# NiubiGEO v0.3.3

## English

This release makes monitoring easier to inspect and export, and consolidates the open issue backlog. It includes the previously merged fixes in PRs #247 and #248.

- Filter keyword history by inclusive UTC dates and model. Inspect mention proportions per observation, with failures and run conditions visible; recommendation/source frequencies remain available per model.
- Export filtered keyword history as CSV/JSON, selected recognition report evidence, and filtered measurement chart points. Missing values and failed samples are preserved. Existing Diff CSV export remains available.
- Run bounded repeated sampling with `keyword-run --repeats 5`; `--dry-run` estimates requests without inference. Currency cost is unknown, and provider retries/search may add charges.
- Show worker status, comparable Diff warnings, failed-model retries, tags and sample controls from PR #248.
- Recognize explicitly configured aliases with Unicode normalization and Latin word boundaries, accept structured recommendation objects, and reject manual execution for archived projects.
- Clarify installation, local Ollama connections, evidence interpretation and the current version in both READMEs.

### Upgrade

Back up the complete actual product data directory and deployment configuration after stopping writers. Update both Web and worker images to `ghcr.io/albert-weasker/niubigeo:v0.3.3`; pull and recreate containers. Start Web first, inspect history, then restore the worker with the same data directory and server-side keys. Browser session keys cannot be used by the worker.

Old history remains readable. New observations use `keyword-monitor-v2`; old-protocol samples cannot be mixed into a retry with the new matcher. Run a new observation instead. Cross-protocol comparisons warn about changed conditions. Roll back using the pre-upgrade backup and previous image.

The release workflow validates Linux amd64 and arm64 images before publishing `v0.3.3` and `latest`. Its acceptance reports also check a v0.3.2 → v0.3.3 upgrade in isolated volumes using explicitly synthetic history, including unchanged original record bytes. Reports are CI artifacts; this is not a test of every historical schema or real provider inference. Earlier security compatibility tags retain their documented v0.3.2 target.

### Limits

No regional proxy simulation, automated email/PDF reports, aggregate competitor dashboard or tag score is added. Charts show observations, not statistical confidence or independent trends across changed conditions/retries. Costs, citations and API/consumer-app differences still require interpretation. Tests use fixtures and do not establish live model quality or local Ollama hardware compatibility.

[Monitoring guide](../monitoring-guide.md) · [Docker](../deployment/docker.en.md) · [Issue ledger](../issue-triage-v0.3.3.md)

## 中文

本版补齐监控历史、数据导出和重复采样入口，并整理归并未关闭的 Issue，包含此前已合并的 PR #247、#248 修复。

- 关键词历史支持 UTC 起止日期、模型筛选和逐次提及比例图；保留失败、运行条件及推荐/来源出现次数。
- CSV/JSON 导出遵循当前筛选，覆盖关键词历史、选中的认知报告和持续测量图表。保留缺失值与失败记录，原有 Diff CSV 导出继续可用。
- CLI 支持 `keyword-run --repeats 5`；`--dry-run` 不调用模型，仅估算请求次数。金额未知，服务商重试与联网可能额外收费。
- 包含 Worker 状态、Diff 可比性提示、失败模型重试、标签和采样次数控制。
- 修复已配置别名的 Unicode 与英文词边界识别、对象形式推荐结果解析，以及归档项目仍可手动运行的问题。
- 补充中英文结果解读、安装及 Ollama 接入说明，统一当前版本入口。

升级前停写并备份真实数据目录与配置，Web 和 worker 同时改用 `ghcr.io/albert-weasker/niubigeo:v0.3.3`，拉取并重建。先启动 Web 检查历史，再恢复 worker。旧记录保持可读；新观测采用 v2 识别协议，旧协议样本不能混合重试，请重新运行一次。回滚使用升级前备份。

发布工作流在 Linux amd64/arm64 上验收镜像，通过后才更新版本标签和 `latest`。另用隔离数据卷检查 v0.3.2 升级后的项目、关键词及明确标注的合成历史，核对原记录未被改写；这不代表所有历史格式、真实模型推理或所有用户数据均已验证。历史安全兼容标签仍指向已说明的 v0.3.2。

本版不增加地域代理模拟、邮件/PDF 自动报告、复杂竞品面板或标签聚合评分。重试与条件变化不能当作独立趋势；来源链接不证明推荐因果。没有执行付费模型实验或本地 Ollama 硬件评测。

[监控解读](../monitoring-guide.zh-CN.md) · [Docker](../deployment/docker.md) · [升级与回滚](../upgrade.md) · [Issue 整理记录](../issue-triage-v0.3.3.md)