🎥 Media & Biometrics

Relata DB 3.0
In 3.0 media processing is outside the core (ETL-first boundary, ADR-0309). The SQL operators FACE_SEARCH, VOICE_MATCH, MATCH_PHOTODNA, MATCH_PDQ and VIDEO_FINGERPRINT_MATCH are removed and fail at parse time, as are the MCP tools face_match and search_video_frames, the face and PDQ matching SDK methods, and POST /ingest/media (answers 410 Gone, problem code media-ingest-removed). Decoding, OCR, transcription, perceptual hashing and embedding run in OpenETL, which loads vectors and MediaSegment rows through bulk ingest; Relata keeps generic similarity retrieval (SIMILAR_IMAGE, SIMILAR TO, HYBRID_SEARCH). The 2.5.8 line, kept in the 2.5.8 documentation snapshot, is described by the older material on this page. See the repository guides media-processing-migration and media-matching-migration.

Four governed SQL operators for face search, image similarity, perceptual-hash matching, and DNS-tunnel detection. (The MCP face_match tool is gated behind ADR-155 — these SQL-reachable operators are the unblocked path.)

client.face_search(
    "gallery-1", [0.12, -0.34, 0.56, ...],   # 128-dim probe embedding
    k=5, threshold=0.6, purpose="investigation",
)
# QueryResult: columns entity_id, gallery_id, score, modality (ranked desc)
 
client.match_pdq(
    "corpus-1", "a1b2c3d4e5f67890...",        # PDQ hex hash
    threshold=0.9, purpose="investigation",
)
# QueryResult: entity_id, media_type, corpus_id, score, matcher
 
client.similar_image(
    "media-42", threshold=0.6, index="ncmec",
    purpose="investigation",
)
# QueryResult: near-duplicate images, ranked desc
 
client.dns_tunnel_detect("ws-finance-01", purpose="security")
# {"data": [{"domain": "exfil.bad.tld", "entropy": 7.8, "count": 1242}]}

Next: Governance & Compliance — GDPR Art. 17 erasure, session commit/review/discard, and data export.