backend/core/style_profile.py

"""Per-tenant writing style profile from the reference corpus."""

from __future__ import annotations

import json
import logging
import threading
from dataclasses import asdict, dataclass, field
from pathlib import Path

from backend.config import settings
from backend.core.rag_store import TIER_REFERENCE, get_rag_store
from backend.llm import openai_client
from backend.utils import tenant_store

logger = logging.getLogger(__name__)

_lock = threading.RLock()


@dataclass
class StyleProfile:
    tone: str = "formal"
    formality_level: str = "professional"
    avg_sentence_complexity: str = "moderate"
    vocabulary_level: str = "technical"
    common_phrases: list[str] = field(default_factory=lambda: [
        "The property",
        "It is noted that",
        "In respect of",
        "At the time of inspection",
        "We would recommend",
    ])
    structural_patterns: list[str] = field(default_factory=lambda: [
        "Topic sentence followed by supporting detail and recommendation",
        "Passive voice preferred for observations; active for recommendations",
    ])
    writing_style_summary: str = (
        "Formal UK RICS survey prose with measured technical vocabulary and professional hedging."
    )
    example_paragraphs: list[str] = field(default_factory=list)

    def to_payload(self) -> dict:
        return asdict(self)


DEFAULT_PROFILE = StyleProfile()


def _cache_path(tenant_id: str) -> Path:
    return tenant_store.tenant_root(tenant_id) / "style_profile.json"


def _sample_reference_text(tenant_id: str, max_chars: int = 4800) -> str:
    store = get_rag_store()
    parts: list[str] = []
    used = 0
    for text in store.sample_chunk_texts(tenant_id, TIER_REFERENCE, limit=60):
        if used + len(text) + 2 > max_chars:
            break
        parts.append(text)
        used += len(text) + 2
    return "\n\n".join(parts)


def _heuristic_profile(sample: str) -> StyleProfile:
    if not sample.strip():
        return DEFAULT_PROFILE
    lower = sample.lower()
    tone = "formal"
    if "recommend" in lower or "advise" in lower:
        tone = "semi-formal"
    vocab = "technical"
    if sample.count("£") + sample.count("mm") + sample.count("dpc") >= 3:
        vocab = "specialist"
    sentences = [s.strip() for s in sample.replace("\n", " ").split(".") if len(s.strip()) > 20]
    avg_len = sum(len(s.split()) for s in sentences) / max(len(sentences), 1)
    complexity = "moderate"
    if avg_len > 22:
        complexity = "complex"
    elif avg_len < 14:
        complexity = "simple"
    examples = []
    for block in sample.split("\n\n"):
        words = block.split()
        if 40 <= len(words) <= 120:
            examples.append(block.strip())
        if len(examples) >= 2:
            break
    return StyleProfile(
        tone=tone,
        avg_sentence_complexity=complexity,
        vocabulary_level=vocab,
        writing_style_summary=(
            "Derived from uploaded past reports: measured UK surveyor voice with "
            f"{complexity} sentences and {vocab} vocabulary."
        ),
        example_paragraphs=examples,
    )


def _llm_profile(sample: str) -> StyleProfile | None:
    if not openai_client.is_available() or len(sample) < 200:
        return None
    prompt = f"""Analyse this UK RICS survey sample and return JSON with keys:
tone, formality_level, avg_sentence_complexity, vocabulary_level,
common_phrases (array), structural_patterns (array), writing_style_summary,
example_paragraphs (array of verbatim short extracts).

SAMPLE:
{sample[:4000]}"""
    try:
        raw = openai_client.chat_json(
            [
                {
                    "role": "system",
                    "content": "You analyse UK RICS report writing style. Output JSON only.",
                },
                {"role": "user", "content": prompt},
            ],
            model=settings.mapping_model,
            max_tokens=800,
        )
        return StyleProfile(
            tone=str(raw.get("tone") or "formal"),
            formality_level=str(raw.get("formality_level") or "professional"),
            avg_sentence_complexity=str(raw.get("avg_sentence_complexity") or "moderate"),
            vocabulary_level=str(raw.get("vocabulary_level") or "technical"),
            common_phrases=list(raw.get("common_phrases") or DEFAULT_PROFILE.common_phrases),
            structural_patterns=list(raw.get("structural_patterns") or DEFAULT_PROFILE.structural_patterns),
            writing_style_summary=str(
                raw.get("writing_style_summary") or DEFAULT_PROFILE.writing_style_summary
            ),
            example_paragraphs=list(raw.get("example_paragraphs") or []),
        )
    except Exception as exc:  # noqa: BLE001
        logger.warning("LLM style profile failed (%s) — using heuristics", exc)
        return None


def get_style_profile(tenant_id: str, *, force_refresh: bool = False) -> StyleProfile:
    with _lock:
        path = _cache_path(tenant_id)
        if not force_refresh and path.is_file():
            try:
                data = json.loads(path.read_text(encoding="utf-8"))
                return StyleProfile(**data)
            except Exception:  # noqa: BLE001
                pass

        sample = _sample_reference_text(tenant_id)
        profile = _llm_profile(sample) or _heuristic_profile(sample)
        path.write_text(json.dumps(profile.to_payload(), indent=2), encoding="utf-8")
        return profile


def invalidate_style_profile(tenant_id: str) -> None:
    path = _cache_path(tenant_id)
    if path.is_file():
        path.unlink(missing_ok=True)