dorfteich/apps/api/src/search/postgres-search.provider.ts

import { Injectable } from '@nestjs/common';
import {
  SEARCH_HIGHLIGHT_END,
  SEARCH_HIGHLIGHT_START,
  SEARCH_RESULT_LIMIT,
  SearchQuery,
  SearchResultView,
  normalizeForSearch,
} from '@dorfteich/shared';
import { Prisma, User } from '@prisma/client';

import { PermissionService } from '../permissions/permission.service';
import { PrismaService } from '../prisma/prisma.service';
import { SearchProvider } from './search.provider';

/** ts_headline options — one fragment, matches wrapped in the shared sentinels. */
const HEADLINE_OPTIONS =
  `StartSel=${SEARCH_HIGHLIGHT_START}, StopSel=${SEARCH_HIGHLIGHT_END}, ` +
  'MaxFragments=1, MaxWords=30, MinWords=8, ShortWord=0';

/** Raw title/labels/body for one page, before normalization. */
interface IndexSource {
  title: string;
  labels: string | null;
  plain_text: string | null;
}

interface SearchRow {
  pageId: string;
  title: string;
  slug: string;
  pondId: string;
  pondSlug: string;
  pondName: string;
  labelIds: string[];
  snippet: string;
}

/**
 * PostgreSQL full-text search (ADR 0010, issue #49). The weighted `tsvector`
 * (title A, labels B, body C) is stored on `page_content_cache.search_vector`
 * and maintained here (and by the collab persistence hook on content changes);
 * both fold text through {@link normalizeForSearch} before `to_tsvector` so
 * matching is diacritic-insensitive without the `unaccent` extension. Queries
 * are folded the same way; results are filtered to the ponds the user may read.
 */
@Injectable()
export class PostgresSearchProvider extends SearchProvider {
  constructor(
    private readonly prisma: PrismaService,
    private readonly permissions: PermissionService,
  ) {
    super();
  }

  async indexPage(pageId: string): Promise<void> {
    const rows = await this.prisma.$queryRaw<IndexSource[]>`
      SELECT p.title,
             c.plain_text,
             (SELECT string_agg(l.name, ' ')
                FROM page_labels pl JOIN labels l ON l.id = pl.label_id
               WHERE pl.page_id = p.id) AS labels
        FROM pages p
        LEFT JOIN page_content_cache c ON c.page_id = p.id
       WHERE p.id = ${pageId}`;
    const source = rows[0];
    if (!source) return;
    await this.writeVector(pageId, source);
  }

  async removePage(pageId: string): Promise<void> {
    await this.prisma
      .$executeRaw`UPDATE page_content_cache SET search_vector = NULL WHERE page_id = ${pageId}`;
  }

  async removePond(pondId: string): Promise<void> {
    await this.prisma.$executeRaw`
      UPDATE page_content_cache SET search_vector = NULL
       WHERE page_id IN (SELECT id FROM pages WHERE pond_id = ${pondId})`;
  }

  async reindexPond(pondId: string): Promise<number> {
    const sources = await this.prisma.$queryRaw<(IndexSource & { page_id: string })[]>`
      SELECT p.id AS page_id, p.title, c.plain_text,
             (SELECT string_agg(l.name, ' ')
                FROM page_labels pl JOIN labels l ON l.id = pl.label_id
               WHERE pl.page_id = p.id) AS labels
        FROM page_content_cache c
        JOIN pages p ON p.id = c.page_id
       WHERE p.pond_id = ${pondId} AND p.deleted_at IS NULL`;
    for (const source of sources) await this.writeVector(source.page_id, source);
    return sources.length;
  }

  async reindexAll(): Promise<number> {
    // Converge to the issue-#195 invariant: trashed content leaves the
    // index entirely, live content is rebuilt.
    await this.prisma.$executeRaw`
      UPDATE page_content_cache c SET search_vector = NULL
        FROM pages p LEFT JOIN ponds po ON po.id = p.pond_id
       WHERE p.id = c.page_id
         AND (p.deleted_at IS NOT NULL OR po.deleted_at IS NOT NULL)`;
    const sources = await this.prisma.$queryRaw<(IndexSource & { page_id: string })[]>`
      SELECT p.id AS page_id, p.title, c.plain_text,
             (SELECT string_agg(l.name, ' ')
                FROM page_labels pl JOIN labels l ON l.id = pl.label_id
               WHERE pl.page_id = p.id) AS labels
        FROM page_content_cache c
        JOIN pages p ON p.id = c.page_id
        JOIN ponds po ON po.id = p.pond_id AND po.deleted_at IS NULL
       WHERE p.deleted_at IS NULL`;
    for (const source of sources) await this.writeVector(source.page_id, source);
    return sources.length;
  }

  /** Writes the weighted, normalized vector for one page (shared by index/reindex). */
  private async writeVector(pageId: string, source: IndexSource): Promise<void> {
    const title = normalizeForSearch(source.title ?? '');
    const labels = normalizeForSearch(source.labels ?? '');
    const body = normalizeForSearch(source.plain_text ?? '');
    await this.prisma.$executeRaw`
      UPDATE page_content_cache SET search_vector =
          setweight(to_tsvector('simple', ${title}), 'A')
        || setweight(to_tsvector('simple', ${labels}), 'B')
        || setweight(to_tsvector('simple', ${body}), 'C')
      WHERE page_id = ${pageId}`;
  }

  async search(query: SearchQuery, user: User): Promise<SearchResultView[]> {
    const normalized = normalizeForSearch(query.q);
    // A phrase that folds to nothing (e.g. only punctuation) matches nothing.
    if (normalized.trim() === '') return [];

    // Substring fallback (M10 follow-up): the tsquery only matches whole
    // words, so a plain LIKE over title/body catches partial words too. The
    // pattern keeps the raw (lowercased) input — umlauts etc. match verbatim;
    // diacritic-insensitive matching stays the FTS branch's job. FTS hits
    // still rank first (ts_rank is 0 for LIKE-only matches).
    const likePattern = `%${query.q
      .trim()
      .toLowerCase()
      .replace(/[\\%_]/g, (char) => `\\${char}`)}%`;

    // Pond-level prefilter (visible ponds only) keeps the LIMIT meaningful;
    // the exact per-page resolution happens below (issue #52, ADR 0010).
    const visiblePondIds = await this.permissions.visiblePondIds(user);
    if (visiblePondIds !== null && visiblePondIds.length === 0) return [];
    const visibility =
      visiblePondIds === null
        ? Prisma.empty
        : Prisma.sql`AND p.pond_id = ANY(${visiblePondIds}::text[])`;

    const scope = query.pondId ? Prisma.sql`AND p.pond_id = ${query.pondId}` : Prisma.empty;
    const labelFilter =
      query.labels && query.labels.length > 0
        ? Prisma.sql`AND EXISTS (SELECT 1 FROM page_labels pl
                                  WHERE pl.page_id = p.id AND pl.label_id = ANY(${query.labels}))`
        : Prisma.empty;

    const rows = await this.prisma.$queryRaw<SearchRow[]>(Prisma.sql`
      SELECT p.id AS "pageId", p.title, p.slug, p.pond_id AS "pondId",
             po.slug AS "pondSlug", po.name AS "pondName",
             COALESCE(
               ARRAY(SELECT pl.label_id FROM page_labels pl WHERE pl.page_id = p.id),
               '{}'
             ) AS "labelIds",
             ts_headline('simple', c.plain_text,
               websearch_to_tsquery('simple', ${query.q}), ${HEADLINE_OPTIONS}) AS snippet
        FROM page_content_cache c
        JOIN pages p ON p.id = c.page_id AND p.deleted_at IS NULL
        JOIN ponds po ON po.id = p.pond_id AND po.deleted_at IS NULL,
             websearch_to_tsquery('simple', ${normalized}) q
       WHERE (c.search_vector @@ q
                OR lower(p.title) LIKE ${likePattern}
                OR lower(c.plain_text) LIKE ${likePattern})
         ${visibility}
         ${scope}
         ${labelFilter}
       ORDER BY ts_rank(c.search_vector, q) DESC, p.updated_at DESC
       LIMIT ${SEARCH_RESULT_LIMIT}`);

    // Per-page resolution (label-/page-scope grants, deny-wins) per pond.
    const rowsByPond = new Map<string, SearchRow[]>();
    for (const row of rows) {
      const group = rowsByPond.get(row.pondId) ?? [];
      group.push(row);
      rowsByPond.set(row.pondId, group);
    }
    const readableByPond = new Map<string, Set<string>>();
    for (const [pondId, group] of rowsByPond) {
      readableByPond.set(
        pondId,
        await this.permissions.filterPages(
          user,
          pondId,
          group.map((row) => ({ id: row.pageId, labelIds: row.labelIds })),
          'read',
        ),
      );
    }

    return rows
      .filter((row) => readableByPond.get(row.pondId)?.has(row.pageId))
      .map((row) => ({
        pageId: row.pageId,
        title: row.title,
        slug: row.slug,
        pondId: row.pondId,
        pondSlug: row.pondSlug,
        pondName: row.pondName,
        labelIds: row.labelIds,
        snippet: row.snippet,
      }));
  }
}