new system rebuild command

harden IngestService
This commit is contained in:
team 1
2026-02-27 12:03:45 +01:00
parent 7441273fad
commit 4761648836
2 changed files with 15 additions and 15 deletions

View File

@@ -13,7 +13,8 @@ final class KnowledgeIngestService
private DocumentLoader $loader, private DocumentLoader $loader,
private SimpleChunker $chunker, private SimpleChunker $chunker,
private DocumentVersionRepository $versionRepo, private DocumentVersionRepository $versionRepo,
) { )
{
} }
/** /**
@@ -39,9 +40,8 @@ final class KnowledgeIngestService
foreach ($chunks as $chunkText) { foreach ($chunks as $chunkText) {
// Titel optional weiterhin prefixen (wenn du das behalten willst)
if ($title !== '' && !str_starts_with($chunkText, $title)) { if ($title !== '' && !str_starts_with($chunkText, $title)) {
$chunkText = $title . "\n\n" . $chunkText; $chunkText = "# Produkt Titel: " . $title . "\n\n --- " . $chunkText;
} }
$chunkText = trim($chunkText); $chunkText = trim($chunkText);

View File

@@ -14,7 +14,7 @@ use App\Vector\VectorSearchClient;
final class NdjsonHybridRetriever implements RetrieverInterface final class NdjsonHybridRetriever implements RetrieverInterface
{ {
private const VECTOR_SCORE_THRESHOLD = 0.4; private const VECTOR_SCORE_THRESHOLD = 0.75;
private const HARD_MAX_CHUNKS = 200; private const HARD_MAX_CHUNKS = 200;
private const HARD_MAX_VECTORK = 200; private const HARD_MAX_VECTORK = 200;
@@ -23,7 +23,7 @@ final class NdjsonHybridRetriever implements RetrieverInterface
* Tags dürfen nur ein kleiner Bonus sein (kein Gate/Filter). * Tags dürfen nur ein kleiner Bonus sein (kein Gate/Filter).
* Enterprise Default: klein halten, sonst dominieren Tags wieder. * Enterprise Default: klein halten, sonst dominieren Tags wieder.
*/ */
private const TAG_SCORE_BONUS = 0.08; private const TAG_SCORE_BONUS = 0.25;
public function __construct( public function __construct(
private readonly NdjsonChunkLookup $lookup, private readonly NdjsonChunkLookup $lookup,