token-reducer

分类: 工具与效率 | 上传者: Madhan230205 | 下载: 0 | 版本: v1.0（最新）

每当用户需要减少上下文膨胀、降低令牌使用、总结大型代码/文档语料库、运行 FTS 加嵌入检索、重新排列顶级块以及在实施工作之前生成紧凑的上下文数据包时，请使用此技能。当用户说上下文太长、减少标记、优化提示大小、检索顶部块、压缩上下文或要求节省成本的提示工作流程时也会触发。

更新日志: Source: GitHub https://github.com/Madhan230205/token-reducer

目录结构

当前层级: tree/main/skills/token-reducer/

📁 references/
- 📄 context7-integration.md 322 B
- 📄 implementation-guide.md 523 B
📄 SKILL.md 1.4 KB

SKILL.md

---
name: token-reducer
description: Use this skill whenever the user needs to reduce context bloat, lower token usage, summarize large code/docs corpora, run FTS plus embeddings retrieval, rerank top chunks, and produce compact context packets before implementation work. Also trigger when users say context is too long, reduce tokens, optimize prompt size, retrieve top chunks, compress context, or ask for cost-saving prompt workflows.
license: MIT
compatibility: Python 3.10+ with SQLite FTS5. Fully local and free by default. Context7 MCP is optional.
allowed-tools: [Read, Glob, Grep, Bash, Task]
metadata:
  author: Madhan230205
  version: "1.0.0"
---

# token-reducer

Cut context size without cutting answer quality.

## Workflow

1. Preprocess large/noisy corpus into overlap-aware chunks.
2. Index chunks into SQLite FTS5 and local embeddings.
3. Retrieve with BM25-first policy.
4. Run vector retrieval in fallback mode by default.
5. Merge + rerank candidates; keep top 3-5.
6. Compress into citation-rich packet and output savings telemetry.

## Commands

- End-to-end run:
  - `python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" run --inputs . --query "${ARGUMENTS}" --hybrid-mode fallback --top-k 5`
- Self-test:
  - `python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" self-test`

## Deep References

- `./references/implementation-guide.md`
- `./references/context7-integration.md`

登录后下载/点赞/收藏 ❤ 5 | ★ 0

token-reducer

目录结构

SKILL.md

举报内容

提示