src/backend/api/endpoints.py

"""API endpoints registered on the Gradio Server.

Every endpoint uses ``@app.api(...)`` so it runs through Gradio's queue
(concurrency-managed, ZeroGPU-safe) and is reachable from ``gradio_client``
and the frontend JS client.

All endpoints are placeholders: they call the placeholder backend functions
and return canned data. No real parsing/embedding/RAG is wired yet.
"""

from __future__ import annotations

from gradio import Server
from gradio.data_classes import FileData

from ..indexing.store import BACKEND
from ..parsing.directive import parse_directive
from ..parsing.lonstatistik import parse_lonstatistik
from ..rag import answer
from ..schemas import (
    ChatMessage,
    Chunk,
    IndexStatus,
    ParseResult,
    Profile,
)


def register(app: Server) -> None:
    """Register all backend endpoints on the given Gradio Server."""

    @app.api(name="parse_directive")
    def parse_directive_endpoint(lang: str = "en") -> list[dict]:
        """Parse Directive 2023/970 into article-level chunks (placeholder)."""
        chunks: list[Chunk] = parse_directive(lang=lang)
        return [c.model_dump() for c in chunks]

    @app.api(name="parse_lonstatistik")
    def parse_lonstatistik_endpoint(
        file: FileData, source: str = "unknown"
    ) -> dict:
        """OCR/table-parse an uploaded lønstatistik PDF (placeholder)."""
        result: ParseResult = parse_lonstatistik(
            file_path=file["path"] if file else None, source=source
        )
        return result.model_dump()

    @app.api(name="index_documents")
    def index_documents_endpoint(source: str = "directive") -> dict:
        """Report the loaded vector index status.

        The index is built offline (``scripts/build_index.py``); this endpoint
        reports how many chunks are loaded, not a live (re)build.
        """
        from ..indexing.store import VectorStore

        store = VectorStore()
        persisted = store.load()
        return IndexStatus(
            source=source,
            indexed_chunks=len(store),
            backend=BACKEND,
            persisted=persisted,
        ).model_dump()

    @app.api(name="warmup")
    def warmup_endpoint() -> dict:
        """Pre-load the chat models (embedder + LLM) on GPU.

        Fired by the frontend when onboarding starts, so the model is ready by
        the time the user reaches the chat. Best-effort: returns a status dict and
        never errors the UI.
        """
        from ..rag import warmup

        return warmup()

    @app.api(name="chat")
    def chat_endpoint(messages: list[dict], lang: str = "en", context: str = "") -> dict:
        """Chat with the RAG assistant: retrieve grounded chunks, cite sources.

        Returns the full ``{reply, citations, disclaimer}`` in one response.
        ``context`` is the user's compact salary-dashboard summary, injected so
        the answer can be specific to their situation.
        """
        msgs = [ChatMessage(**m) for m in messages]
        return answer(msgs, lang=lang, context=context or None).model_dump()

    @app.api(name="dashboard")
    def dashboard_endpoint(profile: dict) -> dict:
        """Build the profile-driven salary dashboard (structured IDA retrieval).

        Takes the onboarding answers (a ``Profile``) and returns the fused
        ``Dashboard`` — per-axis comparison cells, reliability-gated, projected
        to 2026, with the user's percentile and gap. CPU-only (no embedding).
        """
        from ..fusion import build_dashboard

        return build_dashboard(Profile(**profile)).model_dump()