Skip to main content
AI labels are stored separately from manually assigned tags and Spatie Tags. Image analysis is disabled by default and never runs while browsing or searching. Choose Moondream, Google Cloud Vision or Amazon Rekognition for image labels. Video analysis uses a limited set of sampled frames through the selected image provider. The AI page includes a paginated review panel for accepting or rejecting independent AI suggestions. OCR and lexical/hybrid/semantic retrieval are configured separately in Search. All processing is optional and disabled by default; browsing never starts a paid analysis. Set provider quotas and application admission limits before enabling automatic processing.

Install optional tables

These tables are not part of default Studio migrations. They contain version-bound image analyses and independent AI labels; no user tags are created or modified.

Choose one provider

Set an explicit supported Moondream model; Studio does not silently change it. Rekognition requires the optional official aws/aws-sdk-php package and uses the standard AWS credential chain. The independent Studio package does not currently declare/install that optional SDK. Missing SDK/configuration fails closed. No provider credentials reach the browser or analysis tables.

Prepare an analysis derivative

Use intelligence.input to select a ready current-version rendition, the managed original or a registered external variant provider. The default names ai-analysis-v1 and permits a bounded temporary copy when that rendition is missing; you do not have to create a permanent JPEG conversion. Configure missing => 'wait' or 'skip' when using the original as fallback is unwanted. Selected inputs are normalized into a provider-compatible, metadata-stripped temporary copy under byte/pixel limits, then discarded. Ordinary saved AVIF/WebP conversions remain unchanged. See input selection for all strategies and missing-input outcomes.

Run explicitly

The actor alias must be registered in Studio. The command checks workspace/tenant access and studio.files.process before reading, before the provider request and again before publication. A changed/trashed version cannot publish a stale result. External sources are not supported by this command yet. A shared atomic cache lock prevents concurrent identical analysis; completed identical profile/version/derivative results are reused. Output is bounded to64short labels and a2000-character description. Moondream structured output is validated; its label scores remain unknown, not fabricated percentages. Google scores and AWS confidence values are retained on a0–1scale, but are not calibrated equivalently across providers. This version stores no raw provider response or per-object geometry. API timeouts and no automatic HTTP retries reduce accidental repeated charges. An ambiguous network failure or process crash can still cause a repeated paid call on explicit retry; this is not an exactly-once billing guarantee. Use least-privilege credentials, supported provider regions and a shared lock cache. No live-provider quality/cost benchmark has been completed.

Optional automatic upload analysis

Install all optional intelligence migrations, configure a provider, then enable:
Run a worker on that queue with its retry-after greater than the job timeout of 570 seconds and lock lease of 600 seconds. Set the worker timeout consistently; configure finite storage/network timeouts. Admission is attached to managed upload publication (including its shared chunk completion path), after which dispatch occurs only after commit. The default remains off. A synchronous queue never executes automatic analysis inside upload. Oversized source files and a full workspace pending queue are not admitted. Existing/externally imported assets are not automatically backfilled by this hook. fms_analysis_jobs preserves queued/running/ready/failed/dispatch_failed/obsolete state across reloads. Workers check current version before analysis, and the analysis service checks permissions again. One attempt avoids automatic repeated billable requests; job retry/recovery UI and durable dispatch reconciliation are not implemented yet. A process killed without a queue failure callback may leave a running record requiring operator intervention. This is not a complete resilient outbox or a monetary cost budget. Do not enable unbounded external retries. Duplicate automatic admission for the same version creates one record. Analysis copies use existing GD/Imagick resource checks, no enlargement, metadata stripping and private temporary streams cleaned after use. Original download is bounded before decode (default10MiB, hard25MiB), source pixels default16MP/hard40MP. Source checksum must match before encoding. No persistent JPEG delivery conversion is added; provider compatibility conversion is temporary only.

Select the image sent to AI

Use an existing conversion without changing the file shown to users:
type accepts auto, original, rendition or provider. An external provider uses the registered Studio variant-provider contract, with provider and preset identifying the source. This supports a custom conversion system without depending on a particular uploader. Missing conversions can use temporary processing of the original, wait or skip (the latter two do not call AI). An existing derivative must belong to the current version and pass access checks. Regardless of the selected source, provider input is subject to the configured byte/pixel caps and prepared in a compatible temporary format. Saved delivery variants are unchanged. This command processes Studio-managed versions. Application Files and Spatie projections require their own explicit processing adapter; registering a source for viewing does not give Studio ownership of its originals.

Sample video frames

See video analysis and bounded frame sampling for the complete FFmpeg/FFprobe configuration, command and partial-result lifecycle.

Review suggestions

Open AI → AI labels in a workspace after enabling image or video analysis. The panel uses cursor pagination, a virtualized list and a bounded retained window. Filter pending, accepted, rejected or all suggestions. Each item identifies the file and provider; opening it uses the ordinary authorized file preview. studio.ai.view permits inspection. studio.ai.review additionally permits accept/reject actions; configure these using Studio’s Spatie permission mapping. Decisions use a revision check, so an outdated browser cannot silently overwrite a newer decision. Replaced or trashed versions are excluded. Accepting a label never writes a manual tag or a Spatie Tag. Review applies to the selected analysis run of the current file version. A future analysis profile can create new suggestions; cross-profile concept-level moderation is not currently applied. Search indexes the latest current-version description and non-rejected labels. Rejecting a label does not rewrite the provider’s prose description. Synchronize the search index after review; stale indexed fingerprints are rejected immediately. Processing and indexing are separate: publishing an analysis does not mean the search engine has already acknowledged its new document. No provider cost, relevance or million-file throughput claim is implied by enabling this feature.