Content Studio

The Content Studio lets you upload your existing writing, analyze your style, and generate new content that matches your voice.

Writing Style Analysis

Upload content (paste text or import from URL) to build your style profile.

How It Works

1. Upload content via POST /api/content or POST /api/content/import
2. Analyze style via POST /api/content/{id}/analyze
3. The LLM produces a JSON style profile:

{
  "tone": "conversational",
  "avg_sentence_length": "medium",
  "vocabulary_complexity": "moderate",
  "structural_patterns": ["uses headers", "code examples", "bullet points"],
  "use_of_examples": true,
  "use_of_analogies": true,
  "first_person": true,
  "typical_intro_style": "Opens with a relatable problem statement",
  "typical_conclusion_style": "Summarizes key takeaways with action items"
}

Composite Style Profile

When you have multiple uploads, GET /api/content/style-profile merges them into a single composite profile using the LLM. The composite profile is used during content generation.

Content Generation

Output Formats

Format	Description
blog_post	Comprehensive post with intro, headings, code examples, conclusion
youtube_script	Video script with hook, sections, [VISUAL] cues, CTA. 8-12 min target
linkedin_post	Under 3000 chars. Hook, short paragraphs, question/CTA, hashtags
newsletter	Conversational, scannable with headers and bullets, key takeaways

Generation Pipeline

1. Load the target topic and up to 5 related topics from the same group
2. Fetch composite style profile (from your uploaded content)
3. RAG retrieval — find up to 3 similar pieces of your uploaded content via pgvector cosine similarity
4. Build system prompt: format instructions + style profile + RAG examples
5. Call LLM with GENERATION task type
6. Store result in generated_content table

sequenceDiagram
    participant User
    participant API
    participant Gen as ContentGenerationService
    participant Style as StyleAnalysisService
    participant RAG as UserContentService
    participant LLM

    User->>API: POST /api/generate {topicId, format}
    API->>Gen: generate(topicId, format)
    Gen->>Style: getCompositeStyleProfile()
    Gen->>RAG: findSimilarContent(topic, 3)
    Gen->>LLM: complete(GENERATION, system, user)
    LLM-->>Gen: generated text
    Gen-->>API: {id, generated_text}

Regeneration with Feedback

If the output isn't right, regenerate with feedback:

POST /api/generate/{id}/regenerate
{ "feedback": "Make it more technical, add code examples" }

The system uses the original content + your feedback to produce a revised version. Maximum 5 regenerations per topic+format.

Export

Export generated content as Markdown:

GET /api/generate/{id}/export?format=markdown

RAG Workflow

The RAG (Retrieval-Augmented Generation) workflow enhances content generation by referencing your previous writing:

1. Embedding — When you upload content, it's embedded as a 1536-dim vector via the LLM embedding model
2. Retrieval — During generation, the topic text is embedded and compared to your content library via pgvector cosine similarity
3. Context — The top 3 most similar pieces of your content are included in the LLM prompt as style/voice references

Note

RAG requires pgvector to be enabled. Without it, generation still works but without style-matched references.

Content Types

When uploading, specify a content type (informational, used for organization):

Type	Description
blog_post	Blog articles
transcript	Video/podcast transcripts
newsletter	Newsletter issues