Image Intelligence

Interlocute supports JPEG, PNG, WebP, TIFF, BMP, and GIF. Images can be submitted as raw bytes, a public URL, a blob SAS URI, or a provider file ID.

The structural fingerprint runs locally with zero API calls — structural properties, quality metrics, colour analysis, and metadata. Semantic intelligence adds understanding from a multimodal LLM — scene classification, entities, intent, emotion, and design scores. Forensic verification adds adversarial analysis — OCR evidence anchors, claim validation, manipulation detection, and risk assessment.

Yes. The structural fingerprint is fully deterministic and runs locally. It makes no external API calls and incurs no provider costs. It is always executed as the first step of any analysis.

When semantic intelligence produces a confidence score below the profile's escalation threshold (default 0.65), the pipeline automatically runs forensic verification. You can also force forensic verification by setting highStakes to true.

Forensic verification uses a second model pass that cross-examines claims made by the semantic analysis against OCR-extracted evidence. It detects contradictions, manipulation signals (edited images, AI-generated artifacts, lighting inconsistencies), and produces a contextual risk assessment with a tamper-proof evidence hash.

Yes. The Object Detection profile runs independently of the layered pipeline. It detects objects and people with bounding boxes, extracts content tags, and optionally removes the background — all without invoking semantic intelligence or forensic verification.

The structural fingerprint is free. Semantic intelligence incurs LLM token costs plus a platform premium. Forensic verification incurs additional LLM costs for the adversarial pass and vision API costs for OCR and feature extraction. All costs are attributed per-request in your usage ledger.