, int $maxTokens): array
|
|
{
|
|
$raw = $this->azure->chatText([
|
|
['role' => 'system', 'content' => $this->legalJsonSystemPrompt($language)],
|
|
['role' => 'user', 'content' => $prompt],
|
|
], [
|
|
'json' => true,
|
|
'temperature' => 0.1,
|
|
'max_tokens' => $maxTokens,
|
|
]);
|
|
$json = $this->azure->decodeJsonObject($raw);
|
|
if (!$json) {
|
|
dbnToolsAbort('Azure OpenAI did not return valid structured JSON.', 502, 'azure_invalid_json');
|
|
}
|
|
return $json;
|
|
}
|
|
|
|
private function legalJsonSystemPrompt(string $language): string
|
|
{
|
|
$locale = $language === 'no' ? 'Norwegian' : 'English';
|
|
return <<<PROMPT
|
|
You are Do Better Norge Legal Tools in a source-grounded legal preparation workflow.
|
|
Use the DBN legal guardrails:
|
|
- Answer only from provided source excerpts or pasted text.
|
|
- Treat your role as legal information and issue-spotting, not final legal advice.
|
|
- Never invent statutes, paragraph numbers, case names, citations, parties, dates, or sources.
|
|
- If evidence is insufficient, say so plainly.
|
|
- Respond in {$locale}.
|
|
- Return valid JSON only. No markdown fences.
|
|
PROMPT;
|
|
}
|
|
|
|
private function buildEvidenceContext(array $hits): string
|
|
{
|
|
$lines = [];
|
|
foreach ($hits as $idx => $hit) {
|
|
$n = $idx + 1;
|
|
$lines[] = "[{$n}] Title: " . ($hit['title'] ?? 'Untitled');
|
|
if (!empty($hit['section'])) {
|
|
$lines[] = "Section: " . $hit['section'];
|
|
}
|
|
$lines[] = "Corpus/package: " . ($hit['package_or_corpus'] ?? 'unknown');
|
|
$lines[] = "Excerpt: " . ($hit['excerpt'] ?? '');
|
|
}
|
|
return implode("\n", $lines);
|
|
}
|
|
|
|
private function normalizeEvidenceTrail(mixed $trail, array $hits): array
|
|
{
|
|
if (!is_array($trail) || !$trail) {
|
|
return array_map(fn(array $hit): array => [
|
|
'title' => $hit['title'],
|
|
'citation' => $hit['title'],
|
|
'why_it_matters' => dbnToolsExcerpt($hit['excerpt'], 180),
|
|
], array_slice($hits, 0, 4));
|
|
}
|
|
return array_values(array_filter($trail, 'is_array'));
|
|
}
|
|
|
|
private function sourceFromChunk(array $chunk): array
|
|
{
|
|
$title = (string)($chunk['document_title'] ?? $chunk['title'] ?? 'Untitled source');
|
|
$score = isset($chunk['similarity']) ? round((float)$chunk['similarity'], 4) : null;
|
|
return [
|
|
'title' => $title,
|
|
'excerpt' => dbnToolsExcerpt((string)($chunk['content'] ?? ''), 620),
|
|
'package_or_corpus' => (string)($chunk['source_name'] ?? $chunk['source_type'] ?? 'Dave Jr Legal'),
|
|
'score' => $score,
|
|
'document_id' => isset($chunk['document_id']) ? (int)$chunk['document_id'] : null,
|
|
'chunk_id' => isset($chunk['id']) ? (int)$chunk['id'] : null,
|
|
'section' => $chunk['section_title'] ?? null,
|
|
'authority_type' => $chunk['authority_type'] ?? null,
|
|
'jurisdiction' => $chunk['jurisdiction'] ?? null,
|
|
];
|
|
}
|
|
|
|
private function citationConfidence(array $hits): string
|
|
{
|
|
if (!$hits) {
|
|
return 'low';
|
|
}
|
|
$scores = array_values(array_filter(array_map(fn(array $h) => $h['score'] ?? null, $hits), 'is_numeric'));
|
|
$best = $scores ? max($scores) : 0;
|
|
if (count($hits) >= 3 && $best >= 0.35) {
|
|
return 'high';
|
|
}
|
|
if (count($hits) >= 1) {
|
|
return 'medium';
|
|
}
|
|
return 'low';
|
|
}
|
|
|
|
private function fallbackKeywordSearch(int $clientId, array $package, string $query, int $limit): array
|
|
{
|
|
$results = [];
|
|
try {
|
|
$results = array_merge($results, $this->fallbackPrivateSearch($clientId, $query, $limit));
|
|
} catch (Throwable $e) {
|
|
error_log('DBN tools private fallback failed: ' . $e->getMessage());
|
|
}
|
|
try {
|
|
$remaining = max(1, $limit - count($results));
|
|
$results = array_merge($results, $this->fallbackSharedSearch($package, $query, $remaining));
|
|
} catch (Throwable $e) {
|
|
error_log('DBN tools shared fallback failed: ' . $e->getMessage());
|
|
}
|
|
return array_slice($results, 0, $limit);
|
|
}
|
|
|
|
private function fallbackPrivateSearch(int $clientId, string $query, int $limit): array
|
|
{
|
|
$db = dbnToolsDb();
|
|
$terms = $this->searchTerms($query);
|
|
if (!$terms) {
|
|
return [];
|
|
}
|
|
$clauses = [];
|
|
$params = [':client_id' => $clientId];
|
|
foreach ($terms as $i => $term) {
|
|
$key = ':term' . $i;
|
|
$clauses[] = "(cc.content LIKE {$key} OR cd.title LIKE {$key})";
|
|
$params[$key] = '%' . $term . '%';
|
|
}
|
|
$sql = 'SELECT cc.id, cc.document_id, cc.content, cd.title AS document_title, cd.category
|
|
FROM client_chunks cc
|
|
JOIN client_documents cd ON cc.document_id = cd.id
|
|
WHERE cc.client_id = :client_id AND cd.status = "ready" AND (' . implode(' OR ', $clauses) . ')
|
|
LIMIT ' . (int)$limit;
|
|
$stmt = $db->prepare($sql);
|
|
$stmt->execute($params);
|
|
$rows = $stmt->fetchAll(PDO::FETCH_ASSOC);
|
|
foreach ($rows as &$row) {
|
|
$row['similarity'] = 0.25;
|
|
$row['source_name'] = 'Dave Jr Legal private corpus';
|
|
$row['source_type'] = 'private';
|
|
}
|
|
return $rows;
|
|
}
|
|
|
|
private function fallbackSharedSearch(array $package, string $query, int $limit): array
|
|
{
|
|
$ragDb = dbnToolsRagDb();
|
|
$terms = $this->searchTerms($query);
|
|
if (!$terms) {
|
|
return [];
|
|
}
|
|
|
|
$where = ['d.status = "ready"'];
|
|
$params = [];
|
|
|
|
if (!empty($package['corpus_id'])) {
|
|
$where[] = 'd.corpus_id = ?';
|
|
$params[] = (int)$package['corpus_id'];
|
|
}
|
|
|
|
$cats = json_decode((string)($package['category_filter'] ?? '[]'), true) ?: [];
|
|
if ($cats) {
|
|
$where[] = 'd.category IN (' . implode(',', array_fill(0, count($cats), '?')) . ')';
|
|
$params = array_merge($params, $cats);
|
|
}
|
|
|
|
$langs = json_decode((string)($package['language_filter'] ?? '[]'), true) ?: [];
|
|
if ($langs) {
|
|
$where[] = 'd.language IN (' . implode(',', array_fill(0, count($langs), '?')) . ')';
|
|
$params = array_merge($params, $langs);
|
|
}
|
|
|
|
$termClauses = [];
|
|
foreach ($terms as $term) {
|
|
$termClauses[] = '(c.content LIKE ? OR d.title LIKE ?)';
|
|
$params[] = '%' . $term . '%';
|
|
$params[] = '%' . $term . '%';
|
|
}
|
|
$where[] = '(' . implode(' OR ', $termClauses) . ')';
|
|
|
|
$sql = 'SELECT c.id, c.document_id, c.content, c.section_title, d.title AS document_title,
|
|
d.category, d.language
|
|
FROM chunks c
|
|
JOIN documents d ON c.document_id = d.id
|
|
WHERE ' . implode(' AND ', $where) . '
|
|
LIMIT ' . (int)$limit;
|
|
$stmt = $ragDb->prepare($sql);
|
|
$stmt->execute($params);
|
|
$rows = $stmt->fetchAll(PDO::FETCH_ASSOC);
|
|
foreach ($rows as &$row) {
|
|
$row['similarity'] = 0.2;
|
|
$row['source_name'] = (string)($package['name'] ?? 'family-legal');
|
|
$row['source_type'] = 'package';
|
|
}
|
|
return $rows;
|
|
}
|
|
|
|
private function searchTerms(string $query): array
|
|
{
|
|
$parts = preg_split('/[^\p{L}\p{N}]+/u', mb_strtolower($query, 'UTF-8')) ?: [];
|
|
$stop = ['the', 'and', 'for', 'with', 'that', 'this', 'hva', 'har', 'kan', 'jeg', 'som', 'det', 'med', 'til', 'og'];
|
|
$terms = [];
|
|
foreach ($parts as $part) {
|
|
if (mb_strlen($part, 'UTF-8') < 3 || in_array($part, $stop, true)) {
|
|
continue;
|
|
}
|
|
$terms[] = $part;
|
|
}
|
|
return array_slice(array_values(array_unique($terms)), 0, 6);
|
|
}
|
|
|
|
private function requirePasteText(string $text): string
|
|
{
|
|
$text = trim($text);
|
|
if (mb_strlen($text, 'UTF-8') < 20) {
|
|
dbnToolsAbort('Paste at least 20 characters of text.', 422, 'text_too_short');
|
|
}
|
|
if (mb_strlen($text, 'UTF-8') > self::MAX_PASTE_CHARS) {
|
|
dbnToolsAbort('Pasted text is too long for the MVP limit.', 422, 'text_too_long');
|
|
}
|
|
return $text;
|
|
}
|
|
|
|
private function deterministicRedaction(string $text, string $mode): array
|
|
{
|
|
$counts = [
|
|
'email' => 0,
|
|
'phone' => 0,
|
|
'fødselsnummer' => 0,
|
|
'address' => 0,
|
|
'person_or_child_name' => 0,
|
|
];
|
|
|
|
$replace = function (string $pattern, string $category, string $token) use (&$text, &$counts): void {
|
|
$text = preg_replace_callback($pattern, function () use (&$counts, $category, $token): string {
|
|
$counts[$category]++;
|
|
return $token;
|
|
}, $text) ?? $text;
|
|
};
|
|
|
|
$replace('/\b[A-Z0-9._%+\-]+@[A-Z0-9.\-]+\.[A-Z]{2,}\b/i', 'email', '[EMAIL]');
|
|
$replace('/(?<!\d)(?:\d{6}[\s\-]?\d{5}|\d{11})(?!\d)/u', 'fødselsnummer', '[FNR]');
|
|
$replace('/(?<!\d)(?:\+47[\s.\-]?)?(?:\d[\s.\-]?){8}(?!\d)/u', 'phone', '[PHONE]');
|
|
$replace('/\b[A-ZÆØÅ][\p{L}æøåÆØÅ\.\- ]{2,40}\s+(?:gate|gata|vei|veien|plass|street|road|avenue|ave)\s+\d+[A-Z]?\b/iu', 'address', '[ADDRESS]');
|
|
|
|
$text = preg_replace_callback(
|
|
'/\b(Barn|Child|Navn|Name|Mor|Far|Mother|Father|Sønn|Datter)\s*:\s*([^\r\n,.;]+)/iu',
|
|
function (array $m) use (&$counts): string {
|
|
$counts['person_or_child_name']++;
|
|
return $m[1] . ': [PERSON]';
|
|
},
|
|
$text
|
|
) ?? $text;
|
|
|
|
$text = preg_replace_callback(
|
|
'/\b(?:barnet|child|sønn|son|datter|daughter)\s+(?:heter|named|called)?\s*([A-ZÆØÅ][\p{L}æøåÆØÅ\-]{2,})\b/iu',
|
|
function () use (&$counts): string {
|
|
$counts['person_or_child_name']++;
|
|
return '[CHILD_IDENTIFIER]';
|
|
},
|
|
$text
|
|
) ?? $text;
|
|
|
|
if ($mode === 'strict') {
|
|
$replace('/\b[A-ZÆØÅ][\p{L}æøåÆØÅ\-]{2,}\s+[A-ZÆØÅ][\p{L}æøåÆØÅ\-]{2,}\b/u', 'person_or_child_name', '[PERSON]');
|
|
}
|
|
|
|
return [$text, $counts];
|
|
}
|
|
|
|
private function uncertaintySummary(mixed $uncertainty): string
|
|
{
|
|
if (is_array($uncertainty)) {
|
|
$uncertainty = implode(' ', array_map('strval', $uncertainty));
|
|
}
|
|
$uncertainty = trim((string)$uncertainty);
|
|
return $uncertainty !== '' ? dbnToolsExcerpt($uncertainty, 220) : 'No additional uncertainty was supplied by the tool.';
|
|
}
|
|
|
|
private function trace(string $label, string $detail, string $status = 'complete'): array
|
|
{
|
|
return [
|
|
'label' => $label,
|
|
'detail' => $detail,
|
|
'status' => $status,
|
|
];
|
|
}
|
|
}
|