Skills Plugins MCP Prompt Model 博客 我的中心

session-rag-eval

Run and debug Chatbox session attachment RAG model evaluation with synthetic and real long-file fixtures.

DeepseekModel キュレーション済みスキル 品質 優秀 · 90 v1.0.0

取得

https://deepseekmodel.com/api/download.php?id=chatboxai-chatbox-scripts-session-rag-eval-skill-md&format=skill
ダウンロード .skill 標準形式。system_prompt と model_config を収録し、任意の Agent で利用可能
.skill ファイルの system_prompt フィールドの実際の内容。
name session-rag-eval description Run and debug Chatbox session attachment RAG model evaluation with synthetic and real long-file fixtures. Session Attachment RAG Eval Use this skill when validating file question-answering behavior for large chat attachments. Goal Verify both sides of model behavior: The model calls query_session_attachment when the answer depends on an uploaded large file. The model avoids retrieval for clearly unrelated user requests. Prefer the Chatbox conversation-flow harness for product validation because it exercises the real renderer, config, license, local API, file upload, indexing, tool registration, and persisted messages. Fixtures Default fixture repo: ../../chatbox-session-rag-eval-fixtures Fixture types: synthetic long documents with hidden needle facts; Wikipedia-derived real long documents with source attribution; direct positive cases; implicit positive cases where the user does not explicitly ask to search the file; unrelated negative cases. Regenerate fixtures in the fixture repo: node scripts/generate-fixtures.mjs node scripts/fetch-real-fixtures.mjs Chatbox Flow Start the local API before running the harness. Then build with USE_LOCAL_API=true ; the renderer API origin is compiled into the bundle. USE_LOCAL_API= true node ./node_modules/electron-vite/bin/electron-vite.js build --mode development pnpm eval :session-rag:chatbox -- -- case long-citrine-threshold --keep-user-data The harness copies the real config.json into an isolated temporary userDataDir, injects a temporary default chat model if missing, and stores the session RAG sqlite DB at a separate temporary path. Known Failure Modes "This attachment is too large..." means session_attachment_embedding capability is false or unavailable for the active local API/license path. If requests go to production, rebuild with USE_LOCAL_API=true . If Electron hangs at startup, check stale CDP/Electron processes. The harness now allocates a free CDP port. If the assistant says the file is still indexing after DB says ready, inspect the persisted user message. The message file must have sessionAttachmentIndexStatus: "ready" before context building. Do not trust a log line that says the message was submitted unless the textarea cleared and the session contains the user message. Useful Cases pnpm eval :session-rag:chatbox -- -- case long-citrine-threshold pnpm eval :session-rag:chatbox -- -- case implicit-citrine-current-policy pnpm eval :session-rag:chatbox -- -- case real-wiki-apollo-implicit-landing-site pnpm eval :session-rag:chatbox -- -- case multi-turn-real-wiki-apollo-followup pnpm eval :session-rag:chatbox -- -- case unrelated-simple-math pnpm eval :session-rag:chatbox -- -- case real-wiki-unrelated-capital For fast model-behavior iteration without Electron: pnpm eval :session-rag -- --dry-run pnpm eval :session-rag -- -- case implicit-citrine-current-policy
このスキルを起動するキーワード。クリックでコピーできます。

このスキルにはトリガーワードがありません。

ダウンロードした .skill に含まれるフィールド。
フィールド 説明
formatフォーマット識別子(skill/v1)
skill_idスキル固有 ID
nameスキル名
versionバージョン
description説明
categoryカテゴリ(配列)
trigger_wordsトリガーワード
tagsタグ
sourceソース
source_urlソース URL(本ページ)
exported_atエクスポート日時(ダウンロード毎)
system_promptシステムプロンプト本文
model_configモデル設定:provider / model / temperature / max_tokens / top_p
examplesサンプル
install_guide各プラットフォームの導入説明(Coze / Dify / Claude / カスタム)
同じスキルを各プラットフォーム形式で出力できます。
.skill 標準形式。system_prompt と model_config を収録し、任意の Agent で利用可能 ダウンロード
.skillpro 拡張形式。scripts / tools / dependencies / hooks を含む ダウンロード
.json 純粋な JSON 出力。system_prompt とモデル設定のみ ダウンロード
Coze frontmatter 付き Markdown。Coze へのインポート用 ダウンロード
Dify Dify DSL。アプリ作成後にそのままインポート ダウンロード

每日精选 Skill 推荐,免费送到你邮箱

输入邮箱,每天接收一个精选 AI Agent 技能推荐。完全免费,持续更新。

验证码 --

提交后我们会发送一封确认邮件,点击邮件里的链接才会开始收信。

完全免费,取消任意时间。我们不会发送垃圾邮件。