diff --git a/CHANGELOG.md b/CHANGELOG.md index 9428cd5..85093f6 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -2,6 +2,11 @@ ## [Unreleased] +### New Features + +- **`jobradar cache prune-scores`** (`cli.py` / `cache.py`) + Drops match results that no longer reflect the current scoring setup, so the next search or `assess` recomputes them. By default it removes rows whose `prompt_version` differs from the current one — useful right after bumping `matching.PROMPT_VERSION`. `--stale-cv` also drops rows belonging to any CV other than the latest, and `--orphans` drops rows whose job is gone from `job_cache`. The command previews the counts and asks for confirmation before deleting (`--yes` skips the prompt). Categories can overlap, so each is reported separately alongside a de-duplicated total. + ### Changes - **Scores now come from `job_matches` alone** (`schemas.py` / `display.py` / `cli.py` / `search_assessment_stage.py` / `search_prefilter.py` / `cache.py`) diff --git a/README.es.md b/README.es.md index 4bf7d6f..668845c 100644 --- a/README.es.md +++ b/README.es.md @@ -33,6 +33,7 @@ uv run jobradar find cv.docx # Modo CLI | `uv run jobradar assess` | Reejecuta la evaluación LLM sobre JDs en caché | | `uv run jobradar model` | Selecciona interactivamente el proveedor y modelo LLM | | `uv run jobradar cache clear` | Limpia toda la caché | +| `uv run jobradar cache prune-scores` | Elimina resultados de coincidencia de una versión de prompt o CV obsoletos | | `uv run jobradar --version` | Muestra la versión actual | ## Visión General del Pipeline diff --git a/README.md b/README.md index b107d5e..a987d75 100644 --- a/README.md +++ b/README.md @@ -33,6 +33,7 @@ uv run jobradar find cv.docx # CLI mode | `uv run jobradar assess` | Re-run LLM assessment on cached JDs | | `uv run jobradar model` | Interactively choose LLM provider and model | | `uv run jobradar cache clear` | Clear all caches | +| `uv run jobradar cache prune-scores` | Drop match results from an outdated prompt version or CV | | `uv run jobradar --version` | Show current version | ## Pipeline Overview diff --git a/README.zh.md b/README.zh.md index f93333f..1a5e52e 100644 --- a/README.zh.md +++ b/README.zh.md @@ -33,6 +33,7 @@ uv run jobradar find cv.docx # CLI 模式 | `uv run jobradar assess` | 对缓存 JD 单独补跑 LLM 评估 | | `uv run jobradar model` | 交互式选择 LLM provider 和模型 | | `uv run jobradar cache clear` | 清空所有缓存 | +| `uv run jobradar cache prune-scores` | 删除过时 prompt 版本或旧 CV 的匹配结果 | | `uv run jobradar --version` | 显示当前版本号 | ## Pipeline 概览 diff --git a/jobradar/cache.py b/jobradar/cache.py index 5230523..3db06d8 100644 --- a/jobradar/cache.py +++ b/jobradar/cache.py @@ -1464,6 +1464,66 @@ def clear_search_stats() -> None: con.execute("DELETE FROM filter_events") +def _stale_match_conditions( + prompt_version: str, keep_cv_hash: str, drop_orphans: bool +) -> tuple[list[str], list[str]]: + """构造 job_matches 的过时行判定条件，返回 (SQL 片段, 参数)。""" + clauses: list[str] = [] + params: list[str] = [] + if prompt_version: + clauses.append("prompt_version != ?") + params.append(prompt_version) + if keep_cv_hash: + clauses.append("cv_hash != ?") + params.append(keep_cv_hash) + if drop_orphans: + clauses.append("job_id NOT IN (SELECT dedup_key FROM job_cache)") + return clauses, params + + +def prune_job_matches( + *, + prompt_version: str = "", + keep_cv_hash: str = "", + drop_orphans: bool = False, + dry_run: bool = True, +) -> dict[str, int]: + """删除过时的匹配结果，被删除的职位会在下次搜索或 assess 时按当前口径重算。 + + 三类条件独立启用，彼此可重叠（同一行可能既是旧 prompt 版本又属于旧 CV）， + 因此各分类计数之和可能大于 ``total``；``total`` 才是唯一行数。 + ``dry_run=True`` 时只统计不删除。 + """ + clauses, params = _stale_match_conditions(prompt_version, keep_cv_hash, drop_orphans) + result = {"stale_version": 0, "stale_cv": 0, "orphan": 0, "total": 0, "deleted": 0} + if not clauses: + return result + + with _conn() as con: + if prompt_version: + result["stale_version"] = con.execute( + "SELECT count(*) FROM job_matches WHERE prompt_version != ?", (prompt_version,) + ).fetchone()[0] + if keep_cv_hash: + result["stale_cv"] = con.execute( + "SELECT count(*) FROM job_matches WHERE cv_hash != ?", (keep_cv_hash,) + ).fetchone()[0] + if drop_orphans: + result["orphan"] = con.execute( + "SELECT count(*) FROM job_matches WHERE job_id NOT IN (SELECT dedup_key FROM job_cache)" + ).fetchone()[0] + + where = " OR ".join(clauses) + result["total"] = con.execute( + f"SELECT count(*) FROM job_matches WHERE {where}", params + ).fetchone()[0] + if not dry_run: + result["deleted"] = con.execute( + f"DELETE FROM job_matches WHERE {where}", params + ).rowcount + return result + + def clean_expired() -> int: """删除过期 JD 和 Session，返回删除条数。""" now = datetime.utcnow().isoformat() diff --git a/jobradar/cli.py b/jobradar/cli.py index d2e7119..aae70d1 100644 --- a/jobradar/cli.py +++ b/jobradar/cli.py @@ -558,6 +558,54 @@ def cache_clean() -> None: console.print(f"[green]已清理 {count} 条过期记录。[/green]") +@cache_app.command("prune-scores") +def cache_prune_scores( + stale_cv: Annotated[bool, typer.Option("--stale-cv", help="同时删除非当前 CV 的匹配结果")] = False, + orphans: Annotated[bool, typer.Option("--orphans", help="同时删除职位已不存在的匹配结果")] = False, + language: Annotated[str, typer.Option("--language", help="匹配 prompt 使用的语言")] = "zh", + yes: Annotated[bool, typer.Option("--yes", "-y", help="跳过确认")] = False, +) -> None: + """删除过时的匹配结果。被删除的职位会在下次搜索或 assess 时按当前口径重算。""" + from jobradar.matching import match_prompt_version + + current_version = match_prompt_version(language) + keep_cv_hash = cache.get_latest_cv_hash() if stale_cv else "" + if stale_cv and not keep_cv_hash: + console.print("[red]缓存中没有 CVProfile，无法确定当前 CV，--stale-cv 不可用。[/red]") + raise typer.Exit(1) + + preview = cache.prune_job_matches( + prompt_version=current_version, + keep_cv_hash=keep_cv_hash, + drop_orphans=orphans, + dry_run=True, + ) + if preview["total"] == 0: + console.print("[dim]没有需要清理的匹配结果。[/dim]") + return + + console.print(f"\n[bold]将删除 {preview['total']} 条匹配结果：[/bold]") + console.print(f" prompt 版本不是 {current_version}：{preview['stale_version']} 条") + if stale_cv: + console.print(f" 不属于当前 CV（{keep_cv_hash[:8]}）：{preview['stale_cv']} 条") + if orphans: + console.print(f" 职位已不存在：{preview['orphan']} 条") + console.print("[dim]（分类可能重叠，总数为去重后的行数）[/dim]") + + if not yes and not prompt_confirm("确认删除？被删除的职位需要重新评估"): + console.print("[yellow]已取消。[/yellow]") + return + + result = cache.prune_job_matches( + prompt_version=current_version, + keep_cv_hash=keep_cv_hash, + drop_orphans=orphans, + dry_run=False, + ) + console.print(f"[green]已删除 {result['deleted']} 条匹配结果。[/green]") + console.print("[dim]运行 `jobradar assess` 可按当前口径补算。[/dim]") + + # ─── serve 命令 ────────────────────────────────────────────────────────────── diff --git a/tests/test_cache.py b/tests/test_cache.py index 09a86c4..54afd4c 100644 --- a/tests/test_cache.py +++ b/tests/test_cache.py @@ -645,3 +645,102 @@ def fake_evaluate(job, profile, llm, cv_hash="", language="zh"): assert (succeeded, failed) == (3, 0) # 第二次调用必须为空，否则 assess 会陷入每次重评同一批职位的循环。 assert temp_db.get_unassessed_jobs() == [] + + +class TestPruneJobMatches: + def _save(self, temp_db, job_id: str, cv_hash: str, prompt_version: str) -> None: + match = MatchScore( + job_id=job_id, + cv_hash=cv_hash, + overall_score=70, + title_score=70, + seniority_score=70, + must_have_score=70, + nice_to_have_score=70, + domain_score=70, + location_score=100, + language_score=100, + risk_penalty=0, + recommendation="apply", + ) + temp_db.save_job_match(match, "desc", prompt_version=prompt_version) + + def _remaining(self, temp_db) -> set[tuple[str, str]]: + with temp_db._conn() as con: + return {(r["job_id"], r["prompt_version"]) for r in con.execute("SELECT * FROM job_matches")} + + def test_dry_run_reports_without_deleting(self, temp_db): + job = make_job(description_snippet="desc") + temp_db.save_job(job) + self._save(temp_db, job.dedup_key, "cv1", "match_v10:zh") + + result = temp_db.prune_job_matches(prompt_version="match_v11:zh", dry_run=True) + + assert result["stale_version"] == 1 + assert result["total"] == 1 + assert result["deleted"] == 0 + assert len(self._remaining(temp_db)) == 1 + + def test_deletes_only_stale_prompt_versions(self, temp_db): + job = make_job(description_snippet="desc") + other = make_job(company="Other", url="http://example.com/2", description_snippet="desc") + temp_db.save_job(job) + temp_db.save_job(other) + self._save(temp_db, job.dedup_key, "cv1", "match_v10:zh") + self._save(temp_db, other.dedup_key, "cv1", "match_v11:zh") + + result = temp_db.prune_job_matches(prompt_version="match_v11:zh", dry_run=False) + + assert result["deleted"] == 1 + assert self._remaining(temp_db) == {(other.dedup_key, "match_v11:zh")} + + def test_stale_cv_hash_is_optional(self, temp_db): + job = make_job(description_snippet="desc") + temp_db.save_job(job) + self._save(temp_db, job.dedup_key, "cv-old", "match_v11:zh") + + untouched = temp_db.prune_job_matches(prompt_version="match_v11:zh", dry_run=True) + assert untouched["total"] == 0 + + targeted = temp_db.prune_job_matches( + prompt_version="match_v11:zh", keep_cv_hash="cv-current", dry_run=True + ) + assert targeted["stale_cv"] == 1 + assert targeted["total"] == 1 + + def test_orphan_rows_removed_only_when_requested(self, temp_db): + self._save(temp_db, "ghost|role", "cv1", "match_v11:zh") + + assert temp_db.prune_job_matches(prompt_version="match_v11:zh", dry_run=True)["total"] == 0 + + result = temp_db.prune_job_matches( + prompt_version="match_v11:zh", drop_orphans=True, dry_run=False + ) + assert result["orphan"] == 1 + assert result["deleted"] == 1 + assert self._remaining(temp_db) == set() + + def test_overlapping_rows_counted_once_in_total(self, temp_db): + """同一行同时命中多个条件时，total 不应重复计数。""" + job = make_job(description_snippet="desc") + temp_db.save_job(job) + self._save(temp_db, job.dedup_key, "cv-old", "match_v10:zh") + + result = temp_db.prune_job_matches( + prompt_version="match_v11:zh", keep_cv_hash="cv-current", dry_run=False + ) + + assert result["stale_version"] == 1 + assert result["stale_cv"] == 1 + assert result["total"] == 1 + assert result["deleted"] == 1 + + def test_no_conditions_is_a_noop(self, temp_db): + job = make_job(description_snippet="desc") + temp_db.save_job(job) + self._save(temp_db, job.dedup_key, "cv1", "match_v11:zh") + + result = temp_db.prune_job_matches(dry_run=False) + + assert result["total"] == 0 and result["deleted"] == 0 + assert len(self._remaining(temp_db)) == 1