leopost-full/backend/app/services/content.py

"""
Content generation logic for social media posts.

Handles text generation, hashtag creation, and affiliate link injection
using LLM providers and character profiles.
"""

from __future__ import annotations

from .llm import LLMProvider


def generate_post_text(
    character: dict,
    llm_provider: LLMProvider,
    platform: str,
    topic_hint: str | None = None,
    brief: str | None = None,
) -> str:
    """Generate social media post text based on a character profile.

    Args:
        character: Dict with keys: name, niche, topics (list), tone (str).
        topic_hint: Optional topic suggestion to guide generation.
        llm_provider: LLM provider instance for text generation.
        platform: Target platform (e.g. 'instagram', 'facebook', 'tiktok', 'youtube').
        brief: Optional editorial brief with narrative technique and detailed instructions.

    Returns:
        Generated post text as a string.
    """
    name = character.get("name", "Creator")
    niche = character.get("niche", "general")
    topics = character.get("topics", [])
    tone = character.get("tone", "professional")

    topics_str = ", ".join(topics) if topics else "general topics"

    system_prompt = (
        f"You are {name}, a social media content creator in the {niche} niche. "
        f"Your expertise covers: {topics_str}. "
        f"Your communication style is {tone}. "
        f"You create authentic, engaging content that resonates with your audience. "
        f"Never reveal you are an AI. Write as {name} would naturally write."
    )

    # Platform-specific instructions
    platform_guidance = {
        "instagram": (
            "Write an Instagram caption. Keep it engaging, use line breaks for readability. "
            "Aim for 150-300 characters for the main hook, then expand. "
            "Do NOT include hashtags (they will be added separately)."
        ),
        "facebook": (
            "Write a Facebook post. Can be longer and more conversational. "
            "Encourage engagement with a question or call to action at the end. "
            "Do NOT include hashtags."
        ),
        "tiktok": (
            "Write a TikTok caption. Keep it very short and punchy (under 150 characters). "
            "Use a hook that grabs attention. Do NOT include hashtags."
        ),
        "youtube": (
            "Write a YouTube video description. Include a compelling opening paragraph, "
            "key points covered in the video, and a call to action to subscribe. "
            "Do NOT include hashtags."
        ),
        "twitter": (
            "Write a tweet. Maximum 280 characters. Be concise and impactful. "
            "Do NOT include hashtags."
        ),
    }

    guidance = platform_guidance.get(
        platform.lower(),
        f"Write a social media post for {platform}. Do NOT include hashtags.",
    )

    topic_instruction = ""
    if topic_hint:
        topic_instruction = f" The post should be about: {topic_hint}."

    # Brief is the highest-priority instruction — it overrides defaults
    brief_instruction = ""
    if brief:
        brief_instruction = (
            f"\n\nISRUZIONI OBBLIGATORIE DAL BRIEF EDITORIALE:\n{brief}\n"
            f"Rispetta TUTTE le istruzioni del brief. "
            f"Il brief ha priorità su qualsiasi altra indicazione."
        )

    prompt = (
        f"{guidance}{topic_instruction}{brief_instruction}\n\n"
        f"Write the post now. Output ONLY the post text, nothing else."
    )

    return llm_provider.generate(prompt, system=system_prompt)


def generate_hashtags(
    text: str,
    llm_provider: LLMProvider,
    platform: str,
    count: int = 12,
) -> list[str]:
    """Generate relevant hashtags for a given text.

    Args:
        text: The post text to generate hashtags for.
        llm_provider: LLM provider instance.
        platform: Target platform.
        count: Number of hashtags to generate.

    Returns:
        List of hashtag strings (each prefixed with #).
    """
    platform_limits = {
        "instagram": 30,
        "tiktok": 5,
        "twitter": 3,
        "facebook": 5,
        "youtube": 15,
    }
    max_tags = min(count, platform_limits.get(platform.lower(), count))

    system_prompt = (
        "You are a social media hashtag strategist. You generate relevant, "
        "effective hashtags that maximize reach and engagement."
    )

    prompt = (
        f"Generate exactly {max_tags} hashtags for the following {platform} post.\n\n"
        f"Post text:\n{text}\n\n"
        f"Rules:\n"
        f"- Mix popular (high reach) and niche (targeted) hashtags\n"
        f"- Each hashtag must start with #\n"
        f"- No spaces within hashtags, use CamelCase for multi-word\n"
        f"- Output ONLY the hashtags, one per line, nothing else"
    )

    result = llm_provider.generate(prompt, system=system_prompt)

    # Parse hashtags from the response
    hashtags: list[str] = []
    for line in result.strip().splitlines():
        tag = line.strip()
        if not tag:
            continue
        # Ensure it starts with #
        if not tag.startswith("#"):
            tag = f"#{tag}"
        # Remove any trailing punctuation or spaces
        tag = tag.split()[0]  # Take only the first word if extra text
        hashtags.append(tag)

    return hashtags[:max_tags]


def inject_affiliate_links(
    text: str,
    affiliate_links: list[dict],
    topics: list[str],
) -> tuple[str, list[dict]]:
    """Find relevant affiliate links and append them to the post text.

    Matches affiliate links based on topic overlap. Links whose keywords
    overlap with the provided topics are appended naturally at the end.

    Args:
        text: Original post text.
        affiliate_links: List of dicts, each with keys:
            - url (str): The affiliate URL
            - label (str): Display text for the link
            - keywords (list[str]): Topic keywords this link is relevant for
        topics: Current post topics to match against.

    Returns:
        Tuple of (modified_text, links_used) where links_used is the list
        of affiliate link dicts that were injected.
    """
    if not affiliate_links or not topics:
        return text, []

    # Normalize topics to lowercase for matching
    topics_lower = {t.lower() for t in topics}

    # Score each link by keyword overlap
    scored_links: list[tuple[int, dict]] = []
    for link in affiliate_links:
        keywords = link.get("keywords", [])
        keywords_lower = {k.lower() for k in keywords}
        overlap = len(topics_lower & keywords_lower)
        if overlap > 0:
            scored_links.append((overlap, link))

    if not scored_links:
        return text, []

    # Sort by relevance (most overlap first), take top 2
    scored_links.sort(key=lambda x: x[0], reverse=True)
    top_links = [link for _, link in scored_links[:2]]

    # Build the links section
    links_section_parts: list[str] = []
    for link in top_links:
        label = link.get("label", "Check this out")
        url = link.get("url", "")
        links_section_parts.append(f"{label}: {url}")

    links_text = "\n".join(links_section_parts)
    modified_text = f"{text}\n\n{links_text}"

    return modified_text, top_links