@framers/agentos-skills-registry 0.11.0 → 0.13.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@framers/agentos-skills-registry",
3
- "version": "0.11.0",
3
+ "version": "0.13.0",
4
4
  "files": [
5
5
  "dist",
6
6
  "registry",
@@ -1,12 +1,12 @@
1
1
  ---
2
2
  name: image-gen
3
- version: '1.0.0'
4
- description: Generate images from text prompts using AI image generation APIs like DALL-E, Stable Diffusion, or Midjourney.
3
+ version: '2.0.0'
4
+ description: Generate, edit, upscale, and variate images using the AgentOS multi-provider image pipeline with automatic fallback.
5
5
  author: Wunderland
6
6
  namespace: wunderland
7
7
  category: creative
8
- tags: [image-generation, ai-art, dall-e, stable-diffusion, creative, visual]
9
- requires_secrets: [openai.api_key]
8
+ tags: [image-generation, ai-art, dall-e, stable-diffusion, flux, replicate, stability, fal, creative, visual]
9
+ requires_secrets: []
10
10
  requires_tools: [generate_image]
11
11
  metadata:
12
12
  agentos:
@@ -15,36 +15,89 @@ metadata:
15
15
  homepage: https://platform.openai.com/docs/guides/images
16
16
  ---
17
17
 
18
- # AI Image Generation
18
+ # AI Image Generation Workflow
19
19
 
20
- Use the `generate_image` tool to create images from text descriptions. Two providers are supported:
21
- - **DALL-E 3** (OpenAI) — requires `OPENAI_API_KEY`
22
- - **Stability AI** (SDXL) — requires `STABILITY_API_KEY`
20
+ Use this skill when the user wants to create, edit, upscale, or create variations of images. AgentOS provides four high-level APIs that route to any configured provider with automatic fallback when multiple providers have credentials set.
21
+
22
+ ## The Four High-Level APIs
23
+
24
+ 1. **`generateImage()`** — Create new images from text prompts.
25
+ 2. **`editImage()`** — Transform existing images via img2img, inpainting, or outpainting.
26
+ 3. **`upscaleImage()`** — Increase resolution (2x or 4x super-resolution).
27
+ 4. **`variateImage()`** — Generate visual variations of an existing image.
23
28
 
24
29
  If the `generate_image` tool is not loaded, enable it with `extensions_enable image-generation`.
25
30
 
26
- Craft detailed, effective prompts that translate the user's creative vision into high-quality generated images.
31
+ ## Provider Selection Guide
32
+
33
+ Choose the provider based on the user's priority:
34
+
35
+ | Priority | Provider | Env Var | Best For |
36
+ |----------|----------|---------|----------|
37
+ | Quality | **OpenAI** (GPT-Image-1, DALL-E 3) | `OPENAI_API_KEY` | Highest fidelity, prompt adherence, text-in-image |
38
+ | Control | **Stability AI** (SDXL, SD3, Ultra) | `STABILITY_API_KEY` | Negative prompts, style presets, cfg/steps tuning |
39
+ | Speed | **BFL / Flux** (Flux Pro 1.1) | `BFL_API_KEY` | Fast generation with strong quality |
40
+ | Speed | **Fal** (Flux Dev) | `FAL_API_KEY` | Serverless Flux inference, low latency |
41
+ | Variety | **Replicate** (Flux, SDXL, community models) | `REPLICATE_API_TOKEN` | Access to thousands of community models |
42
+ | Cost | **OpenRouter** (routes to cheapest) | `OPENROUTER_API_KEY` | Provider-agnostic routing, best price |
43
+ | Privacy | **Local SD** (A1111 / ComfyUI) | `STABLE_DIFFUSION_LOCAL_BASE_URL` | Fully offline, no data leaves the machine |
44
+
45
+ When multiple providers are configured, AgentOS wraps them in a **FallbackImageProxy** — if the primary provider fails (rate limit, outage, etc.), the request automatically retries on the next available provider in priority order.
46
+
47
+ ## Operation Decision Tree
48
+
49
+ Use this to pick the right API for the user's request:
50
+
51
+ - **"Generate / create / draw / imagine"** -> `generateImage()`
52
+ - **"Edit / change / modify / transform"** -> `editImage()` with `mode: 'img2img'`
53
+ - **"Remove / fill in / fix this area"** -> `editImage()` with `mode: 'inpaint'` + mask
54
+ - **"Extend / expand the borders"** -> `editImage()` with `mode: 'outpaint'`
55
+ - **"Make it higher resolution / sharper"** -> `upscaleImage()` with `scale: 2` or `4`
56
+ - **"Show me variations / alternatives"** -> `variateImage()` with `n: 3-4`
57
+
58
+ ## Prompt Engineering Tips
59
+
60
+ A strong image prompt has five components:
61
+
62
+ 1. **Subject** — What is in the image. Be specific: "a red panda sitting on a mossy branch" not "an animal."
63
+ 2. **Style** — Artistic approach: photorealistic, watercolor, pixel art, oil painting, vector illustration, cinematic, anime.
64
+ 3. **Composition** — Camera angle and framing: close-up portrait, wide establishing shot, overhead flat lay, isometric.
65
+ 4. **Lighting and Color** — Mood through light: golden hour, dramatic side-lighting, neon glow, muted earth tones, high contrast.
66
+ 5. **Atmosphere** — Emotional tone: serene, ominous, whimsical, nostalgic, futuristic.
27
67
 
28
- When generating images, help the user refine their prompt for best results. A good image prompt includes: subject description, style (photorealistic, illustration, watercolor, etc.), composition (close-up, wide shot, overhead), lighting (natural, dramatic, soft), color palette, and mood/atmosphere. Offer prompt suggestions when the user's description is vague or underspecified.
68
+ Additional tips:
69
+ - Front-load the most important elements. Models weight earlier tokens more heavily.
70
+ - Use negative prompts (Stability, Local SD) to exclude unwanted elements: "no text, no watermark, no blurry."
71
+ - For text-in-image, OpenAI GPT-Image-1 is the most reliable. Other models struggle with legible text.
72
+ - Request `quality: 'hd'` for DALL-E 3 when detail matters (doubles cost).
73
+ - For consistent characters across multiple images, describe the character in detail each time or use img2img with a reference.
29
74
 
30
- Support different image sizes and aspect ratios based on the API capabilities (1024x1024, 1792x1024, 1024x1792 for DALL-E 3). For iterative refinement, maintain context from previous generations so the user can say "make it more vibrant" or "change the background to a beach." Save generated images to the filesystem when the user requests it, with descriptive filenames.
75
+ ## Sizes and Aspect Ratios
31
76
 
32
- When the user requests variations or edits of existing images, use the appropriate API endpoints (variations, inpainting) when available. For batch generation, create multiple variations with slightly different prompts to give the user options. Always inform the user of the model and settings used for each generation.
77
+ | Provider | Supported Sizes | Aspect Ratio Support |
78
+ |----------|----------------|---------------------|
79
+ | OpenAI | 1024x1024, 1792x1024, 1024x1792 | Via size selection |
80
+ | Stability | Flexible | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, etc. |
81
+ | Replicate/Flux | Flexible | `aspectRatio` parameter |
82
+ | Local SD | Any (multiples of 64) | Via `width`/`height` |
33
83
 
34
84
  ## Examples
35
85
 
36
- - "Generate an image of a cozy cabin in the mountains at sunset, watercolor style"
37
- - "Create a professional logo for a coffee shop called 'Bean There'"
38
- - "Make the previous image more dramatic with storm clouds"
39
- - "Generate 3 variations of a cyberpunk cityscape at night"
40
- - "Create a 16:9 landscape of a serene Japanese garden in spring"
86
+ - "Generate a photorealistic image of a cozy cabin in the mountains at sunset."
87
+ - "Create a professional logo for a coffee shop called 'Bean There' — vector illustration style, clean lines."
88
+ - "Edit this photo: make the sky more dramatic with storm clouds." (img2img)
89
+ - "Remove the person from the background of this product photo." (inpaint + mask)
90
+ - "Upscale this thumbnail to 4x resolution for print."
91
+ - "Show me 3 variations of this hero image with different color palettes."
92
+ - "Generate a 16:9 cinematic landscape of a neon-lit Tokyo street at night in the rain."
41
93
 
42
94
  ## Constraints
43
95
 
44
96
  - Image generation costs API credits per request; inform the user of approximate costs when possible.
45
- - Content policy restrictions apply: no realistic faces of real people, no violent/explicit content.
46
- - DALL-E 3 does not support exact image editing or inpainting; describe the full desired output.
97
+ - Content policy restrictions apply per provider: no realistic faces of real people, no violent/explicit content.
98
+ - DALL-E 3 does not support native inpainting use GPT-Image-1 or Stability for mask-based editing.
99
+ - Upscaling is not supported by OpenAI or OpenRouter — use Stability, Replicate, or Local SD.
47
100
  - Generated images may not perfectly match the prompt; iterative refinement is expected.
48
- - Maximum prompt length varies by model (DALL-E 3: 4,000 characters).
49
- - Image quality and style depend on the model version and generation parameters.
50
- - Generated images should not be represented as photographs or real events.
101
+ - Maximum prompt length varies by model (DALL-E 3: 4,000 chars; Stability: 2,000 chars).
102
+ - Local SD requires a running A1111 or ComfyUI instance with the API enabled.
103
+ - The fallback chain only activates when the primary provider fails; it does not merge results from multiple providers.
@@ -1,23 +1,153 @@
1
1
  ---
2
2
  name: multimodal-rag
3
- description: Index and search across text, images, audio, video, and PDFs
4
- version: 1.0.0
5
- tags: [rag, multimodal, image, audio, video, pdf, search]
6
- tools_required: [vision-pipeline]
3
+ version: '2.0.0'
4
+ description: Index and search across text, images, audio, video, and PDFs via the multimodal RAG pipeline and HTTP API.
5
+ author: Wunderland
6
+ namespace: wunderland
7
+ category: productivity
8
+ tags: [rag, multimodal, image, audio, video, pdf, search, indexing, memory]
9
+ requires_secrets: []
10
+ requires_tools: [vision-pipeline]
11
+ metadata:
12
+ agentos:
13
+ emoji: "\U0001F50D"
7
14
  ---
8
15
 
9
16
  # Multimodal RAG
10
17
 
11
- Index and retrieve content across all modalities -- text, images, audio, video, and PDFs. Images are described via vision AI, audio is transcribed, video frames are extracted, and PDFs are parsed. All content is embedded and searchable.
18
+ Use this skill when the user wants to index, search, or retrieve content across multiple modalities -- text, images, audio, video, and documents (PDF, DOCX, Markdown, CSV, JSON, XML). All non-text content is converted to a text representation (vision description, STT transcript, document parse) before embedding, so every modality is searchable with the same text query.
19
+
20
+ ## Architecture
21
+
22
+ ```
23
+ Image --> Vision LLM --> description --> embed --> vector store
24
+ Audio --> STT --> transcript --> embed --> vector store
25
+ Video --> ffmpeg (frames + audio) --> vision + STT --> vector store
26
+ PDF --> text extraction + chunking --> embed --> vector store
27
+ ```
28
+
29
+ When cognitive memory is enabled via `MultimodalMemoryBridge`, ingested content also creates memory traces so agents can recall multimodal content during conversation without an explicit search.
12
30
 
13
31
  ## Capabilities
14
- - **Image indexing**: Vision LLM describes -> embed -> search
15
- - **Audio indexing**: STT transcribes -> embed -> search
16
- - **Video indexing**: Frame extraction + audio transcription
17
- - **PDF indexing**: Text + embedded image extraction
18
- - **Cross-modal search**: Query returns results from all modalities
19
-
20
- ## Example
21
- "Find images related to quantum computing"
22
- "Search my audio recordings for mentions of the project deadline"
23
- "What does the diagram in page 3 of the PDF show?"
32
+
33
+ - **Image indexing** Vision LLM describes the image, description is embedded and searchable.
34
+ - **Audio indexing** STT transcribes the audio, transcript is chunked and searchable.
35
+ - **Video indexing** Frame extraction (vision) + audio transcription (STT), both indexed.
36
+ - **Document indexing** PDF, DOCX, TXT, Markdown, CSV, JSON, XML text extracted and indexed.
37
+ - **Cross-modal search** — A single text query returns results from all modalities, ranked by relevance.
38
+ - **Query-by-image** — Upload an image to find similar indexed content.
39
+ - **Query-by-audio** Upload audio to find related indexed content via transcript matching.
40
+
41
+ ## HTTP API Routes
42
+
43
+ All routes are mounted under `/api/agentos/rag/multimodal`. Ingestion routes accept `multipart/form-data`.
44
+
45
+ ### Ingest
46
+
47
+ | Method | Path | Field | Description |
48
+ |--------|------|-------|-------------|
49
+ | POST | `/images/ingest` | `image` | Ingest an image (max 15 MB). Vision LLM generates description. |
50
+ | POST | `/audio/ingest` | `audio` | Ingest audio (max 25 MB). STT generates transcript. |
51
+ | POST | `/documents/ingest` | `document` | Ingest a document (max 30 MB). Text extracted and chunked. |
52
+
53
+ Common form fields for all ingest routes:
54
+
55
+ | Field | Type | Description |
56
+ |-------|------|-------------|
57
+ | `collectionId` | string | Target collection (default: auto) |
58
+ | `assetId` | string | Optional custom ID for the asset |
59
+ | `category` | string | `conversation_memory`, `knowledge_base`, `user_notes`, `system`, `custom` |
60
+ | `tags` | string | Comma-separated or JSON array of tags |
61
+ | `metadata` | string | JSON object with arbitrary metadata |
62
+ | `storePayload` | boolean | Whether to store the raw binary (for later download) |
63
+ | `sourceUrl` | string | Original URL of the content |
64
+ | `textRepresentation` | string | Override auto-generated description/transcript |
65
+ | `userId` | string | Owner user ID |
66
+ | `agentId` | string | Owner agent ID |
67
+
68
+ ### Query
69
+
70
+ | Method | Path | Body / Field | Description |
71
+ |--------|------|-------------|-------------|
72
+ | POST | `/query` | JSON body | Text query across all modalities |
73
+ | POST | `/images/query` | `image` field | Query by uploading an image |
74
+ | POST | `/audio/query` | `audio` field | Query by uploading audio |
75
+
76
+ Text query body:
77
+
78
+ ```json
79
+ {
80
+ "query": "quantum computing diagrams",
81
+ "modalities": ["image", "audio", "document"],
82
+ "collectionIds": ["knowledge-base"],
83
+ "topK": 10,
84
+ "includeMetadata": true
85
+ }
86
+ ```
87
+
88
+ Image/audio query form fields:
89
+
90
+ | Field | Type | Description |
91
+ |-------|------|-------------|
92
+ | `modalities` | string | Comma-separated: `image`, `audio`, `document` |
93
+ | `collectionIds` | string | Comma-separated collection IDs to search |
94
+ | `topK` | number | Max results (default: 5) |
95
+ | `includeMetadata` | boolean | Include stored metadata in results |
96
+ | `retrievalMode` | string | `auto` (default), `text`, `native`, `hybrid` |
97
+
98
+ ### Asset Management
99
+
100
+ | Method | Path | Description |
101
+ |--------|------|-------------|
102
+ | GET | `/assets/:assetId` | Get asset metadata |
103
+ | GET | `/assets/:assetId/content` | Download raw binary (if `storePayload` was true) |
104
+ | DELETE | `/assets/:assetId` | Delete asset and its embeddings |
105
+
106
+ ## Retrieval Modes
107
+
108
+ - **`auto`** (default) — Text-first retrieval with native augmentation when available.
109
+ - **`text`** — Derive a caption/transcript and query the standard text pipeline only.
110
+ - **`native`** — Use modality-native embeddings (e.g. CLIP for images) when available.
111
+ - **`hybrid`** — Combine text and native retrieval, merge and re-rank results.
112
+
113
+ ## Programmatic Usage
114
+
115
+ ```typescript
116
+ import { MultimodalMemoryBridge } from 'agentos/rag/multimodal';
117
+
118
+ // Ingest an image
119
+ await bridge.ingestImage(imageBuffer, { source: 'upload', tags: ['product'] });
120
+
121
+ // Ingest audio
122
+ await bridge.ingestAudio(audioBuffer, { language: 'en' });
123
+
124
+ // Ingest video (requires ffmpeg)
125
+ await bridge.ingestVideo(videoBuffer, { extractFrames: true });
126
+
127
+ // Ingest PDF
128
+ await bridge.ingestPDF(pdfBuffer, { extractImages: true });
129
+
130
+ // Cross-modal search
131
+ const results = await indexer.search('quantum computing', {
132
+ topK: 10,
133
+ modalities: ['image', 'text', 'audio'],
134
+ });
135
+ ```
136
+
137
+ ## Examples
138
+
139
+ - "Index this product photo so I can find it by description later."
140
+ - "Ingest all the PDFs in this folder into my knowledge base."
141
+ - "Search my audio recordings for mentions of the quarterly budget."
142
+ - "Find images related to the network architecture diagram."
143
+ - "What does the chart on page 5 of the annual report show?"
144
+ - "Upload this meeting recording and make it searchable."
145
+
146
+ ## Constraints
147
+
148
+ - Image uploads are capped at 15 MB, audio at 25 MB, documents at 30 MB.
149
+ - Supported audio formats: MP3, MP4, M4A, WAV, WebM, OGG (Whisper-compatible).
150
+ - Supported document formats: PDF, DOCX, TXT, Markdown, CSV, JSON, XML.
151
+ - Video ingestion requires ffmpeg installed on the system.
152
+ - Vision LLM and STT provider must be configured for image/audio indexing respectively.
153
+ - Cross-modal search ranks by cosine similarity of embedded text representations; it does not perform true multimodal embedding fusion unless `retrievalMode: 'native'` is used with a CLIP-like model.