Filestack intelligence

Filestack Intelligence

Filestack exposes AI/ML capabilities as CDN transformation tasks that take a file handle and return JSON. They are backed by AWS Rekognition, Google Cloud Vision, AWS Comprehend/Transcribe, PicScout (copyright), and proprietary models.

Use filestack_analyze to invoke them — one tool, one task param.

Tasks

Task	Input	Output	Backed by	Use case
tags	image handle	{ tags: { auto: { keyword: confidence } } }	AWS Rekognition + Google Vision (best-of-both)	Auto-tagging for search/discovery
sfw	image handle	{ sfw: boolean } (true = safe)	AWS Rekognition moderation labels	User-generated content moderation
caption	image handle	{ caption: "a photo of a..." }	Vision-language model	Auto alt-text, accessibility, content description
ocr	image / PDF handle	{ text: "...", confidence, blocks: [...] }	Google Cloud Vision OCR	Receipts, invoices, signage, document digitization
copyright	image handle	{ copyright: boolean, matches: [...] }	PicScout / Getty Images	Detect stock photos / IP violation before publishing
doc_detection	photo of a document	{ detected: boolean, corners?: [[x,y]...] }	Proprietary model	Mobile scanning UX — detect doc edges in a phone photo
image_sentiment	image handle (with faces)	{ sentiment: "positive|neutral|negative", confidence }	AWS Rekognition emotions	Detect mood / engagement in user photos
text_sentiment	text string (no handle)	{ sentiment, confidence, language }	AWS Comprehend	Analyze captions, comments, reviews

Choosing the right task

• Moderation pipeline (UGC platforms): chain sfw then tags. If sfw=false, reject; if sfw=true but tags contain words like "weapon" / "violence", flag for review.
• Alt-text / accessibility: caption — single natural-language sentence.
• Search/categorization: tags — gives you 20-50 keywords with confidence scores.
• Compliance (stock photo enforcement, DMCA risk): copyright before allowing user uploads to be displayed publicly.
• Mobile doc scanner: doc_detection returns corners; combine with partial_pixelate on sensitive regions.
• Reviews / comments analysis: text_sentiment — no file needed, pass the text directly.

Examples

One-shot SFW + tags moderation check

// In your upload handler, after filestack_upload
const sfw = await callTool('filestack_analyze', { task: 'sfw', handleOrText: handle });
if (!sfw.result.sfw) return reject('NSFW content detected');

const tags = await callTool('filestack_analyze', { task: 'tags', handleOrText: handle });
const auto = tags.result.tags.auto;
const flagged = ['weapon', 'violence', 'drug'].filter(k =>
  Object.keys(auto).some(t => t.toLowerCase().includes(k))
);
if (flagged.length) return flagForReview({ handle, reasons: flagged });

Generate caption + tags in one pass (for SEO alt-text + indexing)

const [cap, tags] = await Promise.all([
  callTool('filestack_analyze', { task: 'caption', handleOrText: handle }),
  callTool('filestack_analyze', { task: 'tags', handleOrText: handle }),
]);
const altText = cap.result.caption;
const keywords = Object.entries(tags.result.tags.auto)
  .filter(([, conf]) => Number(conf) > 70)
  .map(([k]) => k);

OCR a receipt photo

const ocr = await callTool('filestack_analyze', { task: 'ocr', handleOrText: handle });
// ocr.result.text — full text in reading order
// ocr.result.blocks — bounding boxes if you need positional info

Detect document edges before further processing

const doc = await callTool('filestack_analyze', {
  task: 'doc_detection',
  handleOrText: handle,
  options: { coords: true, preprocess: true },
});
// doc.result.corners → use to crop the photo to just the document

Sentiment on user reviews

const sent = await callTool('filestack_analyze', {
  task: 'text_sentiment',
  handleOrText: 'This product changed my life',
  options: { language: 'en' },
});
// sent.result.sentiment === 'positive'

Important notes

1. Security policies must include convert. For security-enabled apps, the policy used to sign the request needs convert in the call array — read alone won't work for intelligence tasks. Generate the policy via filestack_generate_signed_url with call: ['read', 'convert'].

2. tags returns confidence scores 0-100, not 0-1. Threshold at ~60-70 for high-precision filtering.

3. sfw is binary, not a probability. Don't show partial confidence; Filestack does the thresholding server-side.

4. text_sentiment does NOT take a handle — pass the raw text in handleOrText and optionally options.language. Other tasks all take a handle.

5. Caching: All intelligence results are CDN-cached by handle for 24h. Re-calling filestack_analyze with the same handle is free.

6. Cost / limits: Intelligence tasks count against a separate quota in the Filestack Developer Portal. Heavy usage benefits from batching via filestack_run_workflow with a designed pipeline (e.g., upload → sfw → tags → store metadata in one async job).

7. Video intelligence (video_sfw, video_tagging) is NOT exposed via direct transform tasks — it must go through Workflows. Use filestack_run_workflow with a workflow that includes the relevant video intelligence step.