af870bd157
New GET /v1/me/page-search surfacing the analysis worker's output: - repo::page_analysis::page_search + PageSearchQuery: multi-tag AND across (user page_tags ∪ global page_auto_tags) via the unnest double-negative idiom, weighted OCR/scene text ranking (ts_rank over search_doc), and content-warning include/exclude. One row per page with is_nsfw + deduped content_warnings + rank. - domain::PageSearchItem. - api::page_tags: /me/page-search handler with CSV tag/warning parsing (parse_tags_csv reuses normalize_tag; parse_warnings_csv validates the closed vocabulary), requiring at least one positive filter (422 else). This is where the reserved OCR text search lands for pages. Tests: multi-tag AND user∪auto, speech>sfx ranking, cw include/exclude (+ row flags), text-only (no tags), missing-filter 422, unknown-warning 422, auth required. Note: the /me/page-tags/chapters|mangas aggregations keep single-tag behavior + the reserved text=501 for now; page-level search is the primary text surface. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
372 lines
13 KiB
Rust
372 lines
13 KiB
Rust
//! Persistence for AI page-analysis results.
|
||
//!
|
||
//! [`persist_analysis`] is the single transactional writer: it replaces a
|
||
//! page's OCR text, global auto-tags, and content warnings, upserts the
|
||
//! `page_analysis` row, and computes the kind-weighted `search_doc`
|
||
//! tsvector — all atomically, so re-analysis (delete+reinsert) is
|
||
//! idempotent and never leaves a half-written page. It deliberately does
|
||
//! NOT touch the per-user `page_tags` table.
|
||
//!
|
||
//! Auto-tag names resolve to the shared `tags` vocabulary via
|
||
//! [`crate::repo::tag::upsert_by_name`], the same path manga tags and
|
||
//! personal page tags use.
|
||
|
||
use sqlx::PgPool;
|
||
use uuid::Uuid;
|
||
|
||
use crate::crawler::jobs::{self, JobPayload};
|
||
use crate::domain::page_analysis::{
|
||
ContentWarning, OcrKind, PageAnalysis, PageSearchItem, VisionAnalysis,
|
||
};
|
||
use crate::error::AppResult;
|
||
|
||
/// Filter set for [`page_search`]. `tags` are AND-ed (a page must carry
|
||
/// every one, satisfiable by EITHER the user's page tag or a global auto
|
||
/// tag); `text` ranks via the weighted tsvector; `cw_include` requires all
|
||
/// listed warnings; `cw_exclude` rejects any. All string values arrive
|
||
/// pre-normalized (tags lowercased, warnings validated) from the handler.
|
||
#[derive(Debug, Default, Clone)]
|
||
pub struct PageSearchQuery {
|
||
pub user_id: uuid::Uuid,
|
||
pub tags: Vec<String>,
|
||
pub text: Option<String>,
|
||
pub cw_include: Vec<String>,
|
||
pub cw_exclude: Vec<String>,
|
||
pub limit: i64,
|
||
pub offset: i64,
|
||
}
|
||
|
||
/// Longest tag the shared `tags` table accepts (`upsert_by_name` enforces
|
||
/// the same bound). Auto-tags over this are dropped here rather than
|
||
/// aborting the whole page's analysis on one bad model output.
|
||
const MAX_TAG_CHARS: usize = 64;
|
||
|
||
/// Enqueue an `analyze_page` job for `page_id`. `force` re-analyzes a page
|
||
/// that is already `done`. Enqueue is idempotent at the job level only in
|
||
/// that duplicate pending jobs are harmless — processing is idempotent.
|
||
pub async fn enqueue_for_page(pool: &PgPool, page_id: Uuid, force: bool) -> AppResult<()> {
|
||
jobs::enqueue(pool, &JobPayload::AnalyzePage { page_id, force }).await?;
|
||
Ok(())
|
||
}
|
||
|
||
/// Bulk-enqueue `analyze_page` jobs for existing pages — the admin
|
||
/// backfill path for content uploaded/crawled before analysis was
|
||
/// enabled. When `only_unanalyzed` is true, pages that already have a
|
||
/// `done` analysis row are skipped. Pages with a pending/running
|
||
/// `analyze_page` job are always skipped so repeated calls don't pile up
|
||
/// duplicates. Returns the number of jobs enqueued.
|
||
pub async fn enqueue_all_pages(pool: &PgPool, only_unanalyzed: bool) -> AppResult<u64> {
|
||
let result = sqlx::query(
|
||
r#"
|
||
INSERT INTO crawler_jobs (payload)
|
||
SELECT jsonb_build_object('kind', 'analyze_page', 'page_id', p.id, 'force', false)
|
||
FROM pages p
|
||
WHERE ($1 = false OR NOT EXISTS (
|
||
SELECT 1 FROM page_analysis pa
|
||
WHERE pa.page_id = p.id AND pa.status = 'done'))
|
||
AND NOT EXISTS (
|
||
SELECT 1 FROM crawler_jobs j
|
||
WHERE j.payload->>'kind' = 'analyze_page'
|
||
AND j.payload->>'page_id' = p.id::text
|
||
AND j.state IN ('pending', 'running'))
|
||
"#,
|
||
)
|
||
.bind(only_unanalyzed)
|
||
.execute(pool)
|
||
.await?;
|
||
Ok(result.rows_affected())
|
||
}
|
||
|
||
/// Load a page's analysis row, if it has one.
|
||
pub async fn load(pool: &PgPool, page_id: Uuid) -> AppResult<Option<PageAnalysis>> {
|
||
let row = sqlx::query_as::<_, PageAnalysis>(
|
||
r#"
|
||
SELECT page_id, status, scene_description, is_nsfw, model, error, analyzed_at
|
||
FROM page_analysis
|
||
WHERE page_id = $1
|
||
"#,
|
||
)
|
||
.bind(page_id)
|
||
.fetch_optional(pool)
|
||
.await?;
|
||
Ok(row)
|
||
}
|
||
|
||
/// Record that analysis failed terminally for a page (retries exhausted /
|
||
/// dead-lettered). Leaves a `failed` row so the page's state is
|
||
/// observable; it simply contributes nothing to search.
|
||
pub async fn mark_failed(pool: &PgPool, page_id: Uuid, error: &str) -> AppResult<()> {
|
||
sqlx::query(
|
||
r#"
|
||
INSERT INTO page_analysis (page_id, status, error)
|
||
VALUES ($1, 'failed', $2)
|
||
ON CONFLICT (page_id) DO UPDATE
|
||
SET status = 'failed', error = EXCLUDED.error, updated_at = now()
|
||
"#,
|
||
)
|
||
.bind(page_id)
|
||
.bind(error)
|
||
.execute(pool)
|
||
.await?;
|
||
Ok(())
|
||
}
|
||
|
||
/// Persist a completed analysis for a page, replacing any previous result.
|
||
///
|
||
/// The model's free-form `kind` / `content_type` strings are mapped onto
|
||
/// the closed vocabularies here ([`OcrKind::from_model_str`] /
|
||
/// [`ContentWarning::from_model_str`]); unrecognized warnings are dropped,
|
||
/// unknown OCR kinds fall back to `narration`. Empty / over-long tags and
|
||
/// empty OCR text are skipped so one bad item never aborts the page.
|
||
pub async fn persist_analysis(
|
||
pool: &PgPool,
|
||
page_id: Uuid,
|
||
analysis: &VisionAnalysis,
|
||
model: &str,
|
||
) -> AppResult<()> {
|
||
// Kind-weighted text buckets for the tsvector: A=speech/title,
|
||
// B=narration/thought/caption, C=scene, D=sfx.
|
||
let mut bucket_a = String::new();
|
||
let mut bucket_b = String::new();
|
||
let bucket_c = analysis.scene_description.trim().to_string();
|
||
let mut bucket_d = String::new();
|
||
|
||
let mut ocr_rows: Vec<(OcrKind, String)> = Vec::new();
|
||
for r in &analysis.ocr_results {
|
||
let text = r.text.trim();
|
||
if text.is_empty() {
|
||
continue;
|
||
}
|
||
let kind = OcrKind::from_model_str(&r.kind);
|
||
match kind.weight() {
|
||
'A' => push_token(&mut bucket_a, text),
|
||
'B' => push_token(&mut bucket_b, text),
|
||
'D' => push_token(&mut bucket_d, text),
|
||
_ => {}
|
||
}
|
||
ocr_rows.push((kind, text.to_string()));
|
||
}
|
||
|
||
// Dedup tags case-insensitively, preserving first-seen order; drop
|
||
// empties and over-long names rather than failing the transaction.
|
||
let mut seen_tags = std::collections::HashSet::new();
|
||
let mut tags: Vec<String> = Vec::new();
|
||
for raw in &analysis.tagging_results {
|
||
let t = raw.trim();
|
||
if t.is_empty() || t.chars().count() > MAX_TAG_CHARS {
|
||
continue;
|
||
}
|
||
if seen_tags.insert(t.to_lowercase()) {
|
||
tags.push(t.to_string());
|
||
}
|
||
}
|
||
|
||
// Map + dedup warnings.
|
||
let mut seen_warn = std::collections::HashSet::new();
|
||
let mut warnings: Vec<ContentWarning> = Vec::new();
|
||
for raw in &analysis.safety_flag.content_type {
|
||
if let Some(w) = ContentWarning::from_model_str(raw) {
|
||
if seen_warn.insert(w) {
|
||
warnings.push(w);
|
||
}
|
||
}
|
||
}
|
||
|
||
let mut tx = pool.begin().await?;
|
||
|
||
sqlx::query("DELETE FROM page_ocr_text WHERE page_id = $1")
|
||
.bind(page_id)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
sqlx::query("DELETE FROM page_auto_tags WHERE page_id = $1")
|
||
.bind(page_id)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
sqlx::query("DELETE FROM page_content_warnings WHERE page_id = $1")
|
||
.bind(page_id)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
|
||
for (ord, (kind, text)) in ocr_rows.iter().enumerate() {
|
||
sqlx::query(
|
||
"INSERT INTO page_ocr_text (page_id, kind, text, ord) VALUES ($1, $2, $3, $4)",
|
||
)
|
||
.bind(page_id)
|
||
.bind(kind)
|
||
.bind(text)
|
||
.bind(ord as i32)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
}
|
||
|
||
for tag in &tags {
|
||
let tag_row = crate::repo::tag::upsert_by_name(&mut *tx, tag).await?;
|
||
sqlx::query(
|
||
"INSERT INTO page_auto_tags (page_id, tag_id) VALUES ($1, $2) \
|
||
ON CONFLICT (page_id, tag_id) DO NOTHING",
|
||
)
|
||
.bind(page_id)
|
||
.bind(tag_row.id)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
}
|
||
|
||
for w in &warnings {
|
||
sqlx::query(
|
||
"INSERT INTO page_content_warnings (page_id, warning) VALUES ($1, $2) \
|
||
ON CONFLICT (page_id, warning) DO NOTHING",
|
||
)
|
||
.bind(page_id)
|
||
.bind(w)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
}
|
||
|
||
let scene = if bucket_c.is_empty() {
|
||
None
|
||
} else {
|
||
Some(bucket_c.as_str())
|
||
};
|
||
sqlx::query(
|
||
r#"
|
||
INSERT INTO page_analysis
|
||
(page_id, status, scene_description, is_nsfw, model, analyzed_at, search_doc)
|
||
VALUES
|
||
($1, 'done', $2, $3, $4, now(),
|
||
setweight(to_tsvector('simple', $5), 'A') ||
|
||
setweight(to_tsvector('simple', $6), 'B') ||
|
||
setweight(to_tsvector('simple', $7), 'C') ||
|
||
setweight(to_tsvector('simple', $8), 'D'))
|
||
ON CONFLICT (page_id) DO UPDATE SET
|
||
status = 'done',
|
||
scene_description = EXCLUDED.scene_description,
|
||
is_nsfw = EXCLUDED.is_nsfw,
|
||
model = EXCLUDED.model,
|
||
analyzed_at = now(),
|
||
error = NULL,
|
||
search_doc = EXCLUDED.search_doc,
|
||
updated_at = now()
|
||
"#,
|
||
)
|
||
.bind(page_id)
|
||
.bind(scene)
|
||
.bind(analysis.safety_flag.is_nsfw)
|
||
.bind(model)
|
||
.bind(&bucket_a)
|
||
.bind(&bucket_b)
|
||
.bind(&bucket_c)
|
||
.bind(&bucket_d)
|
||
.execute(&mut *tx)
|
||
.await?;
|
||
|
||
tx.commit().await?;
|
||
Ok(())
|
||
}
|
||
|
||
/// Content search over pages: multi-tag AND across (user page tags ∪
|
||
/// global auto tags), optional weighted text search over the OCR/scene
|
||
/// document, and content-warning include/exclude. Returns one row per
|
||
/// matching page plus the total for pagination.
|
||
///
|
||
/// Empty `tags` makes the tag clause vacuously true (text- or
|
||
/// warning-only search), mirroring the manga search's `unnest` idiom. The
|
||
/// caller is responsible for requiring at least one positive filter so
|
||
/// this never degenerates into "every page".
|
||
pub async fn page_search(
|
||
pool: &PgPool,
|
||
q: &PageSearchQuery,
|
||
) -> AppResult<(Vec<PageSearchItem>, i64)> {
|
||
// `text` participates in three places ($3): the rank, the match
|
||
// predicate, and the order key. Empty string disables text filtering.
|
||
let text = q.text.as_deref().unwrap_or("").trim();
|
||
|
||
const WHERE: &str = r#"
|
||
NOT EXISTS (
|
||
SELECT 1 FROM unnest($2::text[]) AS req(name)
|
||
WHERE NOT EXISTS (
|
||
SELECT 1 FROM page_tags ut
|
||
JOIN tags t ON t.id = ut.tag_id
|
||
WHERE ut.page_id = p.id AND ut.user_id = $1 AND lower(t.name) = req.name
|
||
UNION ALL
|
||
SELECT 1 FROM page_auto_tags at
|
||
JOIN tags t ON t.id = at.tag_id
|
||
WHERE at.page_id = p.id AND lower(t.name) = req.name
|
||
)
|
||
)
|
||
AND ($3 = '' OR pa.search_doc @@ plainto_tsquery('simple', $3))
|
||
AND NOT EXISTS (
|
||
SELECT 1 FROM unnest($4::text[]) AS req(w)
|
||
WHERE NOT EXISTS (
|
||
SELECT 1 FROM page_content_warnings pw
|
||
WHERE pw.page_id = p.id AND pw.warning = req.w
|
||
)
|
||
)
|
||
AND NOT EXISTS (
|
||
SELECT 1 FROM page_content_warnings pw
|
||
WHERE pw.page_id = p.id AND pw.warning = ANY($5::text[])
|
||
)
|
||
"#;
|
||
|
||
let rows_sql = format!(
|
||
r#"
|
||
SELECT
|
||
p.id AS page_id,
|
||
p.chapter_id AS chapter_id,
|
||
ch.manga_id AS manga_id,
|
||
p.page_number AS page_number,
|
||
ch.number AS chapter_number,
|
||
ch.title AS chapter_title,
|
||
m.title AS manga_title,
|
||
p.storage_key AS storage_key,
|
||
COALESCE(pa.is_nsfw, false) AS is_nsfw,
|
||
COALESCE(
|
||
(SELECT array_agg(pw.warning ORDER BY pw.warning)
|
||
FROM page_content_warnings pw WHERE pw.page_id = p.id),
|
||
ARRAY[]::text[]
|
||
) AS content_warnings,
|
||
COALESCE(ts_rank(pa.search_doc, plainto_tsquery('simple', $3)), 0)::real AS rank
|
||
FROM pages p
|
||
JOIN chapters ch ON ch.id = p.chapter_id
|
||
JOIN mangas m ON m.id = ch.manga_id
|
||
LEFT JOIN page_analysis pa ON pa.page_id = p.id
|
||
WHERE {WHERE}
|
||
ORDER BY (CASE WHEN $3 = '' THEN 0 ELSE 1 END) DESC, rank DESC, p.id
|
||
LIMIT $6 OFFSET $7
|
||
"#
|
||
);
|
||
let rows = sqlx::query_as::<_, PageSearchItem>(&rows_sql)
|
||
.bind(q.user_id)
|
||
.bind(&q.tags)
|
||
.bind(text)
|
||
.bind(&q.cw_include)
|
||
.bind(&q.cw_exclude)
|
||
.bind(q.limit)
|
||
.bind(q.offset)
|
||
.fetch_all(pool)
|
||
.await?;
|
||
|
||
let count_sql = format!(
|
||
"SELECT count(*) FROM pages p \
|
||
JOIN chapters ch ON ch.id = p.chapter_id \
|
||
JOIN mangas m ON m.id = ch.manga_id \
|
||
LEFT JOIN page_analysis pa ON pa.page_id = p.id \
|
||
WHERE {WHERE}"
|
||
);
|
||
let (total,): (i64,) = sqlx::query_as(&count_sql)
|
||
.bind(q.user_id)
|
||
.bind(&q.tags)
|
||
.bind(text)
|
||
.bind(&q.cw_include)
|
||
.bind(&q.cw_exclude)
|
||
.fetch_one(pool)
|
||
.await?;
|
||
|
||
Ok((rows, total))
|
||
}
|
||
|
||
/// Append a token to a tsvector text bucket with a trailing space.
|
||
fn push_token(bucket: &mut String, text: &str) {
|
||
bucket.push_str(text);
|
||
bucket.push(' ');
|
||
}
|