MobilityOps/backend/app/services/knowledge/demo.py

from __future__ import annotations

import math
import re
from dataclasses import dataclass, field
from pathlib import Path

from app.core.config import get_settings
from app.services.knowledge import GroundedAnswer, KnowledgeHealth, SourceCard
from app.services.knowledge.procedures import parse_frontmatter

SUPPORTED_LANGUAGES = ("nl-BE", "en-GB", "fr-BE")
DEFAULT_LANGUAGE = "en-GB"

STOPWORDS_BY_LANGUAGE: dict[str, set[str]] = {
    "en-GB": {
        "a", "an", "the", "is", "are", "was", "were", "be", "been", "being",
        "to", "of", "in", "on", "at", "for", "and", "or", "but", "if", "then",
        "do", "does", "did", "must", "may", "can", "could", "should", "would",
        "i", "you", "it", "we", "they", "my", "your", "what", "when", "how",
        "with", "without", "this", "that", "these", "those", "not", "no",
    },
    "nl-BE": {
        "een", "de", "het", "is", "zijn", "was", "waren", "worden", "wordt",
        "van", "in", "op", "voor", "en", "of", "maar", "als", "dan",
        "moet", "mag", "kan", "kunnen", "zou", "zouden",
        "ik", "jij", "u", "we", "wij", "zij", "mijn", "jouw", "wat", "wanneer", "hoe",
        "met", "zonder", "dit", "dat", "deze", "die", "niet", "geen",
    },
    "fr-BE": {
        "un", "une", "le", "la", "les", "des", "est", "sont", "était", "être",
        "de", "du", "en", "sur", "pour", "et", "ou", "mais", "si", "alors",
        "doit", "peut", "peuvent", "pourrait", "devrait",
        "je", "tu", "vous", "il", "elle", "nous", "ils", "mon", "votre", "quoi", "quand", "comment",
        "avec", "sans", "ce", "cette", "ces", "cela", "pas", "non",
    },
}

# Includes the Latin-1 accented-letter range (à-ö, ø-ÿ) so French/Dutch words with
# diacritics (véhicule, réservation, geëscaleerd) tokenize as one word instead of
# splitting apart at the accented character -- a plain [a-z0-9]+ pattern silently
# drops every accent and fragments the word either side of it.
_WORD_RE = re.compile(r"[a-zà-öø-ÿ0-9]+")


def _stem(word: str) -> str:
    # Deterministic, intentionally crude suffix stripping — good enough to match "returns"
    # with "return" or "damaged" with "damage" without pulling in a stemming dependency.
    for suffix in ("ing", "edly", "ed", "es", "s"):
        if len(word) > len(suffix) + 2 and word.endswith(suffix):
            return word[: -len(suffix)]
    return word


def _tokenize(text: str, language: str) -> set[str]:
    stopwords = STOPWORDS_BY_LANGUAGE.get(language, STOPWORDS_BY_LANGUAGE[DEFAULT_LANGUAGE])
    words = _WORD_RE.findall(text.lower())
    return {_stem(w) for w in words if w not in stopwords and len(w) > 2}


@dataclass
class Document:
    document_id: str
    title: str
    version: str
    title_tokens: set[str] = field(default_factory=set)


@dataclass
class ScoredSection:
    document: Document
    heading: str
    text: str
    heading_tokens: set[str]
    body_tokens: set[str]


def _split_sections(body: str) -> list[tuple[str, str]]:
    sections: list[tuple[str, str]] = []
    current_heading = "Overview"
    current_lines: list[str] = []
    for line in body.splitlines():
        if line.startswith("## "):
            if current_lines:
                sections.append((current_heading, "\n".join(current_lines).strip()))
            current_heading = line[3:].strip()
            current_lines = []
        elif line.startswith("# "):
            continue
        else:
            current_lines.append(line)
    if current_lines:
        sections.append((current_heading, "\n".join(current_lines).strip()))
    return sections


def _load_sections(procedures_dir: Path, language: str) -> list[ScoredSection]:
    sections: list[ScoredSection] = []
    for path in sorted(procedures_dir.glob("*.md")):
        raw = path.read_text(encoding="utf-8")
        meta, body = parse_frontmatter(raw)
        title = meta.get("title", path.stem)
        doc = Document(
            document_id=meta.get("document_id", path.stem),
            title=title,
            version=meta.get("version", "1.0"),
            title_tokens=_tokenize(title, language),
        )
        for heading, text in _split_sections(body):
            sections.append(
                ScoredSection(
                    document=doc,
                    heading=heading,
                    text=text,
                    heading_tokens=_tokenize(heading, language),
                    body_tokens=_tokenize(text, language),
                )
            )
    return sections


_NO_MATCH_TEXT = {
    "en-GB": "No matching procedure was found for this question.",
    "nl-BE": "Er werd geen passende procedure gevonden voor deze vraag.",
    "fr-BE": "Aucune procédure correspondante n'a été trouvée pour cette question.",
}
_LOW_CONFIDENCE_TEXT = {
    "en-GB": (
        "The available procedures do not clearly answer this question. "
        "The closest matches are included below for review."
    ),
    "nl-BE": (
        "De beschikbare procedures beantwoorden deze vraag niet duidelijk. "
        "De dichtstbijzijnde overeenkomsten staan hieronder ter beoordeling."
    ),
    "fr-BE": (
        "Les procédures disponibles ne répondent pas clairement à cette question. "
        "Les correspondances les plus proches sont indiquées ci-dessous pour examen."
    ),
}
_LEAD_ANSWER_TEMPLATE = {
    "en-GB": 'Per "{title}" (v{version}), section "{heading}": {excerpt}',
    "nl-BE": 'Volgens "{title}" (v{version}), sectie "{heading}": {excerpt}',
    "fr-BE": 'Selon « {title} » (v{version}), section « {heading} » : {excerpt}',
}


class DemoKnowledgeProvider:
    """Deterministic extractive retrieval over the local procedure Markdown files.

    Not a generative model: it scores sections with TF-IDF-weighted keyword overlap
    (downweighting terms common across the whole corpus, like "vehicle", in favor of
    distinctive ones, like "damage") and returns real excerpts, never invented text.

    Each supported UI language has its own translated procedure corpus under
    knowledge/procedures/<language>/ -- retrieval searches only within the requested
    language's corpus so citations always link to a same-language document.
    """

    name = "demo"

    def __init__(self) -> None:
        settings = get_settings()
        self._settings = settings
        base_dir = Path(settings.knowledge_dir)
        self._sections_by_language: dict[str, list[ScoredSection]] = {}
        self._idf_by_language: dict[str, dict[str, float]] = {}
        self._document_count_by_language: dict[str, int] = {}
        for language in SUPPORTED_LANGUAGES:
            lang_dir = base_dir / language
            sections = _load_sections(lang_dir, language) if lang_dir.is_dir() else []
            self._sections_by_language[language] = sections
            self._idf_by_language[language] = self._build_idf(sections)
            self._document_count_by_language[language] = len(
                {s.document.document_id for s in sections}
            )

    @staticmethod
    def _build_idf(sections: list[ScoredSection]) -> dict[str, float]:
        n = len(sections) or 1
        doc_freq: dict[str, int] = {}
        for section in sections:
            doc = section.document
            all_tokens = doc.title_tokens | section.heading_tokens | section.body_tokens
            for token in all_tokens:
                doc_freq[token] = doc_freq.get(token, 0) + 1
        return {token: math.log((n + 1) / (df + 1)) + 1 for token, df in doc_freq.items()}

    def _normalize_language(self, language: str | None) -> str:
        if language in SUPPORTED_LANGUAGES:
            return language
        return DEFAULT_LANGUAGE

    def health(self, language: str = DEFAULT_LANGUAGE) -> KnowledgeHealth:
        language = self._normalize_language(language)
        return KnowledgeHealth(
            provider=self.name,
            available=True,
            detail="Deterministic keyword-matching demo provider; no external service.",
            tenant=self._settings.ragcore_tenant,
            workspace=self._settings.ragcore_workspace,
            collection=self._settings.ragcore_collection,
            document_count=self._document_count_by_language[language],
        )

    def _score(
        self, query_tokens: set[str], section: ScoredSection, idf: dict[str, float]
    ) -> float:
        # The section body is the strongest relevance signal -- it's the actual
        # substance a heading or title can only hint at -- so a body match is weighted
        # *above* heading/title matches, not below them. The previous 3x/2x/1x
        # (heading/title/body) ordering let a single generic word in a heading (e.g.
        # "vehicle", present in nearly every section) or a document's own title
        # outrank a section whose body genuinely covers multiple, more distinctive
        # query terms -- confirmed to misrank the brief's exact validation question in
        # every one of the three languages (see docs/fleet-ops-correction/
        # current-gap-audit.md and i18n-inventory.md): nl-BE picked a checkout section
        # over the damage procedure, en-GB and fr-BE picked the return procedure over
        # the damage procedure, purely from heading/title overlap on common words.
        score = 0.0
        for token in query_tokens:
            token_idf = idf.get(token, 0.0)
            if token_idf == 0.0:
                continue
            if token in section.body_tokens:
                score += 3 * token_idf
            elif token in section.heading_tokens:
                score += 2 * token_idf
            elif token in section.document.title_tokens:
                score += 1.5 * token_idf
        return score

    def ask(
        self, question: str, correlation_id: str, language: str = DEFAULT_LANGUAGE
    ) -> GroundedAnswer:
        language = self._normalize_language(language)
        sections = self._sections_by_language[language]
        idf = self._idf_by_language[language]
        query_tokens = _tokenize(question, language)
        scored = [(self._score(query_tokens, section, idf), section) for section in sections]
        scored = [(score, section) for score, section in scored if score > 0]
        scored.sort(key=lambda item: item[0], reverse=True)
        top = scored[:3]

        if not top:
            return GroundedAnswer(
                answer=_NO_MATCH_TEXT[language],
                evidence_state="insufficient",
                sources=[],
                provider=self.name,
                correlation_id=correlation_id,
            )

        sources = [
            SourceCard(
                document_id=section.document.document_id,
                title=section.document.title,
                version=section.document.version,
                section=section.heading,
                excerpt=(section.text[:400] + "…") if len(section.text) > 400 else section.text,
            )
            for _, section in top
        ]

        if top[0][0] < 3:
            return GroundedAnswer(
                answer=_LOW_CONFIDENCE_TEXT[language],
                evidence_state="insufficient",
                sources=sources,
                provider=self.name,
                correlation_id=correlation_id,
            )

        lead_section = top[0][1]
        answer = _LEAD_ANSWER_TEMPLATE[language].format(
            title=lead_section.document.title,
            version=lead_section.document.version,
            heading=lead_section.heading,
            excerpt=lead_section.text.splitlines()[0][:300],
        )
        return GroundedAnswer(
            answer=answer,
            evidence_state="grounded",
            sources=sources,
            provider=self.name,
            correlation_id=correlation_id,
        )