AI & Content Workflows

AI workflows provide unified interfaces to text generation and embeddings with automatic provider fallback routing.

Text Generation

Workflow: ai/generate

Unified text generation with automatic fallback: Fabric API, then Gemini, then OpenAI.

CLI
REST

fabric run ai/generate \
  --input prompt="Explain quantum computing in simple terms" \
  --input model="gemini-2.5-flash"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=ai/generate" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "prompt": "Explain quantum computing in simple terms",
      "model": "gemini-2.5-flash"
    }
  }'

Input

Parameter	Type	Default	Description
`prompt`	`string`	required	Input prompt
`model`	`string`	resolved	Model identifier
`provider`	`string`	`"auto"`	Force provider: `auto`, `gemini`, `openai`, `fabric`
`system_prompt`	`string`	`""`	System instruction
`temperature`	`float`	`0.7`	Sampling temperature
`max_tokens`	`int`	model default	Max output tokens

Output

{
  "response": "Quantum computing uses quantum bits (qubits)...",
  "usage": {"input_tokens": 12, "output_tokens": 150},
  "provider_used": "gemini",
  "model_used": "gemini-2.5-flash"
}

Embeddings

Workflow: ai/embeddings

Generate vector embeddings for text. Used internally by problem intelligence clustering, hook clustering, and search.

CLI
REST

fabric run ai/embeddings \
  --input texts='["first document", "second document"]'

curl -X POST "$FABRIC_URL/v1/workflows/run?name=ai/embeddings" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "texts": ["first document", "second document"]
    }
  }'

Input

Parameter	Type	Default	Description
`texts`	`string` or `list[str]`	required	Text(s) to embed
`embedding_model`	`string`	`"auto"`	Model: `auto`, specific model ID

Output

{
  "embeddings": [[0.012, -0.034, ...], [0.045, 0.023, ...]],
  "model_used": "text-embedding-3-small"
}

Fallback chain: Fabric API, then OpenAI, then local sentence-transformers.

Vision Analysis

Workflow: ai/vision

Analyze images using local Moondream or cloud Gemini — captioning, visual QA, object detection, and asset tagging. Uses the vision model operation, so quality=prefer-local uses Moondream (free, offline) and quality=standard uses Gemini.

CLI
REST

# Caption + tags (default)
fabric run ai/vision \
  --input image_path="photo.jpg"

# Visual question answering
fabric run ai/vision \
  --input image_path="photo.jpg" \
  --input question="How many people are in this image?"

# Multiple operations with local model
fabric run ai/vision \
  --input image_path="photo.jpg" \
  --input 'vision_operations=["caption","tags","detect:person"]' \
  --input quality=prefer-local

curl -X POST "$FABRIC_URL/v1/workflows/run?name=ai/vision" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "image_path": "photo.jpg",
      "vision_operations": ["caption", "tags", "detect:person"]
    }
  }'

Input

Parameter	Type	Default	Description
`image_path`	`string`	required	Path to image file
`vision_operations`	`list[str]`	`["caption", "tags"]`	Operations: `caption`, `tags`, `detect:<object>`, `query:<question>`
`vision_question`	`string`	`null`	Shorthand for a single visual QA question
`quality`	`string`	`"standard"`	`prefer-local` for Moondream, `standard`/`premium` for Gemini

Operations

caption — Generate a scene description
tags — Extract searchable tags (objects, scenes, colors, emotions)
detect:<object> — Object detection with bounding boxes (local Moondream only)
query:<question> — Visual question answering

Output

{
  "caption": "A bright living room with two people sitting on a beige sofa...",
  "tags": ["living room", "person", "sofa", "warm lighting", "indoor"],
  "detections": [
    {"object": "person", "bbox": {"x_min": 0.12, "y_min": 0.3, "x_max": 0.45, "y_max": 0.9}}
  ],
  "answers": [
    {"question": "How many people?", "answer": "2"}
  ],
  "model_used": "moondream-2b"
}

Vision models

Quality	Model	Cost	Notes
`local` / `prefer-local` / `free`	`moondream-2b`	Free	Local inference, ~2-5s/image on CPU
`standard` / `premium`	`gemini-2.5-flash`	~$0.001/image	Cloud API, faster, better for subjective analysis

Install local vision: pip install -e "sdks/fabric-workflow-sdk[local-vision]"

Content Generation

Workflow: content/generate

Analyzes a topic, generates structured content, and evaluates quality.

CLI
REST

fabric run content/generate \
  --input topic="The future of remote work" \
  --input content_type="blog post" \
  --input audience="tech professionals"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=content/generate" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "The future of remote work",
      "content_type": "blog post",
      "audience": "tech professionals"
    }
  }'

Pipeline

analyze_topic → generate_content → evaluate_quality

Input

Parameter	Type	Default	Description
`topic`	`string`	required	Content topic
`audience`	`string`	`""`	Target audience
`content_type`	`string`	`"blog post"`	Format: `blog post`, `email`, `social`, `video script`

Output

{
  "analysis": {
    "key_points": [...],
    "angles": [...]
  },
  "generated_content": {
    "title": "...",
    "content": "Full article text...",
    "meta_description": "...",
    "tags": ["remote-work", "productivity"]
  },
  "quality_evaluation": {
    "score": 8.5,
    "improvements": [...]
  }
}

YouTube Workflows

Title Generation

Workflow: youtube/titles

Generates 10 viral title options with CTR estimation and style diversity.

CLI
REST

fabric run youtube/titles \
  --input topic="AI coding assistants" \
  --input niche="tech"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=youtube/titles" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "AI coding assistants",
      "niche": "tech"
    }
  }'

Thumbnail Generation

Workflow: image/generate

Platform-specific thumbnails with native text rendering.

CLI
REST

fabric run image/generate \
  --input selected_title="Why AI Can't Replace Developers" \
  --input platform="youtube"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=image/generate" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "selected_title": "Why AI Can't Replace Developers",
      "platform": "youtube"
    }
  }'

Aspect ratios: YouTube (16:9), TikTok/Shorts (9:16), Instagram (1:1, 4:5).

Description Generation

Workflow: youtube/descriptions

Auto-generates YouTube descriptions with chapter timestamps, hashtags, and CTAs from the transcript.

CLI
REST

fabric run youtube/descriptions \
  --input selected_title="Why AI Can't Replace Developers" \
  --input full_text="Transcribed text..."

curl -X POST "$FABRIC_URL/v1/workflows/run?name=youtube/descriptions" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "selected_title": "Why AI Can't Replace Developers",
      "full_text": "Transcribed text..."
    }
  }'