Skip to content

feat(rag): curate legacy character facts and clarify relationship subjects - #60

Merged
Lucas1479 merged 1 commit into
mainfrom
codex/character-knowledge-expansion
Sep 7, 2026
Merged

feat(rag): curate legacy character facts and clarify relationship subjects#60
Lucas1479 merged 1 commit into
mainfrom
codex/character-knowledge-expansion

Conversation

@Lucas1479

Copy link
Copy Markdown
Member

The old 148-entry local corpus contains useful character facts alongside duplicates, incorrect member numbers and unverified anecdotes. This data-only follow-up to #59 audits every legacy row and expands the public corpus from 30 to 52 Chinese/Japanese records, covering 26 topics.

The update corrects Moeka/Ruka/Faris to 005/006/007, adds the Science publication, chopsticks/cooking, father relationship, PhoneWave, Okabe's phrase, supporting characters and the lab building, and deduplicates already covered facts. Twenty old rows have core facts already covered; verified portions of 16 more are incorporated/corrected; 112 are deferred for sourcing or narrative context, not declared universally false. The original local files remain unchanged. Sources and all row dispositions are documented.

The father entry uses the maintainer's suggested subject anchoring: 私(牧瀬紅莉栖)の父親はドクター中鉢。 Other reviewed wording stays intact. A blanket first-person rewrite was tested but added four unrelated-query false hits. No runtime code, static persona, RAG wrapper, default-off setting, threshold, top-k, or dependencies change.

Validation:

  • 70 targeted Python tests passed, including all 67 development retrieval cases and release-tooling tests; Ruff, architecture views and source-release validation passed.
  • Additional held-out retrieval improved from 2/18 with the 30-entry corpus to 11/18; remaining failures and six pre-existing unrelated-query false hits are explicitly recorded.
  • 153 live Flash requests across corpus comparison and follow-up wording tests. Central facts improved, but these are not 153 persona passes. RAG-off and first-person arms both produced an explicit in the focused experiment; shared-history fabrication and unsupported elaboration remain.

Audit, method, results and limitations. Related to #56; this does not close its default-off persona acceptance criterion.


这是 #59 后续的纯资料更新:逐条审看旧本地 148 条,去重、核对及修正后,公开库从 30 条扩为中日文 52 条、26 个主题。20 条的核心事实已有覆盖,16 条中的可核实事实融入,112 条待来源或剧情上下文确认;不代表判定它们全部错误,旧本地文件未改动。

修正萌郁/琉华/菲利斯的编号,补充论文期刊、筷子与料理、父亲、电话微波炉、角色及场所等事实。父亲条目采用明确的“私(牧瀬紅莉栖)の父親は…”写法;全库统一改第一人称新增了四个无关问题误召回,因此没有直接采用。默认 prompt、RAG 包装、开关、阈值、top-k 和依赖不变。

70 项相关 Python 测试通过,包含 67 条真实检索开发回归;额外问法从 2/18 提升至 11/18,失败情况完整保留。另做 153 次真实模型调用,确认事实覆盖改善,但男性口吻、虚构共同经历及附加解释仍有反例,不宣称人物表现全部修好。建议作为默认关闭 RAG 的资料更新审阅,#56 继续开启。

@Lucas1479
Lucas1479 merged commit 39706c0 into main Sep 7, 2026
6 checks passed
@Lucas1479
Lucas1479 deleted the codex/character-knowledge-expansion branch September 7, 2026 00:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant