personal-ai-assistant/backend/app/services/ai_service.py

import json
import uuid
from collections.abc import AsyncGenerator

from anthropic import AsyncAnthropic
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession

from app.config import settings
from app.models.chat import Chat
from app.models.message import Message
from app.models.skill import Skill
from app.services.context_service import DEFAULT_SYSTEM_PROMPT, get_primary_context, get_personal_context
from app.services.chat_service import get_chat, save_message
from app.services.memory_service import get_critical_memories, create_memory, get_user_memories
from app.services.document_service import search_documents
from app.services.notification_service import create_notification
from app.services.ws_manager import manager

client = AsyncAnthropic(api_key=settings.ANTHROPIC_API_KEY)

# --- AI Tool Definitions ---

AI_TOOLS = [
    {
        "name": "save_memory",
        "description": "Save important health information to the user's memory. Use this when the user shares critical health data like conditions, medications, allergies, or important health facts.",
        "input_schema": {
            "type": "object",
            "properties": {
                "category": {
                    "type": "string",
                    "enum": ["condition", "medication", "allergy", "vital", "document_summary", "other"],
                    "description": "Category of the memory entry",
                },
                "title": {"type": "string", "description": "Short title for the memory entry"},
                "content": {"type": "string", "description": "Detailed content of the memory entry"},
                "importance": {
                    "type": "string",
                    "enum": ["critical", "high", "medium", "low"],
                    "description": "Importance level",
                },
            },
            "required": ["category", "title", "content", "importance"],
        },
    },
    {
        "name": "search_documents",
        "description": "Search the user's uploaded health documents for relevant information. Use this when you need to find specific health records, lab results, or consultation notes.",
        "input_schema": {
            "type": "object",
            "properties": {
                "query": {"type": "string", "description": "Search query to find relevant documents"},
            },
            "required": ["query"],
        },
    },
    {
        "name": "get_memory",
        "description": "Retrieve the user's stored health memories filtered by category. Use this to recall previously saved health information.",
        "input_schema": {
            "type": "object",
            "properties": {
                "category": {
                    "type": "string",
                    "enum": ["condition", "medication", "allergy", "vital", "document_summary", "other"],
                    "description": "Optional category filter. Omit to get all memories.",
                },
            },
            "required": [],
        },
    },
    {
        "name": "schedule_notification",
        "description": "Schedule a notification or reminder for the user. Can be immediate or scheduled for a future time.",
        "input_schema": {
            "type": "object",
            "properties": {
                "title": {"type": "string", "description": "Notification title"},
                "body": {"type": "string", "description": "Notification body text"},
                "scheduled_at": {
                    "type": "string",
                    "description": "ISO 8601 datetime for scheduled delivery. Omit for immediate.",
                },
                "type": {
                    "type": "string",
                    "enum": ["reminder", "alert", "info"],
                    "description": "Notification type",
                    "default": "reminder",
                },
            },
            "required": ["title", "body"],
        },
    },
]


async def _execute_tool(
    db: AsyncSession, user_id: uuid.UUID, tool_name: str, tool_input: dict
) -> str:
    """Execute an AI tool and return the result as a string."""
    if tool_name == "save_memory":
        entry = await create_memory(
            db, user_id,
            category=tool_input["category"],
            title=tool_input["title"],
            content=tool_input["content"],
            importance=tool_input["importance"],
        )
        await db.commit()
        return json.dumps({"status": "saved", "id": str(entry.id), "title": entry.title})

    elif tool_name == "search_documents":
        docs = await search_documents(db, user_id, tool_input["query"], limit=5)
        results = []
        for doc in docs:
            excerpt = (doc.extracted_text or "")[:1000]
            results.append({
                "filename": doc.original_filename,
                "doc_type": doc.doc_type,
                "excerpt": excerpt,
            })
        return json.dumps({"results": results, "count": len(results)})

    elif tool_name == "get_memory":
        category = tool_input.get("category")
        entries = await get_user_memories(db, user_id, category=category, is_active=True)
        items = [{"category": e.category, "title": e.title, "content": e.content, "importance": e.importance} for e in entries]
        return json.dumps({"entries": items, "count": len(items)})

    elif tool_name == "schedule_notification":
        from datetime import datetime
        scheduled_at = None
        if tool_input.get("scheduled_at"):
            scheduled_at = datetime.fromisoformat(tool_input["scheduled_at"])

        notif = await create_notification(
            db, user_id,
            title=tool_input["title"],
            body=tool_input["body"],
            type=tool_input.get("type", "reminder"),
            scheduled_at=scheduled_at,
        )
        await db.commit()

        # Push immediately if not scheduled
        if not scheduled_at:
            await manager.send_to_user(user_id, {
                "type": "new_notification",
                "notification": {
                    "id": str(notif.id),
                    "title": notif.title,
                    "body": notif.body,
                    "type": notif.type,
                    "created_at": notif.created_at.isoformat(),
                },
            })

        return json.dumps({
            "status": "scheduled" if scheduled_at else "sent",
            "id": str(notif.id),
            "title": notif.title,
        })

    return json.dumps({"error": f"Unknown tool: {tool_name}"})


async def assemble_context(
    db: AsyncSession, chat_id: uuid.UUID, user_id: uuid.UUID, user_message: str
) -> tuple[str, list[dict]]:
    """Assemble system prompt and messages for Claude API."""
    system_parts = []

    # 1. Primary context
    ctx = await get_primary_context(db)
    system_parts.append(ctx.content if ctx and ctx.content.strip() else DEFAULT_SYSTEM_PROMPT)

    # 2. Personal context
    personal_ctx = await get_personal_context(db, user_id)
    if personal_ctx and personal_ctx.content.strip():
        system_parts.append(f"---\nUser Context:\n{personal_ctx.content}")

    # 3. Active skill system prompt
    chat = await get_chat(db, chat_id, user_id)
    if chat.skill_id:
        result = await db.execute(select(Skill).where(Skill.id == chat.skill_id))
        skill = result.scalar_one_or_none()
        if skill and skill.is_active:
            system_parts.append(f"---\nSpecialist Role ({skill.name}):\n{skill.system_prompt}")

    # 4. Critical memory entries
    memories = await get_critical_memories(db, user_id)
    if memories:
        memory_lines = [f"- [{m.category}] {m.title}: {m.content}" for m in memories]
        system_parts.append(f"---\nUser Health Profile:\n" + "\n".join(memory_lines))

    # 5. Relevant document excerpts (based on user message keywords)
    if user_message.strip():
        docs = await search_documents(db, user_id, user_message, limit=3)
        if docs:
            doc_lines = []
            for d in docs:
                excerpt = (d.extracted_text or "")[:1500]
                doc_lines.append(f"[{d.original_filename} ({d.doc_type})]\n{excerpt}")
            system_parts.append(f"---\nRelevant Document Excerpts:\n" + "\n\n".join(doc_lines))

    system_prompt = "\n\n".join(system_parts)

    # 6. Conversation history
    result = await db.execute(
        select(Message)
        .where(Message.chat_id == chat_id, Message.role.in_(["user", "assistant"]))
        .order_by(Message.created_at.asc())
    )
    history = result.scalars().all()
    messages = [{"role": msg.role, "content": msg.content} for msg in history]

    # 7. Current user message
    messages.append({"role": "user", "content": user_message})

    return system_prompt, messages


def _sse_event(event: str, data: dict) -> str:
    return f"event: {event}\ndata: {json.dumps(data)}\n\n"


async def stream_ai_response(
    db: AsyncSession, chat_id: uuid.UUID, user_id: uuid.UUID, user_message: str
) -> AsyncGenerator[str, None]:
    """Stream AI response as SSE events, with tool use support."""
    # Verify ownership
    chat = await get_chat(db, chat_id, user_id)

    # Save user message
    await save_message(db, chat_id, "user", user_message)
    await db.commit()

    try:
        # Assemble context
        system_prompt, messages = await assemble_context(db, chat_id, user_id, user_message)

        assistant_msg_id = str(uuid.uuid4())
        yield _sse_event("message_start", {"message_id": assistant_msg_id})

        # Tool use loop
        full_content = ""
        max_tool_rounds = 5

        for _ in range(max_tool_rounds):
            response = await client.messages.create(
                model=settings.CLAUDE_MODEL,
                max_tokens=4096,
                system=system_prompt,
                messages=messages,
                tools=AI_TOOLS,
            )

            # Process content blocks
            tool_use_blocks = []
            for block in response.content:
                if block.type == "text":
                    full_content += block.text
                    yield _sse_event("content_delta", {"delta": block.text})
                elif block.type == "tool_use":
                    tool_use_blocks.append(block)
                    yield _sse_event("tool_use", {"tool": block.name, "input": block.input})

            # If no tool use, we're done
            if response.stop_reason != "tool_use" or not tool_use_blocks:
                break

            # Execute tools and continue conversation
            messages.append({"role": "assistant", "content": response.content})
            tool_results = []
            for tool_block in tool_use_blocks:
                result = await _execute_tool(db, user_id, tool_block.name, tool_block.input)
                tool_results.append({
                    "type": "tool_result",
                    "tool_use_id": tool_block.id,
                    "content": result,
                })
                yield _sse_event("tool_result", {"tool": tool_block.name, "result": result})
            messages.append({"role": "user", "content": tool_results})

        metadata = {
            "model": response.model,
            "input_tokens": response.usage.input_tokens,
            "output_tokens": response.usage.output_tokens,
        }

        # Save assistant message
        saved_msg = await save_message(db, chat_id, "assistant", full_content, metadata)
        await db.commit()

        # Update chat title if first exchange
        result = await db.execute(
            select(Message).where(Message.chat_id == chat_id, Message.role == "assistant")
        )
        assistant_count = len(result.scalars().all())
        if assistant_count == 1 and chat.title == "New Chat":
            title = full_content[:50].split("\n")[0].strip()
            if len(title) > 40:
                title = title[:40] + "..."
            chat.title = title
            await db.commit()

        yield _sse_event("message_end", {
            "message_id": str(saved_msg.id),
            "metadata": metadata,
        })

    except Exception as e:
        yield _sse_event("error", {"detail": str(e)})