vkus-ai-v4/backend/app/rag/service.py

"""Оркестрация временного локального RAG и сборки ответа."""

from __future__ import annotations

import re
from typing import TYPE_CHECKING

from app.config import settings
from app.data.knowledge_base import get_colleague_cases_for_system
from app.models import ColleagueCase, ResponsePackage, Source
from app.package_builder import (
    build_warning,
    build_why_suggested,
    ensure_escalation_draft,
    enrich_ticket_draft,
)
from app.rag.indexer import LocalRagIndex, SearchHit, detect_system_from_text

if TYPE_CHECKING:
    from app.llm.client import LlmClient

_index: LocalRagIndex | None = None


def get_index() -> LocalRagIndex:
    global _index
    if _index is None:
        _index = LocalRagIndex()
    return _index


def init_local_rag() -> dict[str, object]:
    """Построить индекс при старте backend."""
    index = get_index()
    if not settings.rag_local_enabled or settings.rag_mode != "local":
        return {"enabled": False, "chunks": 0, "mode": settings.rag_mode}

    count = index.build(settings.rag_documents_path, chunk_size=settings.rag_chunk_size)
    return {
        "enabled": True,
        "mode": "local",
        "chunks": count,
        "documents_path": str(settings.rag_documents_path),
    }


def search_documents(query: str) -> list[SearchHit]:
    if not settings.rag_local_enabled or settings.rag_mode != "local":
        return []
    index = get_index()
    if not index.is_ready:
        init_local_rag()
    return index.search(query, top_k=settings.rag_top_k, min_score=settings.rag_min_score)


def score_to_confidence(score: float) -> int:
    """Нормализация cosine similarity в проценты уверенности."""
    return min(95, max(35, int(score * 120)))


def build_package_from_hits(
    query: str,
    hits: list[SearchHit],
    *,
    system: str | None = None,
    conversation: list[dict[str, str]] | None = None,
) -> ResponsePackage | None:
    if not hits:
        return None

    top = hits[0]
    confidence = score_to_confidence(top.score)
    resolved_system = system or top.chunk.system or detect_system_from_text(query) or "Общее"

    recommendation = _extract_recommendation(top.chunk.text)
    steps = _extract_steps(top.chunk.text, hits)
    sources = _unique_sources(hits)
    colleague_cases = [
        ColleagueCase.model_validate(case) for case in get_colleague_cases_for_system(resolved_system)
    ]
    priority = "Средний" if confidence >= settings.confidence_threshold else "Высокий"
    warning = build_warning(recommendation, colleague_cases) if colleague_cases else None
    why_suggested = build_why_suggested(
        system=resolved_system,
        sources=sources,
        confidence=confidence,
        source_title=top.chunk.source_title,
    )

    package = ResponsePackage(
        recommendation=recommendation,
        steps=steps,
        sources=sources,
        colleague_cases=colleague_cases or None,
        warning=warning,
        why_suggested=why_suggested,
        confidence=confidence,
        system=resolved_system,
        ticket_draft=enrich_ticket_draft(
            query,
            resolved_system,
            priority,
            conversation=conversation,
            sources=sources,
            colleague_cases=colleague_cases,
            recommendation=recommendation,
        ),
    )
    return ensure_escalation_draft(package, query, conversation=conversation)


def build_package_with_llm(
    query: str,
    hits: list[SearchHit],
    llm_client: "LlmClient",
    greeting: str,
    *,
    system: str | None = None,
    conversation: list[dict[str, str]] | None = None,
) -> tuple[str, ResponsePackage] | None:
    context = "\n\n---\n\n".join(
        f"[{h.chunk.source_title}]\n{h.chunk.text}" for h in hits[: settings.rag_top_k]
    )
    llm_text = llm_client.generate_with_context(query, context)
    if not llm_text:
        return None

    package = build_package_from_hits(query, hits, system=system, conversation=conversation)
    if package is None:
        return None

    package.recommendation = llm_text
    if package.colleague_cases:
        package.warning = build_warning(llm_text, package.colleague_cases)
    if package.ticket_draft:
        package.ticket_draft = enrich_ticket_draft(
            query,
            package.system or system or "Общее",
            package.ticket_draft.priority,
            conversation=conversation,
            sources=package.sources,
            colleague_cases=package.colleague_cases,
            recommendation=llm_text,
        )
    package = ensure_escalation_draft(package, query, conversation=conversation)
    threshold = settings.confidence_threshold
    intro = (
        f"{greeting}нашёл материалы в инструкциях и подготовил ответ:"
        if package.confidence >= threshold
        else (
            f"{greeting}уверенность ниже {threshold}% — ответ может быть неполным. "
            "Подготовил черновик заявки, его можно сразу отправить в поддержку:"
        )
    )
    return intro, package


def _extract_recommendation(text: str) -> str:
    lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
    if not lines:
        return text[:400]
    # Первая содержательная строка или абзац
    paragraph = lines[0]
    if len(paragraph) < 80 and len(lines) > 1:
        paragraph = f"{paragraph} {lines[1]}"
    return paragraph[:500]


def _extract_steps(text: str, hits: list[SearchHit]) -> list[str]:
    steps: list[str] = []
    combined = "\n".join(h.chunk.text for h in hits[:3])

    for line in combined.splitlines():
        line = line.strip()
        if not line:
            continue
        if re.match(r"^(\d+[\.\)]|[-•*])\s+", line):
            steps.append(re.sub(r"^(\d+[\.\)]|[-•*])\s+", "", line))
        elif "http" in line.lower() or line.lower().startswith("откройте"):
            steps.append(line)
        if len(steps) >= 6:
            break

    if not steps:
        for line in combined.splitlines():
            line = line.strip()
            if len(line) > 20:
                steps.append(line)
            if len(steps) >= 4:
                break

    return steps[:6]


def _unique_sources(hits: list[SearchHit]) -> list[Source]:
    seen: set[str] = set()
    sources: list[Source] = []
    for hit in hits:
        key = hit.chunk.source_path
        if key in seen:
            continue
        seen.add(key)
        sources.append(Source(title=hit.chunk.source_title, url=None))
        if len(sources) >= 4:
            break
    return sources