28 KiB

Raw Blame History

Banatie API v1 - Technical Changes and Refactoring

Context

Project is in active development with no existing clients. All changes can be made without backward compatibility concerns. Priority: high-quality and correct API implementation.

1. Parameter Naming Cleanup ✅

1.1 POST /api/v1/generations

Current parameters:

assignAlias → rename to alias
assignFlowAlias → rename to flowAlias

Rationale: Shorter, clearer, no need for "assign" prefix when assignment is obvious from endpoint context.

Affected areas:

Request type definitions
Route handlers
Service methods
API documentation

1.2 Reference Images Auto-Detection

Parameter behavior:

referenceImages parameter is optional
If provided (array of aliases or IDs) → use these images as references
If empty or not provided → service must automatically parse prompt and find all aliases

Auto-detection logic:

Prompt parsing:
- Scan prompt text for all alias patterns (@name)
- Extract all found aliases
- Resolve each alias to actual image ID
Manual override:
- If referenceImages parameter is provided and not empty → use only specified images
- Manual list takes precedence over auto-detected aliases
Combined approach:
- If referenceImages provided → add to auto-detected aliases (merge)
- Remove duplicates
- Maintain order: manual references first, then auto-detected

Example:

// Auto-detection (no referenceImages parameter)
{
  "prompt": "A landscape based on @sunset with elements from @mountain"
  // System automatically detects @sunset and @mountain
}

// Manual specification
{
  "prompt": "A landscape",
  "referenceImages": ["@sunset", "image-uuid-123"]
  // System uses only specified images
}

// Combined
{
  "prompt": "A landscape based on @sunset",
  "referenceImages": ["@mountain"]
  // System uses both @mountain (manual) and @sunset (auto-detected)
}

Implementation notes:

Alias detection must use the same validation rules as alias creation
Invalid aliases in prompt should be logged but not cause generation failure
Maximum reference images limit still applies after combining manual + auto-detected

2. Enhanced Prompt Support - Logic Redesign

2.1 Database Schema Changes

Required schema modifications:

Rename field: enhancedPrompt → originalPrompt
Change field semantics:
- prompt - ALWAYS contains the prompt that was used for generation (enhanced or original)
- originalPrompt - contains user's original input ONLY if autoEnhance was used (nullable)

Field population logic:

Case 1: autoEnhance = false
  prompt = user input
  originalPrompt = NULL

Case 2: autoEnhance = true
  prompt = enhanced prompt (used for generation)
  originalPrompt = user input (preserved)

2.2 API Response Format

Response structure:

{
  "prompt": "detailed enhanced prompt...",  // Always the prompt used for generation
  "originalPrompt": "sunset"                // Only present if enhancement was used
}

Affected endpoints:

POST /api/v1/generations response
GET /api/v1/generations/:id response
GET /api/v1/generations list response

3. Regeneration Endpoint Refactoring ✅

3.1 Endpoint Rename

Change:

❌ OLD: POST /api/v1/generations/:id/retry
✅ NEW: POST /api/v1/generations/:id/regenerate

3.2 Remove Status Checks

Remove if (original.status === 'success') throw error check
Remove GENERATION_ALREADY_SUCCEEDED error constant
Allow regeneration for any status (pending, processing, success, failed)

3.3 Remove Retry Logic

Remove retryCount >= MAX_RETRY_COUNT check
Remove retryCount increment
Remove MAX_RETRY_COUNT constant

3.4 Remove Override Parameters

Remove prompt and aspectRatio parameters from request body
Always regenerate with exact same parameters as original

3.5 Image Update Behavior

Update existing image instead of creating new:

Preserve:

imageId (UUID remains the same)
storageKey (MinIO path)
storageUrl
alias (if assigned)
createdAt (original creation timestamp)

Update:

Physical file in MinIO (overwrite)
fileSize (if changed)
updatedAt timestamp

Generation record:

Update status → processing → success/failed
Update processingTimeMs
Keep outputImageId (same value)
Keep flowId (if present)

3.6 Additional Endpoint

Add for Flow:

POST /api/v1/flows/:id/regenerate
Regenerates the most recent generation in flow
Returns FLOW_HAS_NO_GENERATIONS error if flow is empty
Uses parameters from the last generation in flow

4. Flow Auto-Creation (Lazy Flow Pattern)

4.1 Lazy Flow Creation Strategy

Concept:

First request without flowId → return generated flowId in response, but DO NOT create in DB
Any request with valid flowId → create flow in DB if doesn't exist, add this request to flow
If flowAlias specified in request → create flow immediately (eager creation)

4.2 Implementation Details

Flow ID Generation:

When generation/upload has no flowId, generate UUID for potential flow
Return this flowId in response
Save flowId in generation/image record, but DO NOT create flow record

Flow Creation in DB:

Trigger: ANY request with valid flowId value

Logic:

Check if flow record exists in DB
Check if there are existing generations/images with this flowId
If flow doesn't exist:
- Create flow record with provided flowId
- Include all existing records with this flowId
- Maintain chronological order based on createdAt timestamps
If flow exists:
- Add new record to existing flow

Eager creation:

If request includes flowAlias → create flow immediately
Set alias in flow.aliases object

Database Schema:

generations table already has flowId field (foreign key to flows.id)
images table already has flowId field (foreign key to flows.id)
No schema changes needed

Orphan flowId handling:

If flowId exists in generation/image record but not in flows table - this is normal
Such records are called "orphans" and simply not shown in GET /api/v1/flows list
No cleanup job needed
Do NOT delete such records automatically
System works correctly with orphan flowIds until flow record is created

4.3 Endpoint Changes

Remove:

❌ POST /api/v1/flows endpoint (no longer needed)

Modify responses:

POST /api/v1/generations → always return flowId in response (see section 10.1)
POST /api/v1/images/upload → always return flowId in response (see section 10.1)

5. Upload Image Enhancements

5.1 Add Parameters

POST /api/v1/images/upload:

Parameters:

alias (optional, string) - project-scoped alias
flowAlias (optional, string) - flow-scoped alias for uploaded image
flowId (optional, string) - flow association

Behavior:

If flowAlias and flowId specified:
- Ensure flow exists (or create via lazy pattern)
- Add alias to flow.aliases object
If flowAlias WITHOUT flowId:
- Apply lazy flow creation with eager pattern
- Create flow immediately, set flowAlias
If only alias specified:
- Set project-scoped alias on image

5.2 Alias Conflict Resolution

Validation rules:

Technical aliases are forbidden:
- Cannot use: @last, @first, @upload or any reserved technical alias
- Return validation error if attempted
Alias override behavior:
- If alias already exists → new request has higher priority
- Alias points to new image
- Previous image loses its alias but is NOT deleted
- Same logic applies to both project aliases and flow aliases
Applies to both:
- Image upload with alias
- Generation with alias/flowAlias

Example:

State: Image A has alias "@hero"
Request: Upload Image B with alias "@hero"
Result:
  - Image B now has alias "@hero"
  - Image A loses alias (alias = NULL)
  - Image A is NOT deleted

6. Image Alias Management Refactoring

6.1 Endpoint Consolidation

Remove alias handling from:

❌ PUT /api/v1/images/:id (body: { alias, focalPoint, meta })
- Remove alias parameter
- Keep only focalPoint and meta

Single method for project-scoped alias management:

✅ PUT /api/v1/images/:id/alias (body: { alias })
- Set new alias
- Change existing alias
- Remove alias (pass alias: null)

Rationale: Explicit intent, dedicated endpoint for alias operations, simpler validation.

6.2 Alias as Image Identifier

Support alias in path parameters:

Syntax:

UUID: GET /api/v1/images/550e8400-e29b-41d4-a716-446655440000
Alias: GET /api/v1/images/@hero
@ symbol distinguishes alias from UUID (UUIDs never contain @)

UUID validation: UUIDs can NEVER contain @ symbol - this guarantees no conflicts

Flow-scoped resolution:

GET /api/v1/images/@hero?flowId=uuid-123
Searches for alias @hero in context of flow uuid-123
Uses 3-tier precedence (technical → flow → project)

Endpoints with alias support:

GET /api/v1/images/:id_or_alias
PUT /api/v1/images/:id_or_alias (for focalPoint, meta)
PUT /api/v1/images/:id_or_alias/alias
DELETE /api/v1/images/:id_or_alias

Implementation:

Check first character of path parameter
If starts with @ → resolve via AliasService
If doesn't start with @ → treat as UUID
After resolution, work with imageId as usual

6.3 CDN-style Image URLs with Alias Support

Current URL format must be changed.

New standardized URL patterns:

For all generated and uploaded images:

GET /cdn/:orgSlug/:projectSlug/img/:filenameOrAlias

For live URLs:

GET /cdn/:orgSlug/:projectSlug/live/:scope?prompt=...&aspectRatio=...

All image URLs returned by API must follow this pattern.

Resolution Logic:

Check if :filenameOrAlias starts with @
If yes → resolve alias via AliasService
If no → search by filename/storageKey
Return image bytes with proper content-type headers

Response Headers:

Content-Type: image/jpeg (or appropriate MIME type)
Cache-Control: public, max-age=31536000
ETag: based on imageId or fileHash

URL Encoding for prompts:

Spaces can be replaced with underscores _ for convenience
Both prompt=beautiful%20sunset and prompt=beautiful_sunset are valid
System should handle both formats

Examples:

GET /cdn/acme/website/img/@hero        → resolve @hero alias
GET /cdn/acme/website/img/logo.png     → find by filename
GET /cdn/acme/website/img/@product-1   → resolve @product-1 alias

Error Handling:

Alias not found → 404
Filename not found → 404
Multiple matches → alias takes priority over filename

7. Deletion Strategy Overhaul

7.1 Image Deletion (Hard Delete)

DELETE /api/v1/images/:id

Operations:

Delete physical file from MinIO storage
Delete record from images table (hard delete)
Cascade: set outputImageId = NULL in related generations
Cascade: completely remove alias entries from all flow.aliases where imageId is referenced
- Remove entire key-value pairs, not just values
Cascade: remove imageId from generation.referencedImages JSON arrays

Example cascade for flow.aliases:

Before: flow.aliases = { "@hero": "img-123", "@product": "img-456" }
Delete img-123
After: flow.aliases = { "@product": "img-456" }

Rationale: User wants to delete - remove completely, free storage. Alias entries are also completely removed.

7.2 Generation Deletion (Conditional)

DELETE /api/v1/generations/:id

Behavior depends on output image alias:

Case 1: Output image WITHOUT project alias

Delete output image completely (hard delete with MinIO cleanup)
Delete generation record (hard delete)

Case 2: Output image WITH project alias

Keep output image (do not delete)
Delete only generation record (hard delete)
Set generationId = NULL in image record

Decision Logic:

If outputImage.alias !== null → keep image, delete only generation
If outputImage.alias === null → delete both image and generation

Rationale:

Image with project alias is used as standalone asset, preserve it
Image without alias was created only for this generation, delete together

No regeneration of deleted generations - deleted generations cannot be regenerated

7.3 Flow Deletion (Cascade with Alias Protection)

DELETE /api/v1/flows/:id

Operations:

Delete flow record from DB
Cascade: delete all generations associated with this flowId
Cascade: delete all images associated with this flowId EXCEPT images with project alias

Detailed Cascade Logic:

For Generations:

Delete each generation (follows conditional delete from 7.2)
If output image has no alias → delete image
If output image has alias → keep image, set generationId = NULL, set flowId = NULL

For Images (uploaded):

If image has no alias → delete (with MinIO cleanup)
If image has alias → keep, set flowId = NULL

Summary:

Flow record → DELETE
All generations → DELETE
Images without alias → DELETE (with MinIO cleanup)
Images with project alias → KEEP (unlink: flowId = NULL)

Rationale: Flow deletion removes all content except images with project aliases (used globally in project).

7.4 Transactional Delete Pattern

All delete operations must be transactional:

Delete from MinIO storage first
Delete from database (with cascades)
If MinIO delete fails → rollback DB transaction
If DB delete fails → cleanup MinIO file (or rollback if possible)
Log all delete operations for audit trail

Principle: System must be designed so orphaned files in MinIO NEVER occur.

Database Constraints:

ON DELETE CASCADE for appropriate foreign keys
ON DELETE SET NULL where related records must be preserved
Proper referential integrity

No background cleanup jobs needed - system is self-sufficient and always consistent.

8. Live URL System

8.1 Core Concept

Purpose: Permanent URLs that can be immediately inserted into HTML and work forever.

Use Case:

<img src="https://banatie.app/cdn/acme/website/live/hero-section?prompt=beautiful_sunset&aspectRatio=16:9"/>

Key Points:

URL is constructed immediately and used permanently
No preliminary generation through API needed
No signed URLs or tokens in query params
First request → generation, subsequent → cache

8.2 URL Format & Structure

URL Pattern:

/cdn/:orgSlug/:projectSlug/live/:scope?prompt=...&aspectRatio=...

URL Components:

/cdn/acme/website/live/hero-section?prompt=beautiful_sunset&aspectRatio=16:9
     │    │       │    │           │
     │    │       │    │           └─ Generation params (query string)
     │    │       │    └─ Scope identifier
     │    │       └─ "live" prefix
     │    └─ Project slug
     └─ Organization slug

Scope Parameter:

Name: scope (confirmed)
Purpose: logical separation of live URLs within project
Format: alphanumeric + hyphens + underscores
Any user can specify any scope (no validation/signature required)

8.3 First Request Flow

Cache MISS (first request):

Parse orgSlug, projectSlug, scope from URL
Compute cache key: hash(projectId + scope + prompt + params)
Check if image exists in cache
If NOT found:
- Check scope settings (allowNewGenerations, limit)
- Trigger image generation
- Create database records (generation, image, cache entry)
- Wait for generation to complete
- Return image bytes

Response:

Content-Type: image/jpeg
Cache-Control: public, max-age=31536000
X-Cache-Status: MISS
X-Scope: hero-section
X-Image-Id: uuid

Cache HIT (subsequent requests):

Same cache key lookup
Found existing image
Return cached image bytes immediately

Response:

Content-Type: image/jpeg
Cache-Control: public, max-age=31536000
X-Cache-Status: HIT
X-Image-Id: uuid

Generation in Progress:

If image is not in cache but generation is already running:
System must have internal status to track this
Wait for generation to complete
Return image bytes immediately when ready
This ensures consistent behavior for concurrent requests

8.4 Scope Management

Database Table: live_scopes

Create dedicated table with fields:

id (UUID, primary key)
project_id (UUID, foreign key to projects)
slug (TEXT, unique within project) - used in URL
allowNewGenerations (BOOLEAN, default: true) - controls if new generations can be triggered
newGenerationsLimit (INTEGER, default: 30) - max number of generations in this scope
created_at (TIMESTAMP)
updated_at (TIMESTAMP)

Scope Behavior:

allowNewGenerations:

Controls whether new generations can be triggered in this scope
Already generated images are ALWAYS served publicly regardless of this setting
Default: true

newGenerationsLimit:

Limit on number of generations in this scope
Only affects NEW generations, does not affect regeneration
Default: 30

Scope Creation:

Manual: via dedicated endpoint (see below)
Automatic: when new scope is used in live URL (if project allows)

Project-level Settings:

Add to projects table or settings:

allowNewLiveScopes (BOOLEAN, default: true) - allows creating new scopes via live URLs
- If false: new scopes cannot be created via live URL
- If false: scopes can still be created via API endpoint
newLiveScopesGenerationLimit (INTEGER, default: 30) - generation limit for auto-created scopes
- This value is set as newGenerationsLimit for newly created scopes

8.5 Scope Management API

Create scope (manual):

POST /api/v1/live/scopes
Headers: X-API-Key: bnt_project_key
Body: {
  "slug": "hero-section",
  "allowNewGenerations": true,
  "newGenerationsLimit": 50
}

List scopes:

GET /api/v1/live/scopes
Headers: X-API-Key: bnt_project_key
Response: {
  "scopes": [
    {
      "id": "uuid",
      "slug": "hero-section",
      "allowNewGenerations": true,
      "newGenerationsLimit": 50,
      "currentGenerations": 23,
      "lastGeneratedAt": "2024-01-15T10:30:00Z"
    }
  ]
}

Get scope details:

GET /api/v1/live/scopes/:slug
Headers: X-API-Key: bnt_project_key
Response: {
  "id": "uuid",
  "slug": "hero-section",
  "allowNewGenerations": true,
  "newGenerationsLimit": 50,
  "currentGenerations": 23,
  "images": [...]
}

Update scope:

PUT /api/v1/live/scopes/:slug
Headers: X-API-Key: bnt_project_key
Body: {
  "allowNewGenerations": false,
  "newGenerationsLimit": 100
}

Regenerate scope images:

POST /api/v1/live/scopes/:slug/regenerate
Headers: X-API-Key: bnt_project_key
Body: { "imageId": "uuid" } // Optional: regenerate specific image
Response: {
  "regenerated": 1,
  "images": [...]
}

Delete scope:

DELETE /api/v1/live/scopes/:slug
Headers: X-API-Key: bnt_project_key

Deletion behavior: Deletes all images in this scope (follows standard image deletion with alias protection).

8.6 Security & Rate Limiting

Rate Limiting by IP:

Aggressive limits for live URLs (e.g., 10 new generations per hour per IP)
Separate from API key limits
Cache hits do NOT count toward limit
Only new generations count

Scope Quotas:

Maximum N unique prompts per scope (newGenerationsLimit)
After limit reached → return existing images, do not generate new
Regeneration does not count toward limit

8.7 Caching Strategy

Cache Key:

cacheKey = hash(projectId + scope + prompt + aspectRatio + otherParams)

Cache Invalidation:

Manual: via API endpoint regenerate
Automatic: never (images cached forever unless explicitly regenerated)

Scope Naming: scope (confirmed)

URL Encoding:

Prompt in query string: URL-encoded or underscores for spaces
Both formats supported: prompt=beautiful%20sunset and prompt=beautiful_sunset
Scope in path: alphanumeric + hyphens + underscores

8.8 Error Handling

Detailed errors for live URLs:

Invalid scope format → 400 "Invalid scope format. Use alphanumeric characters, hyphens, and underscores"
New scope creation disabled → 403 "Creating new live scopes is disabled for this project"
Generation limit exceeded → 429 "Scope generation limit exceeded. Maximum N generations per scope"
Generation fails → 500 with retry logic
Rate limit by IP exceeded → 429 "Rate limit exceeded. Try again in X seconds" with Retry-After header

9. Generation Modification

9.1 Update Generation Endpoint

New endpoint:

PUT /api/v1/generations/:id

Modifiable Fields:

prompt - change prompt
aspectRatio - change aspect ratio
flowId - change/remove/add flow association
meta - update metadata

Behavior:

Case 1: Non-generative parameters (flowId, meta)

Simply update fields in DB
Do NOT regenerate image

Case 2: Generative parameters (prompt, aspectRatio)

Update fields in DB
Automatically trigger regeneration
Update existing image (same imageId, path, URL)

9.2 FlowId Management

FlowId handling:

flowId: null → detach from flow
flowId: "new-uuid" → attach to different flow
- If flow doesn't exist → create new flow eagerly (with this flowId)
- If flow exists → add generation to existing flow
flowId: undefined → do not change current value

Use Case - "Detach from Flow":

Set flowId: null to detach generation from flow
Output image is preserved (if has alias)
Useful before deleting flow to protect important generations

9.3 Validation Rules

Use existing validation logic from generation creation:

Prompt validation (existing rules)
AspectRatio validation (existing rules)
FlowId validation:
- If provided (not null): must be valid UUID format
- Flow does NOT need to exist (will be created eagerly if missing)
- Allow null explicitly (for detachment)

9.4 Response Format

{
  "success": true,
  "data": {
    "id": "gen-uuid",
    "prompt": "updated prompt",
    "aspectRatio": "16:9",
    "flowId": null,
    "status": "processing",  // If regeneration triggered
    "regenerated": true,      // Flag indicating regeneration started
    "outputImage": { ... }    // Current image (updates when regeneration completes)
  }
}

10. Response Format Consistency

10.1 FlowId in Responses

Rule for flowId in generation and upload responses:

If request has flowId: undefined (not provided):

Generate new flowId
Return in response: "flowId": "new-uuid"

If request has flowId: null (explicitly null):

Do NOT generate flowId
Flow is definitely not needed
Return in response: "flowId": null

If request has flowId: "uuid" (specific value):

Use provided flowId
Return in response: "flowId": "uuid"